V2X System Architecture Utilizing Hybrid Gaussian Process-based Model Structures
Hossein Nourkhiz Mahjoub, Behrad Toghi, S M Osman Gani, Yaser P. Fallah
aa r X i v : . [ ee ss . SP ] M a r V2X System Architecture Utilizing HybridGaussian Process-based Model Structures
Hossein Nourkhiz Mahjoub ∗ , Behrad Toghi ∗ , S M Osman Gani ∗ , Yaser P. Fallah ∗∗ Networked Systems LabDepartment of Electrical and Computer Engineering, University of Central Florida, Orlando, FL { hnmahjoub, toghi, smosman.gani } @knights.ucf.edu, [email protected] Abstract —Scalable communication is of utmost importance forreliable dissemination of time-sensitive information in cooperativevehicular ad-hoc networks (VANETs), which is, in turn, anessential prerequisite for the proper operation of the criticalcooperative safety applications. The model-based communication(MBC) is a recently-explored scalability solution proposed in theliterature, which has shown a promising potential to reduce thechannel congestion to a great extent. In this work, based on theMBC notion, a technology-agnostic hybrid model selection pol-icy for Vehicle-to-Everything (V2X) communication is proposedwhich benefits from the characteristics of the non-parametricBayesian inference techniques, specifically Gaussian Processes.The results show the effectiveness of the proposed communicationarchitecture on both reducing the required message exchange rateand increasing the remote agent tracking precision.
Index Terms —Vehicular ad-hoc network, scalable V2Xcommunication, model-based communication, non-parametricBayesian inference, Gaussian processes.
I. I
NTRODUCTION
In 1999, the Federal Communication Commission (FCC)allocated 75 MHz of spectrum at the 5.9 GHz frequencyfor the emerging field of Intelligent Transportation Systems(ITS). Different vehicular communication solutions such asDedicated Short-Range Communication (DSRC) [1], [2], [3]and Cellular Vehicle-to-Everything (C-V2X) [4], [5] have beenintroduced and developed afterwards, aiming at facilitating theestablishment of critical cooperative safety applications, e.g.,Forward Collision Warning/Avoidance (FCW/A) [6], Coopera-tive Adaptive Cruise Control (CACC) [7], [8], and IntersectionManagement.The fundamental role of the V2X communications isenabling every vehicle in a Vehicular Ad-hoc NETwork(VANET) to frequently inform the surrounding nodes about itsmost recent dynamic states. In general, the V2X architecturecould be broken down into three main categories, i.e., com-munication among vehicles (V2V), communication betweenvehicles and infrastructure (V2I), and communication betweenvehicles and Vulnerable Road Users (VRUs), e.g. V2P [9].The concept of information sharing among nodes resultsin a level of situational awareness for any vehicle/VRU andmakes it aware of its surrounding environment, which is supercrucial for the cooperative safety applications to function
This material is based on work supported in part by the National ScienceFoundation under CAREER Grant 1664968 and in part by the Qatar NationalResearch Fund Project NPRP 8-1531-2-651. properly. The Society of Automotive Engineers (SAE), as themain vehicular regulatory organization in US, has proposeda specific framework through a set of standards in orderto realize the notion of situational awareness in vehicularnetworks. The content of the Basic Safety Message (BSM),which conveys the situational awareness information, has beenspecified by SAE J2735 standard [10]. However, vehicular net-works can potentially experience very dense scenarios whichresult in a congested communication channel and imposesevere performance degradation to the network. Therefore,part of another standard by SAE, i.e. SAE system requirementstandard or SAE J2945/1 [11], explores different congestioncontrol mechanisms such as BSM transmission power andrate control in order to manage the generated load from theinformation beaconing and mitigate the congestion imposedon the communication channel. It is noteworthy that the con-gestion control algorithms defined by SAE J2945/1 standarddo not impose any restrictions on the BSM content or sizesince these parameters are defined through SAE J2735. In thecurrent SAE framework, the message content remains intactfor all broadcast packets and every BSM is filled out withraw information directly captured from CAN-bus or receivedfrom GPS, according to the J2735 dictionary.The congestion control section of the SAE system require-ments standard [11], is the current state-of-the-art congestioncontrol solution accepted by the US vehicular research com-munity as well as US automotive industry. This standard hasbeen developed based on several congestion control algorithmsproposed in the literature, among which one can refer to [12]and [13]. In a nutshell, the rate and power control algorithmsdefined in this standard allow vehicles to broadcast theirmessages at the rate of ∼ The terms “BSM” and “packet” are sometimes used interchangeably inthis paper n comparison with the BSM content structure defined byJ2735 standard. This paper, utilizing non-parametric Bayesianmodeling schemes, proposes a hybrid model structure withinthe MBC framework and integrates it with the congestioncontrol communication policy proposed in J2945/1 standard.The notion of MBC and our proposed communication policywill be explored in more details in the subsequent sections.The rest of this paper is organized as follows. In Section II,an overview of the MBC is presented. Section III is devotedto the system-level architecture design of our proposed model-based communication policy. In this section the proposedhybrid model architecture in addition to the details of ourmodel update policy are thoroughly explained. In section IV,the analysis and evaluation results of the proposed method ispresented before the concluding remarks and future researchdirections stated in Section V.II. M
ODEL -B ASED C OMMUNICATION O VERVIEW
One of the main catalysts behind the MBC frameworkis pursuing a new solution perspective to alleviate the net-work congestion by re-designing the content structure ofthe broadcast messages. As stated earlier, the currently stan-dardized dictionary set stipulates the core content of thebroadcast messages to be directly filled out with the rawvehicle position and dynamic state update data. Therefore, itdoes not explicitly reflect the inherent characteristics of themaneuver in which the vehicle is currently involved. However,considering these conceptual characteristics while a vehiclegenerates its messages could be beneficial for optimizing itsscheduled transmission moments. These characteristics couldbe implicitly utilized to determine the moments at whichthe instant updates are critical and should be transmittedby the maneuvering vehicle, as well as the moments whentransmitting a new packet does not worth. More specifically,in some scenarios, such as abrupt and harsh lane changesor hard brakes an instant update is very critical and highlydemanded for the other vehicles’ safety applications. On theother hand, multiple redundant transmissions by a vehicleare over-occupying the communication channel if the vehicle,for instance, is cruising in a steady state. In the latter case,transmitting consecutive BSMs not only do not provide itsneighbors with any higher degree of situational awareness,but also cause more channel congestion, or equivalently anincrease in the number of collided packets, which in turnresults in the lower level of situational awareness finallyachieved by the neighbouring nodes. From this point of view,the MBC scheme could potentially be capable of improvingthe communication scalability by scheduling the transmissiontimes at more optimized moments, even if its criteria for thisscheduling follows the footsteps of the J2945/1 standard.The transmission rate calculation mechanism in J2945/1 isbasically based on the transmitter estimation of its surroundingnetwork density, in addition to its estimation of the positiontracking accuracy which could be achieved by the informationincluded in its last transmitted BSM. The transmitter keeps track of this tracking precision using an Error-Driven com-munication mechanism. More precisely, at any GPS updateafter each BSM transmission, the transmitter calculates thedifference between the constant-speed coasting of its positionderived from the contents of its last transmitted BSM andits current actual position received via GPS. This differencedefines the position tracking error of the transmitter locationat this time instance for an arbitrary node which has receivedthe latest transmitted packet. Then, transmitter performs acomparison between this error with a predefined thresholdand decides to transmit a new packet if the error exceeds thethreshold. Obviously, this mechanism reduces the transmissionrate compared to the baseline 10 Hz transmission.Now if the transmitted message contains a predictive modelwith high precision for longer prediction time-horizons, pre-dictions made based upon it at the receiver vehicles could lessfrequently reach the same position tracking error thresholddefined in J2975/1 in comparison with the case of constantspeed coasting prediction from raw information received viaJ2735 BSMs. This explanation clarifies the core idea behindthe MBC scheme.The maneuver characteristics, or equivalently driver behav-ioral models, are themselves functions of different factors suchas the driver’s personal driving style, his current mental state,the environmental inputs affecting the driver behavior, e.g.road traffic, other vehicles’ maneuvers, weather condition, etc.Considering these factors and reflecting them into the contentsof the generated packets by any vehicle is the fundamental ideabehind the MBC notion. More specifically, the MBC tries togenerate a mathematical model based on the available noise-free CAN-bus information at the transmitter side which beable to explain and predict the driver actions in the future.Assuming these models give notable higher prediction accu-racy compared to the constant speed coasting scheme, whichis the current default method in the standard, then MBC wouldbe able to avoid several redundant information transmissions.Therefore, the MBC has a two-fold advantage; first it canpotentially shrink the payload size by extracting an abstractrepresentation of the vehicle’s state. In addition, it reducesthe transmission rate by enabling the recipient vehicles topredict their neighbors mobility more accurately in farther timehorizons ahead. The former could be achieved through variousabstraction and dimensionality reduction methods and thelatter would be attained through utilizing different supervisedlearning algorithms. In this work we have explored the lattercase, i.e. the MBC effect on the transmission rate compared tothe raw information communication, while the reduced packetsize effect is part of our future research directions.The initial MBC architecture, illustrated by the author in[14], proposes a stochastic hybrid automata modeling schemeand evaluates its performance on a standard FCW algorithm,known as CAMPLinear [18]. Authors in [15] use hiddenMarkov models (HMMs) to derive an adaptive stochastichybrid system (SHS) in order to capture the non-deterministicnature of driving scenarios. Further enhancements in themodeling approach are presented by authors in [16] and [17]hich include non-parametric Bayesian inference methodssuch as Gaussian processes (GPs) with linear kernels andhierarchical Dirichlet process-hidden Markov models (HDP-HMMs). Results in [14]- [17] demonstrate the significantimprovements in communication rate and tracking accuracymetrics utilizing the MBC approach.Analysis in our previous works in [16] and [17] demonstratethat the highly dynamic and diverse driving behaviors addmore complexity to the modeling process. As an illustration,for the case of a vehicle cruising on a highway, the simplisticconstant speed (CS) model will provide an excellent predictioncapability. On the contrary, if the vehicle is navigating througha Manhattan-grid urban area, the CS model will be totallyobsolete.The above-mentioned phenomena (Figure 1 and Figure 2)gives an intuition of the core idea in this work; we propose ahybrid modeling architecture which switches between different(here two) modeling sub-systems in order to adapt to thevehicle’s dynamic state. Our proposed architecture benefitsfrom a CV modeling sub-system alongside with a GP sub-system with a compound kernel, each of which has shownsignificant prediction performance in specific scenarios. Inaddition, since the change points in the high-level drivingbehaviors on average occur much less frequently comparedto the normal message broadcast rates of the state-of-the-artmethods in the literature, our hybrid-MBC method gives aconspicuous reduction in required communication rate. Thedetails of our proposed architecture is presented in the nextsection.III. H
YBRID
GP-
BASED
MBC A
RCHITECTURE
This section provides the details of our proposed communi-cation system architecture composed of the Gaussian process-based modeling block and the error-driven communicationframework. In the first subsection, a brief explanation of theGaussian processes is presented, while combining the hybridmodel structures with error-driven communication policy isillustrated in the subsequent subsection.
A. Gaussian Processes: A Fully Data Driven non-parametricBayesian Modeling Approach
The record of different vehicle dynamics could be regardedas separate time-series which should be regressed using anappropriate supervised learning method. The regression prob-lem here is equivalent to inferring the characteristics of theunknown target functions which have generated these time-series through their available training sets, which are finitesets of known function output realizations. In this work,following our previous works in [16] [17], a non-parametricBayesian inference framework is proposed to find an appropri-ate representation and abstraction of the driver behavior usinghis observed actions through the recorded time-series of thevehicle dynamics. -2000200 X E NU GP Sub-modelCV Sub-model -2000200 Y E NU -2000200 Long i t ud i na l A cc e l e r a t i on Time (100ms step) -2000200 S t ee r i ng A ng l e Fig. 1: Performance of the proposed hybrid modeling schemein different driving scenarios. Setting a threshold on the track-ing error and utilizing a two-state hybrid modeling scheme,compromised of GP (with RBF + Linear kernel) and CVcomponents, this figure (using different colors) shows themoments when each sub-model satisfies the tracking accuracyconstraint while the other one fails and exceeds the threshold.In general, the main advantage of any non-parametric infer-ence method is relaxing the function-specific characteristicsduring the learning process and letting the model complexityto be derived from and adapted to the available training set. Inother words, a non-parametric inference method finds the bestfunction representation of the observed data without imposingany prior assumption on the form of the underlying function.Gaussian process (GP), as one of the most powerful non-parametric learning methods, puts the Bayesian prior directlyon the function space rather than parameterizing the functionand then putting the priors on the parameters space. This trickmakes the modeling method capable of capturing differentpossible patterns which might occasionally be observed in thetraining data. It is worth mentioning that we use the Gaussianprocess regression to derive the model of the remote vehicleand its driver as a unique object. The outcome is a set offunctions describing the underlying modes which represent thebehavior of this object for a notable time ahead.The formal definition of the Gaussian process is as follows:A Gaussian process defines a distribution over function valuesf(t) at any arbitrary point within the function input range, suchthat any finite subset of the drawn function values from thisdistribution form a multivariate Gaussian random vector (havejoint Gaussian distribution). [19]Posterior distribution is inferred by conditioning the prob-lem on a set of noisy observations as the training data. Gaus-sian process regression model assumes each observed valueas a draw from a normal random variable. Therefore, the setof m observations form an m -dimensional multivariate normalrandom vector. This multivariate random vector is defined byig. 2: Comparison of the prediction precision for GP andCS sub-models. Here, GP outperforms CS because of themaneuver non-linearity. (Courtesy of Google Earth Inc.)a mean vector of length m plus an m -by- m covariance matrix,also called as kernel within the Gaussian process context.The following equations describe the mathematical repre-sentation of GP framework. For more details one can refer to[19]. f ( t ) ∼ gp ( m ( t ) , k ( t, t ′ )) (1) { X i } i =1 , ,...,m = { f ( t i ) } i =1 , ,...,m ∼ N ( µ, Σ) (2) µ = m ( t i ); Σ i,j = k ( t i , t j ) ∀ i, j ∈ { , , ..., m } (3)The kernel matrix defines the correlation between the ele-ments of the marginal distribution. Capturing different patterns (a) Hybrid Gaussian Process architecture(b) Communication system architecture of a host and remote vehicle Fig. 3: Systems Design Illustrationis achievable in GP framework by utilizing different types ofkernels. In this work a compound kernel of RBF and linear hasshown the best performance in examined specific non-linearscenarios.
B. An Error-Driven Communication Architecture for HybridModel Structures
As mentioned in the introduction section, inter-vehicle com-munication in VANETs is our core interest in this work. As-suming a given cooperative vehicular scenario, e.g., a platoonor an intersection, in which each vehicle is equipped withV2X communication devices and interacts with neighboringvehicles, we use the host and remote vehicle naming conven-tions as it is common in the vehicular literature. By definition,the host vehicle (HV) receives situational awareness messagesrom the remote vehicle(s) (RVs) and runs cooperative safetyapplications locally in order to potentially react to the remotevehicles’ actions and maneuvers. From the networking point-of-view, each network node, i.e., each vehicle, can be modeledas a multi-layer stack. The application layer runs on top ofthe lower layers which together enable vehicles to commu-nicate over the air-interface. Considering a RV-HV pair ina network of vehicles, the HV receives multiple situationalawareness messages from vehicles in its communication range.As mentioned above, these messages contain dynamical stateinformation which give the HV an insight to create a real-timemap of it’s surrounding. This map then could potentially beused by the safety applications to avoid collisions or hazardoussituations for both HV and RV(s).Our communication technology-agnostic
MBC architecture,as illustrated in Figure 3, takes place in the application layerand is able to operate independent of the lower network, data-link, and physical layers. Figure 3b illustrates the networkprotocol stack and information flow for an arbitrary HV-RVpair. On the RV side, the Controller Area Network (CAN) busfeeds the application layer with vehicle’s local and sensoryinformation. The GP-based MBC module then trains the GPbased on the last received information. Afterwards, the MBCmodule keeps track of the prediction accuracy of the latestlearned GP model at any new GPS update and comparesit with a certain threshold. Whenever the difference of thelatest learned GP prediction and the actual GPS informationexceeds this threshold MBC module trains a new GP basedon the latest set of sensory inputs. This procedure resultsin generating a new situational awareness messages whichcarry the last updated abstract model of the vehicle’s state.Lower layers schedule and broadcast the message over the air-interface, i.e., communication channel. The corresponding HVnode receives situational awareness messages from all vehiclesin its communication range. The MBC module in HV sidereconstructs the state of the neighboring vehicles and createsa real-time predictive map of the surrounding nodes.In our settings, GPS latitude, longitude and elevation havebeen converted into ENU co-ordinations, then X-ENU and Y-ENU are treated as two separate time-series which should belearned from their own histories. Training window size hasbeen set to 10 latest equally spaced received GPS samplesin time (last 1 second) and a compound kernel type, com-posed of a linear and an RBF kernel, is selected due to ourobservations. Four different position tracking error thresholds,i.e. 20 cm, 30 cm, 40 cm, and 50 cm, are investigated inthis work. These values cover the range between minimumand maximum thresholds specified by SAE J2945 \
1, i.e 20and 50 cm, respectively. The schematic representation of theproposed hybrid model communication policy is presented inFigure 4 and the pseudo-code of our algorithm is illustratedin Algorithm 1. The evaluation results for the proposedframework are presented in the next section.
Algorithm 1
Model-based Communication Algorithm
Require:
Read CAN-bus at time i : S i = { x ,i , ..., x n,i } T ← T start ; i = 0 while T < T end dowhile ( P T E min < th ) or ( i = 0) do i ← i + 1 ; T next ← T + i for kernels ∈ H T nexti do P T E i = getPTE( kernel , S i ); end for P T E min = min ( P T E i ) end whileupdate T ; update S i ; P T E min ← ∞ , i ← end while IV. E
VALUATION
System level performance gain of the MBC architecturestems from its two core components; model-based informationexchange scheme and error-driven message transmit policy. Inthis section we first evaluate the performance gain originatingfrom the communication policy in terms of offered channelload. Tracking accuracy of MBC is then compared againstan error-driven raw information (conventional BSM) exchangepolicy (baseline) to further demonstrate its efficacy.Since the error-driven model update is an integral part of theMBC, we evaluated the same message scheduling policy forour baseline sensor-generated data exchange scheme whichuses constant-speed for position estimation. As mentionedearlier, the message scheduling rate in error-driven exchangedepends on the selected tracking error threshold. This thresh-old may vary depending on the application requirements. Weexperimented with four different thresholds to determine theresulting message generation rate for MBC and baseline, asillustrated in Figure 5. As the tracking error threshold getsstricter the message generation rate increases for both baselineand MBC. However, as the rate is significantly lower inFig. 4: Gaussian Process-based hybrid model update schemeig. 5: Message scheduling rate for different choices oftracking error thresholds. For MBC, the total effective rateis shown here by summing the model update rate and the sub-model identification rate.MBC, it can accommodate higher number of transmittingentities compared to baseline, assuming over-the-air packetlengths of MBC and baseline are similar. Moreover, MBCexperiences lower rate of packet collision comparing to itsbaseline counterpart in different traffic densities.Now that the efficacy of MBC is established in terms ofoffered channel load, we seek to determine its tracking per-formance gain. Tracking performance of the MBC architectureis evaluated for the above mentioned vehicle trip which iscarefully selected from the SPMD data-set [20] on the meritof maneuver counts over the entire trip duration. To ensurefair comparison, message generation rate in baseline is chosento be equal to the average model update rate computed inMBC. Tracking accuracy is determined in terms of positiontracking error (PTE) which is defined as the 2D Euclideandistance between the actual and estimated vehicle position.Actual position at a given time instant is obtained fromGPS logs and position estimation is calculated from receivedmessage information. Since PTE sampling is dependent onthe availability of actual position updates, it can be done atmost at the sampling rate of GPS updates, which is 10 Hzfor SPMD dataset. For position estimation, MBC uses themost recent model parameters and associated model relevancyupdates. In contrast, baseline vehicle position estimates arecomputed by coasting a vehicle’s last received position updatefrom the BSM to the error sampling instant (i.e., the GPSupdate instant), using a constant velocity mobility model.We measured tracking errors for different packet error ratio(PER) levels. PER is indicative of the communication channelquality and is defined as the ratio of the missed packets tothe transmitted packets. The PER metric can be interpretedfrom different perspectives. For a given traffic density, PER istypically an increasing function of sender-receiver separationdistance. Conversely, for a given sender-receiver range, PER Fig. 6: Tracking error comparison for packet loss ratio of40%. Stricter (smaller) tracking error threshold translates tohigher message transmit rate which eventually helps loweringthe tracking error.is an increasing function of traffic density. The “range” interpretation is useful for assessing tracking performance atdifferent ranges, while the “density” interpretation is usefulto evaluate range-specific system performance in differentdriving scenarios such as freeway with peak and off-peak hourtraffic. Another way to interpret PER is based on line-of-sightconditions of sender-receiver pairs where links with dominantLOS results in low PER. This interpretation is applicable forperformance evaluation in different driving environments suchfreeway and urban intersections.90-th percentile position tracking error (PTE), shown inFig 6, evidently suggests that MBC predicts positions moreaccurately than baseline. The higher tracking accuracy of MBCcan be attributed to its capability of capturing higher ordervehicle dynamics resulting from hard brakes and lane changemaneuvers. V. C
ONCLUDING R EMARKS
The notable differences in tracking accuracy of differentdriving maneuvers, resulted from different modeling schemesmotivates us to incorporate more complex model structuresin comparison with what is the current state-of-the-art invehicular society. More specifically, non-parametric Bayesianmethods with different kernels, which are capable of beingadapted to different maneuvers are potentially promising can-didates for this purpose. Therefore, in this work we haveproposed a Hybrid GP-based modeling scheme in combinationwith an error-driven model communication policy and investi-gated its performance against the same error-driven methodof raw-information dissemination. A notable improvementis observed using our scheme against the base-line methodthrough reduction of the required communication load as wellas better tracking precision.
EFERENCES[1] Federal Communications Commission. Amendment of the com-missions rules regarding dedicated short-range communication ser-vices in the 5.850-5.925 ghz band (5.9 ghz band.
FCC 02-302,https://docs.fcc.gov/public/attachments/FCC-02-302A1.doc , November2002.[2] J. B. Kenney. Dedicated short-range communications (dsrc) standardsin the united states.
Proceedings of the IEEE , 99(7):1162–1182, July2011.[3] H. Nourkhiz Mahjoub, A. Tahmasbi-Sarvestani, S. M. Osman Gani,and Y. P. Fallah. Composite α − µ based dsrc channel model usinglarge data set of rssi measurements. IEEE Transactions on IntelligentTransportation Systems , 20(1):205–217, Jan 2019.[4] 3rd Generation Partnership Project (3GPP). 3gpp - 36.885 - release 14.[5] B. Toghi, M. Saifuddin, H. N. Mahjoub, M. O. Mughal, Y. P. Fallah,J. Rao, and S. Das. Multiple access in cellular v2x: Performanceanalysis in highly congested vehicular networks. In , pages 1–8, Dec 2018.[6] S. M. Iranmanesh, H. Nourkhiz Mahjoub, H. Kazemi, and Y. P. Fallah.An adaptive forward collision warning framework design based on driverdistraction.
IEEE Transactions on Intelligent Transportation Systems ,19(12):3925–3934, Dec 2018.[7] X. Lu and S. Shladover. Integrated acc and cacc development for heavy-duty truck partial automation. In , pages 4938–4945, May 2017.[8] H. Kazemi, H. N. Mahjoub, A. Tahmasbi-Sarvestani, and Y. P. Fallah.A learning-based stochastic mpc design for cooperative adaptive cruisecontrol to handle interfering vehicles.
IEEE Transactions on IntelligentVehicles , 3(3):266–275, Sep. 2018.[9] A. Tahmasbi-Sarvestani, H. Nourkhiz Mahjoub, Y. P. Fallah, E. Moradi-Pari, and O. Abuchaar. Implementation and evaluation of a cooperativevehicle-to-pedestrian safety application.
IEEE Intelligent TransportationSystems Magazine , 9(4):62–75, winter 2017.[10] Society of Automotive Engineers (SAE). SAE J2735.[11] Society of Automotive Engineers (SAE). SAE J945/1.[12] C. Huang, Y. P. Fallah, R. Sengupta, and H. Krishnan. Adaptiveintervehicle communication control for cooperative safety systems.
IEEENetwork , 24(1):6–13, Jan 2010.[13] G. Bansal, J. B. Kenney, and C. E. Rohrs. Limeric: A linear adaptivemessage rate algorithm for dsrc congestion control.
IEEE Transactionson Vehicular Technology , 62(9):4182–4197, Nov 2013.[14] Y. P. Fallah. A model-based communication approach for distributedand connected vehicle safety systems. In , pages 1–6, April 2016.[15] E. Moradi-Pari, H. N. Mahjoub, H. Kazemi, Y. P. Fallah, andA. Tahmasbi-Sarvestani. Utilizing model-based communication and con-trol for cooperative automated vehicle applications.
IEEE Transactionson Intelligent Vehicles , 2(1):38–51, March 2017.[16] H. N. Mahjoub, B. Toghi, and Y. P. Fallah. A driver behaviormodeling structure based on non-parametric bayesian stochastic hybridarchitecture.
IEEE Vehicular Technology Conference (VTC) , August2018.[17] H. N. Mahjoub, B. Toghi, and Y. P. Fallah. A stochastic hybridframework for driver behavior modeling based on hierarchical dirichletprocess.
IEEE Connected and Automated Vehicles Symposium (CAVS) ,August 2018.[18] C. A. Flannagan D. J. LeBlanc M. D. Palmer R. K. Deering M.A. Shulman R. Kiefer, M. T. Cassar. ”forward collision warningrequirements project: Refining the camp crash alert timing approach byexamining last-second braking and lane change maneuvers under variouskinematic conditions”.
Report No. DOT-HS-809-574, 2003 .[19] Carl Edward Rasmussen and Christopher K. I. Williams.