TThe Adaptability and Challenges of Autonomous Vehicles toPedestrians in Urban China
Ke Wang a , ∗ ,1 , Gang Li a ,1 , Junlan Chen b , ∗ , Yan Long a , Tao Chen c , Long Chen c and Qin Xia c a School of Automobile Engineering, the Key Lab of Mechanical Transmission, Chongqing University, Chongqing 400044, China b School of Economics & Management, Chongqing Normal University, Chongqing 401331, China c State Key Laboratory of vehicle NVH and Safety Technology, China Automotive Engineering Research Institute Company, Ltd., Chongqing 401122, China
A R T I C L E I N F O
Keywords :Chinese Pedestrian EnvironmentAdaptabilityAutonomous VehiclesRoad Safety
A B S T R A C T
China is the world’s largest automotive market and is ambitious for autonomous vehicles (AVs) de-velopment. As one of the key goals of AVs, pedestrian safety is an important issue in China. Despitethe rapid development of driverless technologies in recent years, there is a lack of researches on theadaptability of AVs to pedestrians. To fill the gap, this study would discuss the adaptability of cur-rent driverless technologies to China urban pedestrians by reviewing the latest researches. The paperfirstly analyzed typical Chinese pedestrian behaviors and summarized the safety demands of pedes-trians for AVs through articles and open database data, which are worked as the evaluation criteria.Then, corresponding driverless technologies are carefully reviewed. Finally, the adaptability would begiven combining the above analyses. Our review found that autonomous vehicles have trouble in theoccluded pedestrian environment and Chinese pedestrians do not accept AVs well. And more explo-rations should be conducted on standard human-machine interaction, interaction information overloadavoidance, occluded pedestrians detection and nation-based receptivity research. The conclusions arevery useful for motor corporations and driverless car researchers to place more attention on the com-plexity of the Chinese pedestrian environment, for transportation experts to protect pedestrian safetyin the context of AVs, and for governors to think about making new pedestrians policies to welcomethe upcoming driverless cars.
1. Introduction
Pedestrians are the vulnerable road users and account for23% of world traffic deaths in 2018 according to WHO(WorldHealth Organization, 2018). Naturally, China has the mostpedestrians around the world due to its largest populationand traffic developing state, and pedestrian safety has al-ways been a problem of China since 26.1% of traffic deathsare pedestrians in 2013, while in America it is 16.1%(WorldHealth Organization, 2016). This terrible result comes frombad road behaviors of Chinese pedestrians, such as red lightrunning; a report(N=31649) showed 18.54% of pedestrianswould run red lights in Changsha(Yan et al., 2016). Thesebad behaviors make the Chinese traffic environment verychallenging. Fig.1 is the common scene at Chinese cross-walks.AVs are deemed as promising solutions for safer roadtransportation in the future(Yang et al., 2018; Chen et al.,2018) and China is expected to become one of the largestmarkets for AVs(Wang et al., 2014). Therefore, it will bemeaningful to analyze the adaptability of AVs to better pro-tect pedestrians and give corporations and governments someuseful guidance.Driverless technologies are developing rapidly and ac-cording to the ability of self-driving,the Society of Automo- ∗ Corresponding author. [email protected] (K. Wang); [email protected] (G. Li); [email protected] (J. Chen); [email protected] (Y. Long); [email protected] (Y. Long) These authors contributed equally to this work.
Figure 1:
The challgening Chinese pedestrian environment tive Engineers(SAE) divides self-driving into six levels(L0-L5: from no automation to full automation)(SAE On-RoadAutomated Driving Committee, 2016), corresponding to driv-ing in pedestrians traffic of different difficulty levels. How-ever, the adaptability of driverless technologies to pedestri-ans is largely unknown. In previous researches, the manuscriptspaid attention to pedestrian detection(Hussein et al., 2016;Ouyang et al., 2018; Li et al., 2019), interaction(Rasouli andTsotsos, 2019a), receptivity(Deb et al., 2017; Jayaraman et al.,2018), behavior prediction(Volz et al., 2018), pose estima-tion(Fang and LÃşpez, 2019; Wang et al., 2019), etc. Theseresearches are related to the adaptability but did not directlypoint out the adaptability.
Ke Wang et al.:
Preprint submitted to Elsevier
Page 1 of 24 a r X i v : . [ c s . H C ] J u l ccident analysis and preventtion Similarly, review articles focused more on technology aswell. Generally, pedestrian detection is reviewed to sum-marize and compare the detection algorithms according tothe used sensors (Hurney et al., 2015; Lai and Teoh, 2014;Li et al., 2012, 2006; Simonnet et al., 2012; Yadav et al.,2015). Additionally, Deb et al. summarized the factors thatinfluence the pedestrians’ behaviors, public acceptance offully automated vehicles as well as current interacting in-terfaces between pedestrians and autonomous vehicles(Debet al., 2018a). Cao et al. conducted a review on differentmethods to model crowd of pedestrians(Cao et al., 2015).Daniela et al. explored the ways pedestrians’ intention esti-mation has been studied, evaluated, and evolved(Ridel et al.,2018). Kardi et al. provided a review of a microscopic pedes-trian simulation model(Teknomo et al., 2016). Sarker et al.reviewed human factors that would influence the acceptanceof users to AVs including user comfort, trust, reliability, andpreferences (Sarker et al., 2019). In these reviews, the detec-tion methods, influencing factors, and interacting interfaceswere carefully summarized and these are closely related toadaptability. Nevertheless, there is a lack of systematic anddirect analyses of the adaptability to pedestrians.Therefore,the adaptability analyses are valuable and needs to be sup-plemented.We have three contributions: firstly, the paper is the firstto summarize the Chinese pedestrians’ phenomena throughabundant data and comparision with foreign countries andanalyze the key safety demands for AVs. Secondly, we con-ducted a comprehensive literature review on the newest driver-less technologies to pedestrians, including detection, inter-action as well as receptivity. Thirdly, we executed the firsttry on the adaptability of AVs to Chinese pedestrians andsummed up the challenges as well as opportunities. The pa-per is useful for driverless researchers who care about pedes-trian safety and researchers from other areas, such as trafficsafety and public policy, who want to conduct researches re-lated to AVs, for we offer them guidance and research direc-tions.The rest of the paper is organized as follows. Section2 show the methods that we review and conduct our analy-ses. Section 3 analyzed three typical phenomena of Chinesepedestrians. A comprehensive literature review of driverlesstechnologies for pedestrians and adaptability analyses wereprovided in Section 4. In Section 5, emerging challenges andopportunities for future pedestrian research were proposed.Section 6 is the conclusion of this paper.
2. Methods
The objective of this article is to figure out the adaptabil-ity of AVs to Chinese pedestrians by reviewing articles. Thefirst question should be how to evaluate the adaptability log-ically. To analyze the adaptability step by step, we dividedit into two main questions: what are the Chinese pedestri-ans’ characteristics and their technical demands for AVs, and what are the current abilities of AVs when facing pedestri-ans? By solving the first, we know the safety demands forChinese pedestrians, which would be worked as our criteriato evaluate adaptability. Then, through reviewing articles,the current abilities of AVs would be summarized. By com-bining the two answers, the adaptability would be analyzedand develop the answer to the adaptability of AVs to Chineseurban pedestrians. The methodology is shown in figure 2.Firstly, pedestrain behaviors data was collected from theChinese government, research articles and websites as theleft side of figure 2 showed. In the searching process, weused terms like ’Chinese urban pedestrians’, ’pedestrian badbehaviors’, ’pedestrian accidents’, ’pedestrian safety’ to searchresults in engine (baidu, baidu scholar, google, google scholar)and filter them by reading the abstract. After collecting data,some typical behaviors of Chinese pedestrians are gained.Then, based on our technical background of AVs, the influ-ence of these pedestrian behaviors on AVs would be care-fully discussed. More importantly, some technical demandsfor AVs are proposed to protect pedestrian safety. In our re-search, these demands are worked as the criteria to evaluatethe adaptability. To make it concrete, we divide it into threeclasses: excellent, OK, bad, to evaluate the adaptability.Secondly, according to the demands, corresponding driver-less technologies are searched and summarized through scholarengine and authoritative websites. In the searching process,we combined terms like ’autonomous vehicles’, ’driverlesstechnologies’, ’pedestrians’, and ’pedestrian safety’ to searchprofessional articles and filtered them according to our de-mands. After that, the abilities of AVs are categorized intogroups and summarized by tables and figures. Consideringthe demands, we evaluate the abilities of AVs respectively.
One important thing is that our reviewed technologies arethe newest researches and are hard to tell which autonomouslevel they belong to. Therefore, in this article, we discussthe adaptability based on single technology rather than au-tonomous levels. Finally, the adaptability summary of everysingle technology would be our analyses of adaptability.Thirdly, during the process, we found some challengingbut promising problems and we analyzed them in a simpleway after the adaptability analyses, to give some guidance topractitioners.
3. Phenomena & Analyses
Pedestrians are vulnerable road users around the worldand recognized as the worst victims because they are di-rectly exposed to the impact of traffic crashes compared tovehicle passengers(Lee et al., 2019a; World Health Organi-zation, 2018). In developed countries, pedestrians would besafer. According to Road trauma Australia: 2018 statisti-cal summary(Bureau of Infrastructure, Transport, RegionalDevelopment and Communications of Austrilia, 2018), thedeaths-to-injuries rate of Australian pedestrians is 4%. In
Ke Wang et al.:
Preprint submitted to Elsevier
Page 2 of 24ccident analysis and preventtion
Figure 2:
The content structure of this paper
America, 5,977 pedestrians were killed with 71,000 injuriesin 2017 according to road traffic statistics(National HighwayTraffic Safety Administration, 2019). While in England, onefatal accident would take place in 50 accidents(Departmentfor Transport of the UK, 2019). However, as the biggest de-veloping country, China has more severe pedestrian safetyproblems. According to the statistics of National Bureauof Statistics of China(The National Bureau of Statistics ofChina, 2019), the deaths-to-accidents ratio of pedestrians isabout 50%, which is far worse than that of car passengers aswell as cyclists as Table 1 shows; in other words, there willbe a person dead in every two pedestrian accidents. There-fore, protecting pedestrians is an essential problem of Chi-nese traffic safety.However, due to the low average education level and aware-ness of obeying rules, Chinese pedestrians often violate traf-fic rules and cause accidents, adding complexity to the Chi- nese driving environment. The driverless car is a potentialsolution to promote traffic safety, however, whether AVs aresuitable for the Chinese pedestrian environment is largelyunknown. To assess the adaptability, three typical behaviorsof Chinese pedestrians were summarized in the followingthrough literature as well as open databases and we foundChina is one of serious countries of these behaviors. Af-ter that, the characteristics of Chinese pedestrians were an-alyzed. On the basis, the key technical demands for AVswere put forward as the evaluating criteria of adaptability.The followings are the analyses of three typical behaviors.
Red lights running of pedestrians greatly influences theirsafety. In Lille of France, one third of pedestrians are re-ported running red lights in 2015(Dommes et al., 2015). InAmerica, 4.5% of pedestrians traffic deaths happened when
Ke Wang et al.:
Preprint submitted to Elsevier
Page 3 of 24ccident analysis and preventtion
Table 1
The ratio of deaths to accidents of Chinese traffic participants(The National Bureau ofStatistics of China, 2019)Traffic participants data 2017 2016 2015 2014 2013 2012 2011 2010All Accidents 203049 212846 187781 196812 198394 204196 210812 219521Deaths 63772 63093 58022 58523 58539 59997 62387 65225
Deaths/accidents 0.31 0.30 0.31 0.30 0.30 0.30 0.29 0.30
Cyclists Accidents 1576 1460 1369 1393 1304 1433 1522 1978Deaths 350 341 304 289 300 279 315 447
Deaths/accidents 0.22 0.23 0.22 0.21 0.23 0.19 0.21 0.23
Car passengers Accidents 139412 145820 129155 136386 138113 142995 145338 148367Deaths 46817 45990 42388 42847 42927 44679 46100 46878
Deaths/accidents 0.34 0.32 0.33 0.31 0.31 0.31 0.31 0.32
Pedestrians Accidents 2470 2443 2137 2242 2088 2063 2277 2565Deaths 1322 1304 1192 1247 1185 1075 1134 1222
Deaths/accidents 0.54 0.53 0.56 0.56 0.57 0.52 0.50 0.48Table 2
Statistics of red lights running of pedestrians in Hangzhou (thefirst column is the different possibility to run red lights and thesecond is corresponding percentages of pedestrians, N=200)(Mengyun et al., 2013)Possibility of running red light 80% 50% 10% 0%Percentage 2.5% 17.5% 88% 1%
Table 3
The impact of groups on running red lights of pedestrians inYulin(Single denotes pedestrians run red lights alone and groupmeans pedestrians together run reds lights, N=3460) (Kedenget al., 2015))Pedestrians state Obey the rules Run red lightsSingle 76.7% 23.3%Group 70.2% 29.8%
Table 4
PedestriansâĂŹ attitudes toward running red lights inXian(N=425)(Qiang et al., 2017)Attitudes Agree to run redlights DisagreePercentage 60% 40% they failed to obey traffic signals according to national statis-tics in 2017(National Highway Traffic Safety Administra-tion, 2019). However, running lights running is more seriousin China and is nicknamed the Chinese style of road crossing.Table 2, 3, 4 as well as 5 are statistics of running red lightsdone in four main cities in China. From these data, runningred lights is a regular behavior when pedestrians are crossingthe crosswalk with a possibility up to 70%. What is more, theattitudes of pedestrians toward obeying light rules are nega-tive because 60% of pedestrians are approved of running redlights as table 4 shows. Table 3 reflects that the number ofpedestrians influences the violating behaviors since walkingin group increases the possibility to violate lights, which isrelated to the conformity phenomenon of psychology. The Fig. 3a, b, c are common scenes in Chinese cross-walks, which shows three levels of running red lights ac-cording to the pedestrian number. The conflict in the scenesis that cars have to cross the crosswalk but to keep pedes-trians safe. When there are minority persons violating thelights, their bodies, trajectories, tendencies as well as intentare clear and cars could slow down to avoid at sacrifice ofefficiency. To make matters worse, influenced by the con-formity phenomenon, persons would violate lights in groupas 3b shows, hence leading to pedestrian occlusion. The oc-clusion would cause a misunderstanding of pedestrian intentand wrong detection of pedestrians, possibly causing acci-dents. In rush hours, the pedestrian flow is even walking onthe vehicle road. In this case, it is difficult for the cars tomove forward but wait.Analyzing the behaviors, we conclude detection plays an es-sential role in red lights running scenarios. Good detectionmeans cars know the right position as well as features ofpedestrians, and could predict the intent. Based on detec-tion, cars could decide whether they should yield or keeptheir movements when pedestrians are running red lights.Additionally, special care would be given to those specialpedestrians like the elderly, the disabled, the distracted aswell as the wheelman according to the detected features. Inthe context of traditional cars, drivers could detect almost allpedestrians in the first level of red lights running accordingto human intelligence. When more pedestrians violate thelights, drivers impossibly detect all pedestrians as a result oflimited attention and occlusion. AVs could understand theworld through multiple sensors, so could they detect pedes-trians well in the red lights running scenarios?Additionally, intent communication between pedestriansand vehicles is of vital importance in the process. Generally,intent communication includes the moving state of pedestri-ans and cars and their observation of each other. Through de-tection, drivers know pedestrian position and features, how-ever, who should take the road right to go need negotiating,which is tackled by communication. In the current driving
Ke Wang et al.:
Preprint submitted to Elsevier
Page 4 of 24ccident analysis and preventtion
Figure 3:
Three typical bad behaviors of Chinese pedestrians: a), b), c) running red lights;d), e), f) jaywalking; g), h), i) distraction (phone utilization)
Table 5
Frequency of running red lights of pedestrians in Shanghai(N=500)(Zhichan, 2017)Behaviors Run red lights See others run red lightsFrequency Usually Occalisonally Never Usually Occasionally NeverPercentage 12% 65% 23% 67% 30% 3% environment, there already exist some nonverbal methods tocommunicate pedestrians with cars, such as car light (head-lights, turn lights, rear lights), distance, voice of the cars.In addition, drivers play an important role in communica-tion. During the process, drivers could interact with pedes-trians by utilizing eye contact, gestures as well as voice to as-sign road right (Charisi et al., 2017)as shown in Fig. 4a,andLee et al.(Lee et al., 2019b) concluded that the existence ofdrivers would strength the safety-in-numbers effect to pro-tect pedestrians by increasing the interaction between pedes-trians and vehicles. More importantly, it will give pedestri-ans more trust to communicate with humans rather than ma- chines. Compared with traditional cars, there will be no realdriver in AVs as Fig. 4b shows, so how do AVs commu-nicate with pedestrians and whether pedestrians would trustAVs?
Jaywalking means pedestrians pass a road illegally forconvenience where there is no crosswalk or some leadingsigns, possible with a baluster-passing behaviors. In the UK,nearly 50% of pedestrians crossed the road ignoring the traf-fic signals according to YouGov poll 2013(Jordan, 2013). InAmerica, a similar YouGov poll in 2014 found that only 13%of Americans said they had never jaywalked(Moore, 2014)
Ke Wang et al.:
Preprint submitted to Elsevier
Page 5 of 24ccident analysis and preventtion
Figure 4:
Intent communication methods between pedestrians and vehicles: a) traditionalvehicles and pedestrians b) AVs and pedestrians and America national statistics showed that 21.2% of pedes-trians traffic deaths came from pedestrians’ improper cross-ing of roadway(National Highway Traffic Safety Adminis-tration, 2019). The Singapore government fined 2,049 jay-walkers in the first quarter of 2013(Ren et al., 2011). Simi-larly, jaywalking is very common in China because of unrea-sonable crosswalk settings and a lack of awareness to obeyrules. From Table 6, we could see jaywalking is very dan-gerous for 68% pedestrian accidents are related to accidents.What is more, the research of (Jun et al., 2012)showed thatno matter in urban, suburbs or countryside, jaywalking is thebiggest cause of car accidents according to Fig.5.Fig. 3d, e, f are common scenes of jaywalking in urbanChina, from which we could observe jaywalking often oc-curs in the area without signs, even with a fence in the mid-dle of the road. In these scenes, the trajectories of pedestri-ans are irregular and difficult to predict because there are nosigns to regulate their movements, and they are faster thanon the crosswalks. For drivers, they would be partly lack-ing in concentration as a result of no expectation for pedes-trians; additionally, cars would be faster due to the higherspeed limit on these sections. For the road itself, some spe-cial sections, such as curved sections, cars and pedestriansare naturally difficult to detect each other.These features contribute to most of jaywalking accidents,which could be divided into three types. The first type of ac-cidents is the failure to detect each other. The drivers do notobserve pedestrians cross the road and hence they crash intopedestrians directly because of no expectation for pedestri-ans or on special sections. The second is untimely detectionto each other. In this case, it is difficult for drivers to avoida collision as a result of vast inertia out of high speed. Thethird is poor communication between drivers and pedestri-ans. With high speed, both drivers and pedestrians have verylittle time to negotiate the road rights, often leading to poorcommunication.Similarly, detection is essential to keep pedestrians safe in
Figure 5:
The causes of pedestrian accidents in three areas inChongqing(Jun et al., 2012) the jaywalking. In this scenario, cars must detect pedestriansfrom long distance because cars are so fast that drivers needmore time and longer distance to stop. Moreover, enoughcommunication is required to transfer the moving state. Withhigh speed and far distance, the communicating methods be-tween pedestrians and drivers would be limited and easilymisunderstood. Therefore, more communicating methodsare needed to better transmit the state. Traditional cars aretroubled by remote communication and small pedestrian de-tection. As a high technology, how do AVs interact with re-mote pedestrians and could they detect remote small pedes-trians?
Ke Wang et al.:
Preprint submitted to Elsevier
Page 6 of 24ccident analysis and preventtion
Table 6
The behaviors of pedestrians just before pedestrian accidents took place(N=181)(Shiboet al., 2018)Behaviors
Jaywalk
Walk oncrosswalk inthe section Walk oncrosswalk intheintersection Stand onthe road Work bythe road Walk bythe road othersPercentage
Table 7
The attitudes of using phone on crosswalks of pedestrians in Hefei(questionnaire,N=405)(Feiyang, 2017)Attitudes Use phone when crossing the roadthis week Involved in accidents because ofphone utilization in crosswalk Whether to be punishedPositive 162(40%) 22(5.4%) 206(51.7%)Negative 243(60%) 383(94.6%) 199(48.3%)
Distraction refers to pedestrians doing other things andnot paying full attention to the environment when they arepassing the road, and the possible reasons are reading, eat-ing, using smartphones, talking, and the influence of alcohol,drugs, or medication. As Judith et al. showed(Mwakalongeet al., 2015), phone usage has been the major reason of pedes-trian distraction recently, therefore, the following mainly talkabout phone usage.In America, a report of Liberty Mutual Insurance in 2013surveyed 1,004 adults and found that 60% of them had usedphones when crossing the crosswalks(Liberty Mutual Insur-ance, 2013). Moreover, the Consumer Product Safety Com-mission of America estimated that 3.76% of American in-juries are related to phone usage and Nasar et al. (Nasarand Troyer, 2013a) thought injuries caused by phone usagewould be much bigger. In England, a report in 2019 foundthat 31.37% of adolescent pedestrians crossed the crosswalksusing phones(Baswail et al., 2019). Similarly, phone us-age of pedestrians has been an important problem of Chi-nese pedestrian safety.The table 7 as well as 8 are statis-tics of phone usage on crosswalks in Hefei and Wuhan re-spectively while Fig. 6 shows data collected in Beijing andChongqing. From these data, it could be concluded thatabout 12% of pedestrians would really utilize phones whencrossing the road. What is more, about half of the pedes-trians thought they should not be punished because of us-ing phones(Feiyang, 2017). This reflects a lack of aware-ness to obey traffic rules of Chinese pedestrians. As canbe seen from table 8, watching telephone screens, having acall and listening to music are three aspects of phone usageand watching phones is the most common behavior. Dur-ing these processes, the attention of pedestrians is occupiedso they hardly perceive surroundings and communicate roadrights with vehicles.Ling et al. emphasized that the time to cross a road wouldincrease and the times pedestrians watch around would de-crease when they use phones(Feiyang, 2017). According to a research(Nasar and Troyer, 2013b), the proportion of pedes-trians killed while using phones has increased by more than3.5% in 2010. Additionally, pedestrians using phones aremore likely to conflict with vehicles than pedestrians not us-ing phones.There are two scenarios where pedestrians use their mobilephones at the crosswalk; the first is when the pedestrian traf-fic light is green and the second is when the pedestrian trafficlight is red. In the first case, pedestrians might be relativelysafe since pedestrians have the road right while cars have towait. However, drivers might run red lights when pedestri-ans cross the road using phones. In this situation, distractedpedestrians cannot avoid a collision. In the second case, itwill be extremely unsafe for pedestrians. Firstly, cars maynot realize pedestrians are utilizing phones and think pedes-trians would behave like normal persons. Furthermore, thedistraction would hinder the intent communication betweenpedestrians and vehicles. As a result, drivers do not knowhow to respond to the behaviors accordingly and cause acci-dents.From the analyses above, we could conclude the unsafetycomes from little communication and poor detection whenpedestrians are utilizing phones, which highlights the impor-tance of communication as well as detection again. In thecontext of traditional cars, drivers could partially recognizewho are using phones and slow down to keep safe. Never-theless, the intent communication remains a problem. Underthe context of AVs, could AVs detect phone utilization andcommunicate with pedestrians well?
Based on the above phenomena and analyses, pedestriansafety requires good detection and communication betweencars and pedestrians. In the following, the technical demandsof pedestrians for AVs would be summarized. Simultane-ously, we propose criteria to evaluate the ability of each tech-nology.In the detection part, in order to protect the safety of normal
Ke Wang et al.:
Preprint submitted to Elsevier
Page 7 of 24ccident analysis and preventtion
Table 8
The statistics of using phone on crosswalks in Wuhan(empirical research,N=2901)(CunBao et al., 2018)Behaviors Use phones when crossing the road Not use phoneswhen crossing the roadPercentage Watch phone Have a call Listen to music 88.24%6.65% 2.83% 2.28%
Figure 6:
The phone usage of pedestrians in different scenarios: a) Beijing (empiricalresearch)(Cun et al., 2019) b) Chongqing (questionnaire)(Jinhua, 2017) pedestrians (under no bad behaviors), vehicles should knowthe location and features of pedestrians. Moreover, occludedpedestrian detection is required to solve the problems of oc-clusion, which often takes place in red lights running. Ad-ditionally, remote(small) pedestrian detection is essential toprotect pedestrians when pedestrians are jaywalking. Lastbut not the least, phone utilization detection is required todistinguish normal pedestrians and distracted ones. To eval-uate the detection ability, the precision of location and clas-sification will be reviewed and summarized.In terms of communication, other than existing physical in-formation transferring, more interfaces should be designedfor interaction. For example, more external interfaces areneeded to transfer the intent of cars to pedestrians in the sce-narios of occlusion as well as jaywalking. As mentionedabove, we suggested the state of AVs should be well trans-ferred, which would be worked as our evaluating metric. Tomake it concrete, we propose the intent transferring com-prised of stopping, accelerating, slowing down, as well asthe observation of pedestrians, to evaluate the ability of com-munication.Importantly, there is a hidden factor, receptivity, that playsan important role in the relationship between AVs and pedes-trians. If pedestrians cannot accept AVs, the communica-tion is meaningless for pedestrians could not trust the infor- mation. Furthermore, there will be significant obstacles tobring AVs into the market. In this article, we have consid-ered more about Chinese pedestrians, hence, we take the re-ceptivity of Chinese pedestrians into account.In summary, the adaptability of AVs depends on the adapt-ability of communication, detection as well as receptivityand hence we evaluate the adaptability of the three technolo-gies to figure out the adaptability of AVs. The followingsection 3 will talk about the adaptability of AVs to pedestri-ans in detail. At the same time, the adaptability relationshipwould change the current mentality of pedestrians and askfor the governmentâĂŹs regulations adjustment. Based onthe adaptability, we would simply analyze the influence thatAVs would bring to current pedestrians mentality and gov-ernment regulations.
4. The Adaptability of Driverless Technologiesto Pedestrians
China has a challenging driving environment of pedestri-ans, in which running red lights, jaywalking as well as phoneusage take place usually. These behaviors lead to crowd, oc-clusion and poor communication, greatly harming the safetyof pedestrians. Analyzing the scenarios, detection, commu-nication and receptivity are summed up as key technical de-
Ke Wang et al.:
Preprint submitted to Elsevier
Page 8 of 24ccident analysis and preventtion mands for safety. As a promising technology, AVs could per-ceive the world through multiple sensors, so can AVs solvethese problems? What are the essential driverless technolo-gies?The adaptability could be divided into two parts: AVs adaptto pedestrians and pedestrians adapt to AVs. Correspondingto the demands, driverless technologies that are dedicated tosolving these problems are divided into detection, interac-tion, and receptivity. In these researches, detection appliesvarious sensors and algorithms to determine whether thereare pedestrians and find their location, which shows the abil-ity of AVs to feel the world and hence describes the adapt-ability of AVs to pedestrians. According to the demands,normal pedestrian detection, occluded pedestrian detection,small pedestrian detection and distracted pedestrian are re-viewed and analyzed respectively. Furthermore, the tech-nology of interaction utilizes different kinds of external in-terfaces to transfer the state of AVs to pedestrians, whichshows the ability of pedestrians to understand AVs and hencedescribe the adaptability of pedestrians to AVs. And theresearches of receptivity are committed to seeking factorsthat influence the acceptance of pedestrians to AVs; this re-flects the pedestriansâĂŹ trust in AVs and also describes theadaptability of pedestrians to AVs.The following will examine these technologies by re-viewing articles. After that, we would give out the adapt-ability by combining the analyse of phenomenon with tech-nologies. To get an intuitive result, three levels of descrip-tions are applied to evaluate the adaptability: bad is used forpoor research status and terrible adaptability; OK is used forgood research status but there is still room to improve; ex-cellent is used for splendid work and good adaptability.Furthermore, after summarizing the adaptability, the influ-ence on these driverless technologies to the mentality of pedes-trians and government regulations would be talked about ina simple manner, to offer some guidance to Practitioners.
As discussed above, detection is important to guaran-tee pedestrian safety. As a promising technology, AVs paymore attention to the detection task to recognize and local-ize pedestrians. However, the occlusion, small size as wellas distraction of pedestrians make it a challenging task inChina. To recognize the capabilities, the detection task wouldbe divided into four subtasks including normal, occluded,small and distracted pedestrian detection.In order to better review the ability, we first summarizedthe pedestrian dataset.
The popularity of data-based methods promotes the im-portance of datasets. The improvement of dataset also re-flects the technical tendency of pedestrian detection. Theauthoritative pedestrian datasets are summarized in Table 9.These datasets are mainly used to evaluate the capabili-ties of detectors, of which Caltech dataset as well as KITTI dataset are popular datasets in recent years. The numberof pedestrians is related to the number of ground truth andframes, representing the difficulty of detection. Caltech datasetranks first with 350,000 pedestrians at 1,000,000 frames. Ad-ditionally, the dataset released later often has a larger num-ber of pedestrians because data-based methods dominate themainstream. As the year increases, occluded labels are addedto the dataset more because occlusion usually occurs in realscenarios, which is why KITTI as well as Caltech datasetsbecome popular currently. The occlusion level could be eval-uated by the fraction of occlusion, which can be calculatedas one minus the fraction of visible area over the total area.However, KITTI and Caltech datasets have different occlu-sion labels. KITTI dataset divided scenes into three levels:easy, moderate and heavy whereas Caltech grouped theminto no, partial occlusion as well as heavy occlusion basedon occlusion ratio, which is shown in the third column. Nat-urally, the two datasets both take the occlusion into accountwhen evaluating detectors.To emphasize different things, the metrics utilized to eval-uate the abilities of detectors are different. The ROC curveevaluates the detectors in terms of true positive rate and falsenegative rate in the early dataset. The MR-FPPI curve usedby Caltech and ETH reflects the relationship between missrate (MR) and false positive per image (FPPI), and used missrate value when the FPPI value is 10-1 as the evaluationmetric. In order to emphasize precision and recall, KITTIdataset used the P-R curve to evaluate the performance ofthe detectors and utilize the area between the P-R curve andcoordinates as the evaluation metric. Recently, the P-R andMR-PFFI curves are utilized more in the detection task dueto the popularity of KITTI and Caltech datasets. Reason-ably, each dataset has an optimal detector shown in the fifthcolumn. We cannot get a specific optimal detector in somedatasets so we use a dash to represent. In the early dataset,detectors gain a splendid detection rate even around 100%while detectors get worse in recent as a result of increasingdifficulty in datasets. Additionally, the popularity of neuralnetwork-based methods are reflected in terms of the name ofoptimal detectors. More importantly, the websites of thesedatasets would rank detection methods to help researchersfollow the latest development and this would be our approachto evaluate the capabilities.Last but not the least, KITTI dataset collected pictures fromcampus, city, road, residential and person scenes, which iscloser to real scenarios in an autonomous driving environ-ment. However, the urban Chinese pedestrians would beonly talked in this paper so Caltech dataset would be moresuitable.
Normal pedestrians are pedestrians who do not performbad behaviors. We set the occlusion level, size as well asthe distraction of these pedestrians to normal values. Nor-mal pedestrian detection could assess basic detection capa-
Ke Wang et al.:
Preprint submitted to Elsevier
Page 9 of 24ccident analysis and preventtion
Table 9
Comparison of pedestrian detection datasetsDataset Pedestriannum-ber Occlusion labels EvaluationMetrics Best detector &performances YearMIT(Papageorgiouand Poggio, 2000) 924 YES MR-FPPW (HOG,100%) 2000USC(Wu andNevatia, 2005) 816 NO ROC - 2005INRIA(Dalal andTriggs, 2005) 1774 NO MR-FPPW (F-DNN, 7%) 2005Daimler(Munderand Gavrila, 2006) 4000 NO ROC (MLS,28%) 2006CVC(GerÃşnimoet al., 2007) 1000 NO ROC - 2007ETH(Ess et al.,2007) 12,000 NO R-FPPI (RPN+BF, 30%) 2007TUD-Brussels(Wojeket al., 2009) 3247 NO P-R (SpatialPooling,47%) 2009Caltech(Dollár et al., 2009)(Dollar et al., 2011) 350,000 YES MR-FPPI (ARPed,6%) 2009Noocclusion(0%) Partial(1%-35%) Heavy (35%- 80%)KITTI(Geiger et al.,2012) 80,000 YES P-R (FichaDL, 81.73%) 2012(2015)Easy(0-15%) Moderate(15%-30%) Hard(35%-50%) MR, FPPW, FPPI, R and P respectively infer to miss rate, false positive per window, false positive per image, recall andprecision. the higher, the better in MR-FPPW and P-R curve whereas the lower, the better in ROC, MR-FPPI, MR-FPPWand R-FPPI curve. ROC curve takes true positive rate and false negative rate as coordinate bilities because normal pedestrians are the largest parts ofpedestrians. To check the ability, the performance of detec-tors is summed up based on four popular pedestrian datasetsin Fig. 7. The moderate and reasonable occluded levels ofKITTI and Caltech datasets are selected to meet the demandsof normal pedestrians.It is observed that detectors in Caltech are the best with thelowest miss rate of 6% among all the four datasets. The bestdetector in INRIA is slightly behind with a 7% precision.However, the optimal detector named F-DNN2+SS in ETHdataset achieved the best miss rate of just 30%. This is pos-sibly because it is so old-fashioned that the latest detectorswould not validate performance on it. Interestingly, the dif-ference between Caltech training dataset and Caltech testingdataset is huge, which is because data-based methods use atesting dataset to validate while features-based methods usea training dataset. This also reflects the superiority of data-based methods.In KITTI dataset, the area of the curve is utilized to evaluatedetection performance and the best detector achieves an av-erage moderate precision of 81.73%. Roughly speaking, wecould say the best detectors in Caltech with 7% miss rate is better than in the best detectors in KITTI with 81.73% pre-cision. As talked above, the scenarios in Caltech dataset aremore suitable for the setting of Chinese urban pedestrians,hence, the result of Caltech is more reliable in this article.In conclusion, the capabilities of normal pedestrian detec-tion are excellent with a miss rate of 6% and the driverlesstechnology of normal pedestrian detection is adaptive to ur-ban China.
Pedestrian occlusion usually happens in driving scenar-ios and about 70% of pedestrians have occlusions accordingto Caltech dataset(Dollár et al., 2009; Dollar et al., 2011),which is among the toughest problems in pedestrian detec-tion. Since Chinese pedestrians perform badly on crosswalks,occlusion has always been a serious problem. Comparedto no occlusion, occluded pedestrian needs two boundingboxes to jointly label, one for visible portion and one forfull extent. Furthermore, occlusion could be divided intotwo categories: pedestrians are occluded by other objects,which often causes information missing and leads to false
Ke Wang et al.:
Preprint submitted to Elsevier
Page 10 of 24ccident analysis and preventtion
Figure 7:
The performances of 10 top detectors in five popular datasets: a) Caltechtraining (Dollár et al., 2009; Dollar et al., 2011) b) Caltech testing(Dollár et al., 2009;Dollar et al., 2011) c) ETH (Ess et al., 2007) d) INRIA (Dalal and Triggs, 2005) e) KITTI(Geiger et al., 2012)(Data is summarized in October 1, 2019) negatives, and pedestrians are occluded by other pedestri-ans, which brings lots of interfering information and leadsto false positives(Wang et al., 2018b).To figure out the capabilities, the top 10 detectors in KITTIand Caltech datasets are summarized in terms of different oc-clusion levels in Table 10. The hard level of KITTI datasetand the heavy level of Caltech dataset are the most occludedscenarios.It is reasonable that when occlusion levels improve, ac-curacies would significantly decrease. The detectors in easyocclusion of KITTI has an excellent performance since 88.27%of pedestrians are detected (the same is true in no occlusionin Caltech). However, the performance becomes particu-larly poor, especially in heavy occlusion in Caltech (hard inKITTI), for 49% of pedestrians are missed. More impor-tantly, the change rate is huge when the difficulty of scenar-ios changes from easy to hard, especially in Caltech datasetwith a change rate of about 40%. This is because Caltechdataset has a more complicated occlusion status. However,occlusion is common and sometimes there is heavy occlu-sion taking place in jaywalking or crosswalk crossing sce-narios. Therefore, both the easy level and heavy level ofocclusion detection are important in the Chinese urban envi-ronment. Above all, driverless technologies perform excel- lent detection in easy occlusion whereas detectors performvery poorly in the heavy level. Considering the importanceof easy occlusion and heavy occlusion, we think driverlesstechnology has overall bad research in occluded pedestriandetection due to the unacceptable result of heavy occlusion.
Small pedestrians mean that pedestrians take up smallparts of pixels in the eyes of AVs. As analyzed above, jay-walking is very common and has caused a number of ac-cidents in the Chinese environment. In the scenarios, thespeed of vehicles is fast and they need to detect pedestri-ans at long distances to take timely measures. Therefore,small pedestrian detection is important to protect pedestri-ans. However, small pedestrian detection is complicated dueto low resolution and noisy presentation(Li et al., 2017).In the authoritative dataset, COCO of Microsoft(Lin et al.,2014), they define objects smaller than 32*32 pixels as smallobjects. In the KITTI dataset, researchers set a minimizedbounding box height of 25 pixels for moderate as well ashard level while Caltech dataset categorized them into nearscale, medium scale and far scale. Additionally, Dollar etal. (Dollár et al., 2009)thought that pedestrians from 30 to80 pixels are most important for automotive settings, respec-tively correspond to 20m and 60m away from the pedestrians
Ke Wang et al.:
Preprint submitted to Elsevier
Page 11 of 24ccident analysis and preventtion
Table 10
The performances of 10 top detectors of KITTI & Caltech datasets in different occlusionlevels(data is summarized in October 1, 2019)Rangking & Methods KITTI dataset Rangking & Methods Caltech dataset Easy Moderate/Change Hard/Change No Patial/Change Heavy/Change1. FichaDL 88.27% 81.73% /6.54% 75.29%/12.98% 1. AR-Ped(Nasarand Troyer,2013b) 5.00% 12.00%/7% 49.00%/44%2.Alibaba-CityBrain 88.13% 80.90%/7.23% 74.08%/14.05% 2.SDS-RCNN(Brazilet al., 2017) 6.00% 15.00%/9% 59.00%/53%3. ExtAtt 87.95% 79.63%/8.32% 74.78%/13.17% 3.F-DNN+SS(Duet al., 2017) 7.00% 15.00%/8% 54.00%/47%4. DGIST-CellBox 87.77% 79.54%/8.23% 75.70%/12.07% 4. F-DNN(Duet al., 2017) 7.00% 15.00%/8% 55.00%/48%5. DH-ARI 87.43% 78.29%/8.32% 69.91%/17.52% 5. PCN(Wanget al., 2018a) 7.00% 16.00%/9% 56.00%/49%6. EM-FPS 84.93% 77.61%9.14% 72.52%/12.41% 6.F-DNN2+SS(Duet al., 2018) 6.00% 16.00%/10% 40.00%/34%7. F-PointNet(Qiet al., 2018) 87.81% 77.25%/7.32% 74.46%/13.35% 7. GDFL(Linet al., 2018) 6.00% 17.00%/11% 43.00%/37%8. TuSimple(Heet al., 2016; Yanget al., 2016) 86.78% 77.04%/10.56% 72.40%/14.38% 8. ADM(Zhanget al., 2018) 7.00% 18.00%/11% 30.00%/23%9. THICV-YDM 87.27% 76.91%/9.74% 69.02%/18.25% 9. TLL-TFA(Songet al., 2018) 6.00% 18.00%/12% 29.00%/23%10. Argus-detection-v1 83.49% 75.51%/10.36% 71.24%/12.25% 10. MS-CNN(Caiet al., 2016) 8.00% 19.00%/11% 60.00%/52% KITTI dataset uses P-R curve to evaluate detectors and the higher, the better. Caltech dataset uses MR-FPPI curve to evaluate detecors and the lower, the better. Change denotes the value of current level minus the value of first column of each dataset;
Figure 8:
Small pedestrian detection in three scales in Caltech dataset(Data is summarizedin October 1, 2019) at urban speed of 15m/s. Above all, we would focus moreon pedestrians with the range from 30 to 80 pixels, whichinclude small pedestrians in autonomous settings.In the COCO dataset, general small object detection reflectsan average precision of 0.343, compared to medium objectsof 0.556 and large objects of 0.660. This also shows the dif-ficulty of small object detection. As for small pedestrian de- tection, Caltech researchers summed up the performance ofdetectors in three scales shown in Fig. 8. Reasonably, detec-tors in near scale has the best performance with the lowestmiss rate of nearly 0%. However, when it comes to mediumscale, the performance drops dramatically with the lowestmiss rate of 23%. To make matters worse, detectors in thefar scale have a terrible miss rate of only 60%. As analyzed
Ke Wang et al.:
Preprint submitted to Elsevier
Page 12 of 24ccident analysis and preventtion above, the medium-scale (30-80 pixels) pedestrian data ismore suitable for Chinese urban pedestrians. Therefore, thebest performance is the TLL-TFA algorithm with a miss rateof 23%. In summary, driverless technology is OK to detectsmall pedestrians but has some room to improve.
Distraction usually happens when pedestrians are read-ing books, talking to others as well as using phones andphone utilization is the main reason. Distraction can seri-ously harm the safety of pedestrians as discussed above. Inthe context of AVs, it is essential to detect pedestrians usingphones and accordingly decide how to respond.Although the phone utilization of drivers gains much atten-tion in research, there is a lack of researches about pedes-trians using phones. Akshay et al. proposed a vision-basedframework to classify whether a pedestrian is using a phone,with the highest accuracy of 91.20%(Rangesh et al., 2016).At the same time, a pedestrian dataset using phones was pro-posed as well, in which pedestrians are in high resolution.However, there is little research to further the study.Yet, surprisingly, phone utilization hasnâĂŹt always beenterrible for AVs. Vehicle-to-pedestrians(V2P) communica-tion is thought as a method to guarantee active safety, inwhich smartphone works as a receiving device. Ahmed etal.(Hussein et al., 2016) proposed a V2P-based applicationto broadcast alert information, including possible collisiondistance and time, to both phone users and AVs. At the sametime, Pooya et al.(Rahimian et al., 2018) proposed a modelusing phones to send alert warnings to pedestrians when theyare around the crosswalk. Furthermore, He et al.(He et al.,2017) proposed a V2P model utilizing WIFI, Bluetooth aswell as DSRC technology to establish interaction betweenAVs and pedestrians. In these cases, the smartphone playsan important role to avoid collision rather than lead to acci-dents.All in all, there exists research with excellent phone uti-lization detection of 91.20%, nevertheless, no other researchesare conducted to support and further the study. Moreover,the smartphone has played a totally new role to interact withthe pedestrians rather than ruin the safety, and the methodperforms well. In conclusion, there is not systematic workto support the accurate detection to distracted pedestrians butthe phone has become another way to protect safety. Hence,we think the driverless technology is OK and has some roomto promote.
To summarize the above conclusions, normal pedestriandetection is excellent. Occluded pedestrian detection is bad.Distracted pedestrian detection is OK. Small pedestrian de-tection is OK.
According to Gunnar(Johannsen, 2009), human-machineinteraction(HMI) transfers communication information be- tween human users and machines via human-machine in-terfaces. Under the context of traffic, the interaction takesplace between vehicles and road users, such as pedestrians,and human-machine interfaces are utilized as the interactingmethods. When pedestrians enter the road network, they be-gin constant information exchange with the traffic environ-ment (Deb et al., 2018b). Traditionally, drivers interact withpedestrians to negotiate road rights. While under the con-text of AVs, drivers do not have the control right of the cars(above Level 4) and cannot know the state. Sometimes theperson sitting on the seat might be distracted, leaving pedes-trians to infer the state of AVs alone. In this case, pedes-trians could get limited information by observing the speedand distance. There is a special need for alternative commu-nication techniques of AVs, which must be able to substitutefor the gaze of the driver(Saleh et al., 2017a). Furthermore,the interaction helps to boost acceptance and develop propermental models when AVs are originally put into markets.More importantly, it is crucial to inform pedestrians whenAVs have a failure state(Steinfeld et al., 2018).
Recently, a lot of researches has been conducted on in-teraction, focusing on the human-machine interfaces. Theused interfaces and transferring information are summarizedin Table 11. Google developed a patent utilizing LED dis-plays to notify pedestrians of the state of AVs(Urmson et al.,2015). Tobias et al. (LagstrÃűm and Lundgren, 2015a) usedLED strips to conduct interaction, in which lights are laid ina line and different locations as well as light numbers showdifferent information. The middle four lights light up to ex-press autonomous mode is on; lights expand to the sides, in-dicating AVs notice pedestrians while the lights shrink fromthe sides to the middle, which means AVs is about to start;all lights light up to indicate AVs are resting.Karthik et al. proposed a fusion interface to communi-cate with pedestrians. In their studies, infrastructures andelectric devices can also become important interfaces of in-teraction(Mahadevan et al., 2018). They have proposed fourprototypes and prototype 1 utilized a speaker coupled withLED strips on cars. In order to transfer intent, the LED stripwas mounted on the vehicle and exhibit four states. Solid redlights indicate the pedestrian shouldnâĂŹt cross as the ve-hicle would not stop. Blinking blue lights mean the vehiclewas aware of the pedestrian. Green lights moving from leftto right indicate the vehicle had fully halted and it was safeto cross. Purple lights moving from right to the left meat thevehicle would start soon.Milecia et al. proposed a method fusing the speaker,LED word displays as well as strobe lights to transfer in-tent and corresponding experiments showed pedestriansâĂŹtrust to AVs was greatly improved(Matthews et al., 2017).Furthermore, based on human-robot interaction, three novelmethods were proposed by Nicole et al(Mirnig et al., 2017).
Ke Wang et al.:
Preprint submitted to Elsevier
Page 13 of 24ccident analysis and preventtion
First, Gaze and gestures of conventional drivers are loggedand projected on wild shield screens to interact with pedes-trians. Second, elements in the front of cars, such as head-lights, radiator grill as well as side mirrors, were also utilizedto make gestures like drivers to inform pedestrians. Third, arobot driver was utilized to imitate the eye contact and ges-tures of drivers.In experiments of Rahimian et al.(Rahimian et al., 2018),pedestrians could receive an alarm auditory from the AVswhen they are crossing the road, alerting not using phonesand paying attention to driving environment. However, phonesare not always terrible during the interaction process. Ahmedet al.(Hussein et al., 2016) applied phones to interact withpedestrians. Based on GPS data, phone application and AVscould calculate the distance and time to the collision pointas well as danger indexes, therefore, the application couldwarn pedestrians and AVs by displaying a warning messageor vibrating. Additionally, He et al.(He et al., 2017) pro-posed a V2P model which utilized WIFI, Bluetooth as wellas DSRC communication technology to build interaction be-tween AVs and pedestrians. Last but not the least, it will beessential to transfer information to pedestrians under the col-lapse of AVs, so Aaron et al.(Steinfeld et al., 2018) utilizedthe words of iPad and LED strips to send faulty informationto pedestrians.Through these external interfaces, pedestrians argued theyknow more about the state of AVs in (Deb et al., 2018b; Flo-rentine et al., 2016; Jayaraman et al., 2018; LagstrÃűm andLundgren, 2015b; Mahadevan et al., 2018; Matthews et al.,2017; Reig et al., 2018a) and thereafter they trust AVs morein (Creech et al., 2017; Deb et al., 2017, 2018b; Florentineet al., 2016; Hussein et al., 2016; Jayaraman et al., 2018;Matthews et al., 2017). To show the process of interaction,the features and transferring information are partially plottedin Fig. 9.
Nevertheless, every interface has its inevitable shortages.Visual interfaces are the most commonly used interfaces butit is not useful for people with color blindness, visual im-pairment and distraction. Auditory interfaces are thought tobe more like a command and hence pedestrians dislike it. Inaddition, pedestrians would be confused when there are lotsof AVs and other sounds. Phone vibration isnâĂŹt preferredby pedestrians for there are other functions that would causevibration in a phone (Mahadevan et al., 2018).To tackle the shortages, a fusion of multiple features is apromising method to transfer more information and reachhigher robustness(Mahadevan et al., 2018). In KarthikâĂŹsexperiments, mixed interfaces got the best score. However,it doesnâĂŹt mean the more interfaces mixed, the better in-teraction it gets. However, information overload takes placewhen there are too many interfaces(Mahadevan et al., 2018).In the situations, pedestrians tend to check all interafces andthen allow a go-head, thus leading to inefficiency. There-fore, researchers should consider possible information over- load when fusing different interfaces. Furthermore, interac-tion needs a standard language to decide what interfaces touse and how to use. At above, different interfaces and differ-ent infomration transferrring methods are adopted by variousresearchers. Considering the future with AVs, pedestrianswould get confused when interacting with AVs of differentbrands(Habibovic et al., 2019). Therefore, a standard inter-action language is extremely needed for better welcomingAVs.
In conclusion, equipped with external features, AVs couldtransfer the following state: about to start, about to stop,slowing down, starting AV mode as well as observing thepedestrians, etc. Furthermore, phone utilization would bewarned and pedestrians could get the failure state of AVs.Through these, pedestrians could know the driving state ofAVs and hence decide what and how to do. Compared totraditional interaction methods, the external features of AVsseem to be more abundant and intuitive. Yet, there still exitssome shortages in these researches and more efforts shouldbe taken in the area of making a uniform interacting languageand fusing multiple external interfaces. Above all, we con-clude driverless technology is OK on interacting with pedes-trians but still has room to improve. Moreover, we boldlypredict that AVs could even reach a better interaction com-pared to traditional cars by accomplishing the current limi-tations.
Receptivity was originally defined as the willingness toaccept uncertain, unfamiliar, or paradoxical idea(Smith, 1988).Therefore, the receptivity of pedestrians to AVs is the will-ingness of pedestrians to accept AVs, similar to the conceptof trust and acceptance.The high receptivity of pedestrians is important. On theone hand, as Vahidi concluded, the main obstacle to achievea place in the market is not only technical issues but alsothe lack of acceptance of new ideas, which is important togradually push AVs into markets(Vahidi and Eskandarian,2003). On the other hand, pedestrians would be more willingto interact with AVs and push the technology to evolve.
Recent researches have paid attention to factors that in-fluence receptivity, which is summarized in Table 12. De-mography (including age & gender), reflects some basic prop-erties of persons and is thought to influence the receptivity.Shuchisnigdha found males and younger people tend to trustAVs due to more interest in new technologies(Deb et al.,2017). However, Reig thought demographic variables werenot meaningfully related to beliefs and perceptions(Reig et al.,2018b).Moreover, understanding to AVs has a strong relationshipwith receptivity. Samantha et al.(Reig et al., 2018b) foundan insufficient understanding of AVs would lead to mistrust
Ke Wang et al.:
Preprint submitted to Elsevier
Page 14 of 24ccident analysis and preventtion and made-up explanations for the behaviors of AVs. Monikaet al.(Hengstler et al., 2016) showed that peopleâĂŹs under-standing of driverless algorithms improves receptivity andpedestrians now know a little how AVs conduct their move-ments
Ke Wang et al.:
Preprint submitted to Elsevier
Page 15 of 24ccident analysis and preventtion T a b l e : I n t e r ac ti ng m e t hod s o f d r i v e r l e ss t ec hno l og i e s t o c o mm un i ca t e w it hp e d e s t r i a n s R e f e r e n ce E xp e r i m e n t a l t yp e D a t e I n t e rf ace s I n f o r m a ti on C on c l u s i on K a r t h i k e t a l . ( M a h a d e v a n e t a l ., ) E m p i r i ca l e xp e r i m e n t & qu e s ti onn a i r e . S p ea k e r + LE D s t r i p2 . S p ea k e r + LE D li gh t s ( i n s t r ee t ) . A n i m a t e d f ace + phon e h a p ti c ( p e d e s t r i a n s ) . P r i n t e dh a nd + phon ea ud i o + LE D li gh t s ( s t r ee t )( AV s ) a bou tt o s t a r t; a bou t t o s t op ; f u ll y s t op ; no ti ce t h e p e d e s t r i a n s ; ( )I n t e rf ace s h e l pp e d e s t r i a n s a tt e m p tt o c r o ss ; ( )I n t e rf ace s ca n e x i s ti n t h e e nv i r on m e n t; ( ) AV ss hou l du s ea c o m b i n a ti ono f v i s u a l , a ud it o r y , a nd phy s i ca li n t e rf ace s . M il ec i ae t a l . ( M a tt h e w s e t a l ., ) E m p i r i ca l e xp e r i m e n t & s i m u l a ti ng e xp e r i m e n t s LE Dw o r dd i s p l a y + s p ea k e r s + s t r ob e li gh t s ( P e d e s t r i a n s ) : c r o ss no w ; s t op ; w a itt o c r o ss ( ) H u m a n s r eac t po s iti v e l y a nd m o r e p r e d i c t a b l y w h e n t h e i n t e n t o f t h e v e h i c l e i s c o mm un i ca t e d ( ) P e d e s t r i a n s t r u s t AV s m o r e w h e n t h e r ea r e i n t e rf ace s o r t h e y h a v e p r i o r kno w l e dg e t o AV s A a r on e t a l . ( S t e i n f e l d e t a l ., ) E m p i r i ca l e xp e r i m e n t & s i m u l a ti ng e xp e r i m e n t s & i n t e r v i e w LE D s t r ob e li gh t s + i P a d d i s p l a y ( P e d e s t r i a n s ) : p l ea s e w a it; s a f e t o c r o ss ( ) T h e r ee x i s t s po ss i b l ec on f u s i ono f i n t e rf ace s ( ) P e d e s t r i a n s w a n tt okno w t h e s t a t e o f AV s r a t h e r t h a n w h a tt h e y s hou l ddo N i c o l ee t a l . ( M i r n i g e t a l ., ) N / A . W i nd s h i e l d s c r ee n2 . H ea d li gh t s , r a d i a t o r g r ill , t h e s i d e m i rr o r s . R obo t d r i v e r H u m a n - li k e g azea nd g e s t u r e s N / A T ob i a s e t a l . ( L a g s t r à ű m a nd L undg r e n , a ) A W i za r do f O za pp r o ac h2015 LE D li gh t s t r i p I n AV m od e ; a bou tt oy i e l d ; a bou tt o s t a r t;i s r e s ti ng I n t e rf ace s ca np r o m o t e t h e i n t e r ac ti on b e t w ee np e d e s t r i a n s a ndv e h i c l e s . A h m e d e t a l . ( H u ss e i n e t a l ., ) U s e r s t udy & s i m u l a ti ng e xp e r i m e n t s S m a r t phon e s T i m e t o c o lli s i on ; v e l o c it y ; d i s t a n ce t o c o lli s i on e t c . E x t e r n a li n t e rf ace s c r ea t e good p e rf o r m a n ce h i ghd e t ec ti on r a t ea ndu s e r s a ti s f ac ti on U r m s on e t a l . ( U r m s on e t a l ., ) N / A ( p a t e n t ) A n e l ec t r on i c s i gno r li gh t s , a s p ea k e r W h a t AV s a r e do i ngo r go i ng t odo T h e i n t e rf ace s c ou l d r e p l ace t h e i n t e r ac ti onb e t w ee np e d e s t r i a n s a nd d r i v e r s Ke Wang et al.:
Preprint submitted to Elsevier
Page 16 of 24ccident analysis and preventtion P ooy ae t a l . ( R a h i m i a n e t a l ., ) S i m u l a ti ng e xp e r i m e n t s S m a r t phon e s W a r np e d e s t r i a n s t h a t " T a k e ca r ea r ound a nd t h e r ea r e AV s " I n f o r m i ngp e d e s t r i a n s byphon e s c ou l d i m p r ov e t h e s a f e t y H ee t a l . ( H ee t a l ., ) S i m u l a ti ng e xp e r i m e n t & ca s e s t udy2016 S m a r t phon e s T h e l o ca ti ono f AV s a nd p e d e s t r i a n s B l u e t oo t h t ec hno l ogy c o m b i n e d w it h D S RC c ou l db e w o r k a b l e f o r ac ti v e p e d e s t r i a np r o t ec ti on . E v e l yn e t a l . ( F l o - r e n ti n e e t a l ., ) E m p i r i ca l e xp e r i m e n t S p ea k e r + LE D d i s p l a y1 . W h e t h e r p e d e s t r i a n s â Ă Ź p r e s e n ce h a s b ee np e r ce i v e d by t h ea u t ono m ou s v e h i c l e . B r o a d ca s tt h e d e t ec ti on r e s u lt o f L i DA R T h ea ud i o c u e s a nd LE D s t r i p s a r e u s e f u l f o r i n t e r ac ti on a ndp r o m o t e t r u s t . S hu c h i s n i gdh a e t a l . ( D e b e t a l ., ) V R e xp e r i m e n t S p ea k e r + LE D d i s p l a y1 . AV s a r e b r a k i ng2 . P e d e s t r i a n s a r e s a f e t o c r o ss ( ) . F a m ili a r s i gn f o r p e d e s t r i a n , c l ea r t e x t a nd c l ea r v e r b a l m e ss a g e i s p r e f e rr e d ( ) . T h e i n c l u s i on i n t e rf ace s i n c r ea s e p e d e s t r i a n s â Ă Ź r ece p ti v it y A z r ae t a l . ( H a b i - bov i c e t a l ., ) N / A LE D li gh t s t r i p s I n AV o r m a nu a l m od e ; a bou tt o s t a r t; a bou tt oy i e l d E x t e r n a li n t e rf ace s h e l p t o a ddup t r u s tt o AV s ; F ea t u r e s ca ll f o r s t a nd a r d i za ti on ; Ke Wang et al.:
Preprint submitted to Elsevier
Page 17 of 24ccident analysis and preventtion
Figure 9:
External interfaces and transferring information for interaction in driverlesstechnology agreeing with previous research of Rogers(Rogers, 2003).Furthermore, Monika et al. emphasized the usability as wellas trialability would take pedestrians closer to AVs. The ex-periments of Samantha et al.(Reig et al., 2018b)found themore interest pedestrians have, the higher level pedestrianswould trust AVs. Actually, usability, trialability and interestimprove the receptivity by promoting the understanding ofpedestrians.The interaction between AVs and pedestrians also promotesthe receptivity(Habibovic et al., 2019). In the context ofAVs, interaction is conducted based on external interfaces.Samantha et al. (Reig et al., 2018b)emphasized the impor-tance of external interfaces to promote receptivity due totransferring the intent as well as the current state of AVs.Saleh et al. highlighted the intent communication and calledit the most critical sign to win the trust(Rasouli and Tsotsos,2019b). Ahmed et al.(Hussein et al., 2016)adopted a V2Papplication to warn persons and found a promotion in trusttoward AVs. Azrac et al. discovered that after interactingwith AVs once, pedestrians could gain a close understandingof AVs, hence promoting trust and acceptance. Similarly, pedestrians trust AVs more in (Deb et al., 2018b; Florentineet al., 2016; Matthews et al., 2017) after interacting with ex-ternal interfaces.Interestingly, the brands of companies influence the recep-tivity to AVs. Samantha et al.(Reig et al., 2018b) discoveredpedestrians trust the AVs of Uber because of its largeness,engineering quality while others doubt Uber for its fast speedto develop AVs .Accidents related to AVs harm the receptivity because nega-tive events are more visible compared to positive events(Leeand See, 2004). Small events, even totally no responsibil-ity from AVs, could set off a number of resistances(Slovic,1993). Actually, most of the accidents should not blame AVsas FavarÚ et al. showed in (Favaro et al., 2017) that con-tributing factors come more from other the manoeuvers ofother vehicles.
Above all, demography, understanding, the ability of AVs,interaction, the brands of companies and accidents would in-
Ke Wang et al.:
Preprint submitted to Elsevier
Page 18 of 24ccident analysis and preventtion
Table 12
Influencing factors of pedestriansâĂŹ receptivity to AVsTitle Reference Method Year FactorsA Field Study of Pedestrians andAutonomous Vehicles (Reig et al.,2018b) Questionnaire 2018 Understanding& Ability &Brand&DemographyApplied artificial intelligence andtrustâĂŤThe case of autonomous vehiclesand medical assistance devices (Hengstleret al., 2016) Case study 2016 UnderstandingTrust in AV: An Uncertainty ReductionModel of AV-Pedestrian Interactions (Jayaramanet al., 2018) VR 2018 AbilityTowards Trusted Autonomous Vehicles fromVulnerable Road Users Perspective (Saleh et al.,2017b) Model (trustmodel) 2017 AbilityDevelopment and validation of aquestionnaire to assess pedestrianreceptivity toward fully autonomous vehicles (Deb et al.,2017) Questionnaire 2017 Interaction &DemographyP2V and V2P Communication forPedestrian Warning on the basis ofAutonomous Vehicles (Husseinet al., 2016) Model(application) 2016 InteractionExternal Vehicle Interfaces forCommunication with Other Road Users? (Habibovicet al., 2019) Model(externalinterfaces) 2018 InteractionIntent Communication betweenAutonomous Vehicles and Pedestrians (Matthewset al., 2017) Model(externalinterfaces) 2017 InteractionPedestrian Notification Methods inAutonomous Vehicles for Multi-ClassMobility-on-Demand Service (Florentineet al., 2016) Model(externalinterfaces) 2018 InteractionInvestigating pedestrian suggestions forexternal interfaces on fully autonomousvehicles: A virtual reality experiment (Deb et al.,2018b) VR 2018 Interaction fluence the receptivity of AVs. By analyzing these factors,interaction has taken up largely in the influencing factors,which emphasize the importance of interaction and interact-ing experiences again. Additionally, the understanding andthe ability also contribute to the receptivity. More impor-tantly, the ability of AVs, knowledge and interaction couldpromote each other and the commonality of these factors isto stress the real experiences with AVs. It is known that realexperiences occur in close contact with pedestrians. How-ever, we seldom see researches on Chinese pedestrians; wehardly see interacting experiment is carried out in China; wenever see waymo (google) or cruise (general motor) test theirAVs in the Chinese environment. Therefore, the receptivityof Chinese pedestrians to AVs must be very low. In conclu-sion, the receptivity of Chinese pedestrians is bad and notadaptive to China.
According to analyses above, we summarize the adapt-ability of AVs to pedestrians in Table 13. The interactionfeatures could transfer the state of AVs to pedestrians butneed a standard language to manage external interfaces. Driver-less technologies could detect normal pedestrians well. More-over, in small and distracted pedestrian detection, the preci-sion is OK and there is room to promote. The worse scenesare pedestrians with heavy occlusion and detectors have very bad performance in the situation. Additionally, the influenc-ing factors of receptivity reflects the importance of real ex-periences with AVs but little is conducted in China. Hence,the receptivity of Chinese pedestrians is so bad.
AVs would greatly change the traditional situation of pedes-trian traffic and take a change to pedestrian mentality andregulations. In the scholar research, There are some arti-cles talking about and the mentality of drivers in AVs andlegal obstacles of pushing AVs into mass production, suchas issue of compliance, issues of liability, issues of informa-tion governance(National Highway Traffic Safety Adminis-tration, 2016; Gurney, 2015; Lin, 2016; Smith, 2016; Wu,2019). However, there is a lack of research on the change ofpedestrian mentality and government regulations to pedes-trians when AVs enter the road. Therefore, based on the ana-lyzed behaviors of pedestrians and adaptability analysse, theAVsâĂŹ influence on pedestrian mentality and governmentregulations would be discussed in the following.Firstly, the mentality of pedestrians to conduct these bad be-haviors would change. Currently, AVs are troubled by oc-cluded pedestrian detection, which is to say, AVs would notpass the road when there are a few pedestrians. Furthermore,
Ke Wang et al.:
Preprint submitted to Elsevier
Page 19 of 24ccident analysis and preventtion
Table 13
The adaptability of AVs to pedestrians in urban ChinaTechnical demands AVs adapt to pedestrians: detection Pedestrians adpat to AVsNormalpedestri-ans Occludedpedestrians Smallpedestri-ans Distratedpedestrians Interaction(HMI) ReceptivityEvaluation Excellent Bad OK OK OK BadLimitations andsuggestions N/A Greatlyimproveprecision inheavy occlu-sion(lowestmissrate:49%) Improvedetectionpreci-sion(lowestmissrate:23%) Moreresearchesabout phonedistractiondetection;moreresearchesfocusing onotherdistractionfactors Call for standard HMIdesign language;avoidinformation overloadof HMI; fuse differentinterface modalities Totally lackreceptivityresearch in China;conduct driverlessexperimentsinvolvingpedestrians the human-machine interface would accurately transfer thisstate to pedestrians that AVs would stop. In the situation,pedestrians would get road rights with no fear of accidentsand learn experience from it, which could cause a more se-rious red light running take place. Similarly, jaywalkingwould be safer for pedestrians because they could find theAVs are about to stop directly compared to gaining infor-mation from traditional drivers. As for distraction, pedestri-ans would have to pay attention to AVs because they are notfamiliar with this new technology and the human-machineinterface should read by them themselves, which possiblyreduces the distracted pedestrians. In summary, pedestrianswould pay more attention to the traffic but turn bold to vio-late rules.Under the context of AVs, pedestrian regulations should beadjusted accordingly. We conclude that pedestrians could bebolder to violate the rules above. In the situation, pedestri-ans would be safe but AVs would lose their ability to maketraffic efficiency. In most of existing regulations, violatedpedestrians are hard to caught and the punishment is light,which could not prevent bad behaviors in an efficient way.Therefore, the government must take stronger and more effi-cient measures to limit pedestrians violate the rules; for ex-ample, record illegal records into personal credit reports orbuild pedestrian bridges. Only by strong or efficient mea-sures could reduce pedestrian violations and let AVs bringtraffic more efficiency.
5. Challenges & Opportunities
Based on the above analyses, we summarize some chal-lenging but promising researches in the Chinese pedestrianenvironment.
As discussed above, a number of researches emerge onthe external interfaces to interact AVs with pedestrians butthere is not a common language. What information to trans-fer and how to transfer depends on researchers. We could imagine it will be a mess if different AVs use different inter-acting ways to inform pedestrians, causing a possible mis-understanding. Therefore, we call for a standard interact-ing language with two parts. Firstly, the kind of transfer-ring information should be standardized. All AVs only trans-fer specific kinds of information, such as startup, observingthe pedestrians, etc. Secondly, the information transferringmethods should be made standard. For example, AVs wouldonly indicate they are starting up by lighting red lights. Un-der the system, we believe pedestrians would have a betterunderstanding of driverless technologies and push AVs tobecome real traffic participants in the long run.
Single modality interface, such as LED strips, might of-fer limited interaction information and have some shortagesin its nature. Fusing interfaces of different modalities wouldbe a promising solution, which would promote the preci-sion and robustness of interaction. For example, pedestrianscould observe all the interfaces and decide what to do. Es-pecially, pedestrians could still judge the state of AVs whenpart of interfaces collapse, which is essential to protect pedes-trian safety in the future with AVs. It is suggested that re-searchers should combine the advantages of different inter-faces and eliminate the shortages by adding other interfaces.
Interfaces fusion of HMI design has been a trend, how-ever, overfull interfaces would lead to interaction informa-tion overload. In the situation, pedestrians tend to checkall interafces and then decide what to do. This would con-fuse pedestrians and cause an inefficiency in traffic becausepedestrians need much time to observe interfaces’ state. There-fore, information overload on HMI design is suggested toavoid and the reriewed articles showed three interfaces areenough.
Ke Wang et al.:
Preprint submitted to Elsevier
Page 20 of 24ccident analysis and preventtion
Small pedestrian detection is important to protect safetyaround the world. In the Chinese driving environment, thereare a number of small pedestrians due to jaywalking. How-ever, the optimal detector only has a 23% miss rate of smallpedestrian detection, which is far from practical use. Wesuggest more efforts should be taken to improve the perfor-mance.
Occlusion often happens when pedestrians are crossingthe crosswalk and is a challenging problem in the Chinesedriving environment. Nevertheless, the detector performspoorly in heavily occluded pedestrians. Hence, we call onefforts to tackle the issues of heavily occluded pedestrian de-tection.
We have talked about the distracted pedestrian detec-tion above and focus on smartphone usage. However, thereis only one paper implemented to detect pedestrians usingphones. Further research is required to push phone utiliza-tion detection. Moreover, other distracted reasons also callfor researches in the context of AVs. Therefore, we suggestmore distracted pedestrian research.
In the part of receptivity, we concluded factors that in-fluence the receptivity of pedestrians, such as demography,knowledge and interaction, which highlight the real expe-riences between AVs and pedestrians. Therefore, the re-searches based on American would be not adaptive to othercountries. However, as the largest market of cars, few re-searches are conducted on the receptivity in China. There-fore, we suggest nation-based receptivity research, especiallyin China, to get close knowledge to develop AVs.
We conclude that real interacting experience of pedes-trians greatly promote the trust to AVs. However, few re-searches are implemented involving pedestrians. To achievereceptivity, why not invite more pedestrians to join the em-pirical experiments? By doing so, not only do researchersgain practical feedback but also pedestrians could interactwith AVs. In the process, pedestrians know more knowl-edge about AVs and hence promote the receptivity. Overall,it is wise for researchers to conduct experiments with pedes-trians.
6. Conclusions
The objective of this paper is to survey the adaptability ofautonomous vehicles to pedestrians in urban China. Chinahas a complicated pedestrian environment, however, little isknown about the adaptability of future driverless technology.To this end, we analyzed three typical pedestrian behaviorsin urban China and summed up the key technical demands for AVs. Then by reviewing the latest driverless technolo-gies, we conclude the adaptability of autonomous vehicles topedestrians respectively. Finally, we summarized the chal-lenging problems and opportunities in Chinese pedestrianenvironment.As we talked above, the adaptability of autonomous vehi-cles depends on the adaptability of interaction, detection aswell as receptivity. In conclusion, driverless technologiesperform well in normal pedestrian detection. Additionally,driverless technologies have an OK work in small pedes-trian detection, distracted pedestrian detection and interac-tion. However, occluded pedestrian detection and the recep-tivity of pedestrians are not adaptive to China. We notedsome challenging but promising areas for Chinese pedes-trian environment, such as standard interaction languagesand nation-based receptivity research. These aspects couldbe our future research or conducted by other practitioners.
Acknowledgment
This research was funded by National Natural ScienceFoundation of China (Grant No. 51605054), State Key Lab-oratory of Vehicle NVH and Safety Technology (NVHSKL-202008 and NVHSKL-202010), The Science and Technol-ogy Research Program of Chongqing Education Commis-sion of China (KJQN201800517 and KJQN201800107), Fun-damental Research Funds for the Central Universities (No:2019CDXYQC003), Chongqing Social Science Planning Project(No:2018QNJJ16), Key Technical Innovation Projects of ChongqingArtificial Intelligent Technology (cstc2017rgzn-zdyfX0039)
References
Baswail, A., Allinson, L., Goddard, P., Pfeffer, K., 2019. AdolescentsâĂŹmobile phone use while crossing the road. Safety 5, 27.Brazil, G., Yin, X., Liu, X., 2017. Illuminating pedestrians via simultaneousdetection & segmentation, in: Proceedings of the IEEE InternationalConference on Computer Vision, pp. 4950–4959.Bureau of Infrastructure, Transport, Regional Development and Commu-nications of Austrilia, 2018. Road trauma Australia 2018 statisticalsummary. URL: .Cai, Z., Fan, Q., Feris, R.S., Vasconcelos, N., 2016. A unified multi-scaledeep convolutional neural network for fast object detection, in: europeanconference on computer vision, Springer. pp. 354–370.Cao, K., Chen, Y., Stuart, D., Yue, D., 2015. Cyber-physical model-ing and control of crowd of pedestrians: a review and new framework.IEEE/CAA Journal of Automatica Sinica 2, 334–344.Charisi, V., Habibovic, A., Andersson, J., Li, J., Evers, V., 2017. Children’sviews on identification and intention communication of self-driving ve-hicles, in: Proceedings of the 2017 Conference on Interaction Designand Children, ACM. pp. 399–404.Chen, J.L., Wang, K., Xiong, Z.B., 2018. Collision probability predictionalgorithm for cooperative overtaking based on ttc and conflict proba-bility estimation method. International Journal of Vehicle Design 77,195–210. URL:
Ke Wang et al.:
Preprint submitted to Elsevier
Page 21 of 24ccident analysis and preventtion
CunBao, Z., Feng, C., Yuanyuan, W., Hualong, Z., 2018. Impacts onstreet crossing behaviors of pedestrians using phones in unsignal sec-tion(in chinese). Transportation System Engineering and Informationv.18, 140–145.Dalal, N., Triggs, B., 2005. Histograms of oriented gradients for human de-tection, in: 2005 IEEE computer society conference on computer visionand pattern recognition (CVPR’05), IEEE. pp. 886–893.Deb, S., Rahman, M.M., Strawderman, L.J., Garrison, T.M., 2018a. Pedes-triansâĂŹ receptivity toward fully automated vehicles: Research reviewand roadmap for future research. IEEE Transactions on Human-MachineSystems 48, 279–290.Deb, S., Strawderman, L., Carruth, D.W., DuBien, J., Smith, B., Garrison,T.M., 2017. Development and validation of a questionnaire to assesspedestrian receptivity toward fully autonomous vehicles. TransportationResearch Part C: Emerging Technologies 84, 178–195. doi: .Deb, S., Strawderman, L.J., Carruth, D.W., 2018b. Investigating pedestriansuggestions for external features on fully autonomous vehicles: A virtualreality experiment. Transportation Research Part F: Traffic Psychologyand Behaviour 59, 135–149. doi: .Department for Transport of the UK, 2019. Reported road casualtiesGreat Britain, annual report: 2018. URL: https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/834585/reported-road-casualties-annual-report-2018.pdf .Dollár, P., Wojek, C., Schiele, B., Perona, P., 2009. Pedestrian detection: Abenchmark, in: 2009 IEEE Conference on Computer Vision and PatternRecognition, IEEE. pp. 304–311.Dollar, P., Wojek, C., Schiele, B., Perona, P., 2011. Pedestrian detection:An evaluation of the state of the art. IEEE transactions on pattern anal-ysis and machine intelligence 34, 743–761.Dommes, A., GraniÃľ, M.A., Cloutier, M.S., Coquelet, C., Huguenin-Richard, F., 2015. Red light violations by adult pedestrians and othersafety-related behaviors at signalized crosswalks. Accident Analysis &Prevention 80, 67–75.Du, X., El-Khamy, M., Lee, J., Davis, L., 2017. Fused dnn: A deep neuralnetwork fusion approach to fast and robust pedestrian detection, in: 2017IEEE winter conference on applications of computer vision (WACV),IEEE. pp. 953–961.Du, X., El-Khamy, M., Morariu, V.I., Lee, J., Davis, L., 2018. Fuseddeep neural networks for efficient pedestrian detection. arXiv preprintarXiv:.08688 .Ess, A., Leibe, B., Van Gool, L., 2007. Depth and appearance for mobilescene analysis, in: 2007 IEEE 11th International Conference on Com-puter Vision, IEEE. pp. 1–8.Fang, Z., LÃşpez, A.M., 2019. Intention recognition of pedestrians andcyclists by 2d pose estimation. IEEE Transactions on Intelligent Trans-portation Systems .Favaro, F.M., Nader, N., Eurich, S.O., Tripp, M., Varadaraju, N., 2017.Examining accident reports involving autonomous vehicles in california.PLoS One 12, e0184952. doi: .Feiyang, L., 2017. Researches on behavior features and controlling methodsof pedestrians using phones when crossing the road(In Chinese). Thesis.Hefei University of Technology.Florentine, E., Ang, M.A., Pendleton, S.D., Andersen, H., Ang Jr, M.H.,2016. Pedestrian notification methods in autonomous vehicles for multi-class mobility-on-demand service, in: Proceedings of the Fourth Inter-national Conference on Human Agent Interaction, pp. 387–392.Geiger, A., Lenz, P., Urtasun, R., 2012. Are we ready for autonomousdriving? the kitti vision benchmark suite, in: 2012 IEEE Conference onComputer Vision and Pattern Recognition, IEEE. pp. 3354–3361.GerÃşnimo, D., Sappa, A.D., LÃşpez, A., Ponsa, D., 2007. Adaptive imagesampling and windows classification for on-board pedestrian detection,in: International Conference on Computer Vision Systems: Proceed-ings.Gurney, J.K., 2015. Crashing into the unknown: An examination of crash-optimization algorithms through the two lanes of ethics and law. Alb. L.Rev. 79, 183.Habibovic, A., Andersson, J., Lundgren, V.M., KlingegÃěrd, M., Englund, C., Larsson, S., 2019. External vehicle interfaces for communica-tion with other road users?, in: Automated Vehicles Symposium 2018,Springer. pp. 91–102.He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for imagerecognition, in: Proceedings of the IEEE conference on computer visionand pattern recognition, pp. 770–778.He, S., Li, J., Qiu, T.Z., 2017. Vehicle-to-pedestrian communication mod-eling and collision avoiding method in connected vehicle environment.Transportation Research Record 2621, 21–30.Hengstler, c., Enkel, E., Duelli, S., 2016. Applied artificial intelligenceand trustâĂŤthe case of autonomous vehicles and medical assistancedevices. Technological Forecasting and Social Change 105, 105–120.doi: .Hurney, P., Waldron, P., Morgan, F., Jones, E., Glavin, M., 2015. Review ofpedestrian detection techniques in automotive far-infrared video. IEEETransactions on Intelligent Transportation Systems 9, 824–832.Hussein, A., GarcÃŋa, F., Armingol, J.M., Olaverri-Monreal, C., 2016.P2v and v2p communication for pedestrian warning on the basis of au-tonomous vehicles, in: 2016 IEEE 19th International Conference on In-telligent Transportation Systems (ITSC), IEEE. pp. 2034–2039.Jayaraman, S.K., Creech, C., Robert Jr, L.P., Tilbury, D.M., Yang, X.J.,Pradhan, A.K., Tsui, K.M., 2018. Trust in av: An uncertainty reduc-tion model of av-pedestrian interactions, in: Companion of the 2018ACM/IEEE International Conference on Human-Robot Interaction, pp.133–134.Jinhua, L., 2017. Research on Impact of Pedestrian Crossing Using MobilePhones on Traffic Safety Based on Logistic Model (In Chinese). Thesis.Chongqing University.Johannsen, G., 2009. Human-machine interaction. Control Systems,Robotics and Automation 21, 132–62.Jordan, W., 2013. Green man gets the red light. URL: https://yougov.co.uk/topics/politics/articles-reports/2013/09/06/green-man-gets-red-light .Jun, Q., Zhiquan, J., Danfeng, Y., 2012. Analysis of pedestrian road trafficinjury in chongqing (in cinese). Chinese Journal of Trauma 28, 24–27.Kedeng, H., Xiaofei, H., Qiwei, C., Ye, W., 2015. Mathematical model forchinese style of crossing management (in chinese). Journal of Yulin Nor-mal University 02, 26–32+36. URL: .LagstrÃűm, T., Lundgren, V.M., 2015a. AVIPâĂŞAutonomous Vehi-cles Interaction with Pedestrians: An Investigation of Pedestrian-DriverCommunication and Development of a Vehicle External Interface. The-sis. Master thesis, Chalmers University of Technology.LagstrÃűm, T., Lundgren, V.M., 2015b. An investigation of pedestrian-driver communication and development of a vehicle external interface.Human Factors 84.Lai, C.Q., Teoh, S.S., 2014. A review on pedestrian detection techniquesbased on histogram of oriented gradient feature, in: 2014 IEEE StudentConference on Research and Development, IEEE. pp. 1–6.Lee, J., Abdel-Aty, M., Shah, I., 2019a. Evaluation of surrogate measuresfor pedestrian trips at intersections and crash modeling. Accident Anal-ysis & Prevention 130, 91–98.Lee, J., Abdel-Aty, M., Xu, P., Gong, Y., 2019b. Is the safety-in-numberseffect still observed in areas with low pedestrian activities? a case studyof a suburban area in the united states. Accident Analysis & Prevention125, 116–123.Lee, J.D., See, K.A., 2004. Trust in automation: Designing for appropriatereliance. Human Factors 46, 50–80.Li, B., Yao, Q., Wang, K., 2012. A review on vision-based pedestrian de-tection in intelligent transportation systems, in: Proceedings of 20129th IEEE international conference on networking, sensing and control,IEEE. pp. 393–398.Li, G., Yang, Y., Qu, X., 2019. Deep learning approaches on pedestriandetection in hazy weather. IEEE Transactions on Industrial ElectronicsPP, 1–1. URL:
Ke Wang et al.:
Preprint submitted to Elsevier
Page 22 of 24ccident analysis and preventtion https://today.yougov.com/topics/lifestyle/articles-reports/2014/09/25/almost-everyone-jaywalks .Munder, S., Gavrila, D.M., 2006. An experimental study on pedestrianclassification. IEEE transactions on pattern analysis and intelligence,machine 28, 1863–1868.Mwakalonge, J., Siuhi, S., White, J., 2015. Distracted walking: Examiningthe extent to pedestrian safety problems. Journal of traffic and trans-portation engineering (English edition) 2, 327–337.Nasar, J.L., Troyer, D., 2013a. Pedestrian injuries due to mobile phone usein public places. Accident Analysis & Prevention 57, 91–95.Nasar, J.L., Troyer, D., 2013b. Pedestrian injuries due to mobile phone usein public places. Accident Analysis and Prevention 57.National Highway Traffic Safety Administration, 2016. Federal automatedvehicles policy: Accelerating the next revolution in roadway safety. USDepartment of Transportation.National Highway Traffic Safety Administration, 2019. Traffic Safety Facts2017: A Compilation of Motor Vehicle Crash Data. volume 812806.Ouyang, W., Zhou, H., Li, H., Li, Q., Yan, J., Wang, X., 2018. Jointlylearning deep features, deformable parts, occlusion and classification forpedestrian detection. IEEE transactions on pattern analysis and machineintelligence 40, 1874–1887.Papageorgiou, C., Poggio, T., 2000. A trainable system for object detection.International journal of computer vision 38, 15–33.Qi, C.R., Liu, W., Wu, C., Su, H., Guibas, L.J., 2018. Frustum pointnetsfor 3d object detection from rgb-d data, in: Proceedings of the IEEEConference on Computer Vision and Pattern Recognition, pp. 918–927.Qiang, W., Zhaojun, H., Mei, L., 2017. Survey on red light run-ning on the background of chinese style of crossing toward xian (inchinese), in: Decision Forum-Academic Symposium on Scientificand Democratic Decisions. URL: http://cpfd.cnki.com.cn/Article/CPFDTOTAL-KJQY201701002094.htm .Rahimian, P., OâĂŹNeal, E.E., Zhou, S., Plumert, J.M., Kearney, J.K.,2018. Harnessing vehicle-to-pedestrian (v2p) communication technol-ogy: sending traffic warnings to texting pedestrians. Human factors 60,833–843.Rangesh, A., Ohn-Bar, E., Yuen, K., Trivedi, M.M., 2016. Pedestriansand their phones-detecting phone-based activities of pedestrians for au-tonomous vehicles, in: 2016 IEEE 19th International Conference on In-telligent Transportation Systems (ITSC), IEEE. pp. 1882–1887. Rasouli, A., Tsotsos, J.K., 2019a. Autonomous vehicles that interact withpedestrians: A survey of theory and practice. IEEE Transactions onIntelligent Transportation Systems .Rasouli, A., Tsotsos, J.K., 2019b. Autonomous vehicles that interact withpedestrians: A survey of theory and practice. IEEE transactions on in-telligent transportation systems .Reig, S., Norman, S., Morales, C.G., Das, S., Steinfeld, A., Forlizzi, J.,2018a. A field study of pedestrians and autonomous vehicles, in: Pro-ceedings of the 10th International Conference on Automotive User In-terfaces and Interactive Vehicular Applications, ACM. pp. 198–209.Reig, S., Norman, S., Morales, C.G., Das, S., Steinfeld, A., Forlizzi, J.,2018b. A field study of pedestrians and autonomous vehicles, in: Pro-ceedings of the 10th international conference on automotive user inter-faces and interactive vehicular applications, pp. 198–209.Ren, G., Zhou, Z., Wang, W., Zhang, Y., Wang, W., 2011. Cross-ing behaviors of pedestrians at signalized intersections: Observationalstudy and survey in china. Transportation Research Record 2264,65–73. URL: https://journals.sagepub.com/doi/abs/10.3141/2264-08 ,doi: .Ridel, D., Rehder, E., Lauer, M., Stiller, C., Wolf, D., 2018. A literaturereview on the prediction of pedestrian behavior in urban scenarios, in:2018 21st International Conference on Intelligent Transportation Sys-tems (ITSC), IEEE. pp. 3105–3112.Rogers, E.M., 2003. Diffusion of innovations/everett m. rogers. NY: Simonand Schuster 576.SAE On-Road Automated Driving Committee, 2016. SAE J3016. Taxon-omy and Definitions for Terms Related to Driving Automation Systemsfor On-Road Motor Vehicles. Report. SAE International.Saleh, K., Hossny, M., Nahavandi, S., 2017a. Towards trusted autonomousvehicles from vulnerable road users perspective, in: 2017 Annual IEEEInternational Systems Conference (SysCon), IEEE. pp. 1–7.Saleh, K., Hossny, M., Nahavandi, S., 2017b. Towards trusted autonomousvehicles from vulnerable road users perspective, in: 2017 Annual IEEEInternational Systems Conference (SysCon), IEEE. pp. 1–7.Sarker, A., Shen, H., Rahman, M., Chowdhury, M., Dey, K., Li, F., Wang,Y., Narman, H.S., 2019. A review of sensing and communication, hu-man factors, and controller aspects for information-aware connected andautomated vehicles. IEEE Transactions on Intelligent TransportationSystems .Shibo, Z., Lan, L., Pingfei, L., 2018. Study on the characteristics andcauses of pedestrian fatal traffic accidentsâĂŤâĂŤbased on 181 casesof in-depth investigation of accidents (in chinese). Traffic Informa-tion and Safety 36, 22–29. URL: .Simonnet, D., Velastin, S., Turkbeyler, E., Orwell, 2012. Backgroundlessdetection of pedestrians in cluttered conditions based on monocular im-ages: a review. IET Computer Vision 6, 540–550.Slovic, P., 1993. Perceived risk, trust, and democracy. Risk analysis 13,675–682.Smith, B.W., 2016. The trolley and the pinto: Cost-benefit analysis in au-tomated driving and other cyber-physical systems. Tex. A&M L. Rev.4, 197.Smith, J.C., 1988. Steps toward a cognitive-behavioral model of relaxation.Biofeedback and Self-regulation 13, 307–329.Song, T., Sun, L., Xie, D., Sun, H., Pu, S., 2018. Small-scale pedestriandetection based on somatic topology localization and temporal featureaggregation. arXiv preprint arXiv:.01438 .Steinfeld, A., Forlizzi, J., Das, S., Morales, C.G., Norman, S., Reig, S.,2018. Bystander interactions with failing vehicle autonomy .Teknomo, K., Takeyama, Y., Inamura, H., 2016. Review on microscopicpedestrian simulation model. arXiv preprint arXiv:.01808 .The National Bureau of Statistics of China, 2019. National data nationalbureau of statistics of china (in chinese) URL: http://data.stats.gov.cn/english/easyquery.htm?cn=C01 .Urmson, C.P., Mahon, I.J., Dolgov, D.A., Zhu, J., 2015. Pedestrian notifi-cations. US Patent 8,954,252.Vahidi, A., Eskandarian, A., 2003. Research advances in intelligent colli-sion avoidance and adaptive cruise control. IEEE Transactions on Intelli-
Ke Wang et al.:
Preprint submitted to Elsevier
Page 23 of 24ccident analysis and preventtion gent Transportation Systems 4, 143–153. doi: .Volz, B., Mielenz, H., Gilitschenski, I., Siegwart, R., Nieto, J., 2018. In-ferring pedestrian motions at urban crosswalks. IEEE Transactions onIntelligent Transportation Systems 20, 544–555.Wang, K., Huang, X., Chen, J.L., Cao, C., Xiong, Z.B., Chen, L., 2019.Forward and backward visual fusion approach to motion estimation withhigh robustness and low cost. Remote Sensing 11. URL:
ARTN213910.3390/rs11182139 .Wang, K., Huang, Z., Zhong, Z.H., 2014. Simultaneous multi-vehicledetection and tracking framework with pavement constraints basedon machine learning and particle filter algorithm. Chinese Journalof Mechanical Engineering 27, 1169–1177. URL:
Ke Wang et al.: