Kei Hiroi

Indoor positioning method integrating pedestrian Dead Reckoning with magnetic field and WiFi fingerprints

In this paper, we propose a high accuracy indoor positioning method that uses residual magnetism in addition to Pedestrian Dead Reckoning (PDR) and WiFi-based localization methods. Our proposed method needs WiFi and magnetic field fingerprints, which are created by measuring in advance the WiFi radio waves and the magnetic field in the target map. The fingerprints are represented by a Gaussian Mixture Models (GMMs) to reduce the amount of computation. Our proposed method estimates positions by comparing the pedestrian sensor and fingerprint values by particle filters. We evaluated this method in real environments and confirmed that it provides accurate indoor positioning with a mean error less than 8 m and more accurate position detection than existing techniques.

Pedestrian dead reckoning based on human activity sensing knowledge

This research addresses improvement of the accuracy of pedestrian dead reckoning (PDR), which is one effective technique to estimate indoor positions using smartphone sensors. Even though various techniques using step lengths and their number have been previously proposed for PDR, insufficient accuracy is gotten from smartphone sensors. In this research, we define human activity sensing knowledge and propose improvements to PDR accuracy based on it. Human activity sensing knowledge consists of four kinds of information: pedestrian, environmental, activity, and terminal. Previous studies separately used these kinds of information; however, no study has systematically arranged them for use in PDR. We improved PDR accuracy by adjusting the step length in passages and on stairs and revised activity recognition error with human activity sensing knowledge. To investigate the effectiveness of that strategy, we used HASC-IPSC, which is an indoor pedestrian sensing corpus. After our investigation, activity recognition accuracy improved from 71.2% to 91.4%, and the distance estimation error was reduced from approximately 27 m to approximately 7 m using human activity sensing knowledge.

PIEM: Path Independent Evaluation Metric for Relative Localization

There are many methods for indoor positioning. These methods are divided into the relative localization and absolute localization. In the relative localization, one widely used method is Pedestrian Dead Reckoning (PDR). Relative localization estimates the moving distance, orientation, and height of the pedestrian. However, relative localization has a problem caused by an accumulated error: the longer the path, the worse the accuracy of relative localization. There is another problem in the existing evaluating metrics: they compare only the actual location and the estimated location of the destination. Relative localization also has this evaluation problem. We propose PIEM: Path Independent Evaluation Metric for Relative Localization. PIEM is a path independent evaluation metric, considering the complexity of the path; distance, orientation, and height. Then we evaluate these three factors of relative localization in addition to the position. Our proposed method showed more consistent results for the complexity of the path than the existing methods of relative localization evaluation.

HASC-PAC2016: large scale human pedestrian activity corpus and its baseline recognition

Human activity recognition by wearable sensors will enable a next-generation human-oriented ubiquitous computing. However, most of the existing research on human activity recognition is based on a small number of subjects, and lab-created-data. To overcome this problem, we hold HASC Challenge as a technical challenge to collect the data for activity recognition. In addition to HASC Challenge, we collected indoor pedestrian sensing data of 107 people with a balance of gender and age (HASC-IPSC). Through these data collection, we gathered 111,968 sensor files of 510 subjects. For the convenience of the future researchers in this field, we combined them as a single corpus named HASC-PAC2016 and make it public. Baseline recognition result of HASC-PAC2016 segmented data is 73.4% accuracy for overall, 81.4% for limited by terminal position, and 85.1% with file-based recognition. For sequence data, we only get 73.4% even for limited subjects. This shows we need further research of activity recognition using HASC-PAC2016.

A pedestrian passage detection method by using spinning magnets on corridors

The passage event on the specific spot is one of the useful information for position estimate. If we can detect the passage of the specific spot, we could contribute to the field of the position estimate because it is available for movement course identification, and the correction of the position estimate error. We suggest pedestrian passage detection methods by using magnets. We make the characteristic magnetic field as a marker. We detect the pedestrian passage by reading a marker by a smartphone device. We generate a magnetic field marker by rotating magnets. We can acquire a passage direction by rotating two magnets at different frequencies. We can detect a passage with an accuracy rate of 100 percent and a passage direction with an accuracy rate of 94 percent when the distance between magnets and a smartphone is less than 75cm.

Design and implementation of algorithm for estimation of elevator travel distance using smartphone accelerometer

Since the advent of smartphones equipped with sophisticated sensing hardware, human activity recognition research has moved from utilizing dedicated sensing devices to using commercial smartphones. This paper presents the design of an algorithm to recognize and estimate travel distance when riding an elevator and its corresponding implementation within an app for a smartphone running Apples mobile operating system (iOS). The algorithm receives solely the signal of the smartphones accelerometer, recognizes that it belongs to an elevator ride, and proceeds to calculate the distance traveled by the rider. The algorithm has been designed in a way that simplifies the necessary operations to calculate the travel distance with the objective of minimizing processing power, while keeping the corresponding estimations highly accurate. This is an initial attempt towards the building of a robust, but simple and fast, real-time human activity recognition service on a wearable platform.

A proposal of IndoorGML extended data model for pedestrian-oriented voice navigation system

We propose Landmark-Conscious Voice Navigation as one type of a pedestrian navigation system, which navigate users by only voice guidance. It is necessary to standardize data model in order to use this system widely. In a previous paper[1], we constructed a basic voice navigation system, which uses Open Street Map based data model. In this paper, at first, we conduct an experiment of voice navigation at an underground shopping area of Nagoya Station with two types of landmark descriptions. After that, we discuss what data structure is necessary to describe landmark information for voice navigation. Therefore, we propose to extend IndoorGML1.0 by adding landmark space as a new defined data model for voice navigation. The main contribution of this paper is that we conduct an experiment of voice navigation and research how different landmark descriptions affect users; furthermore, we discuss a IndoorGML extended data model for voice navigation.

Non-Local Dictionary Based Japanese Dish Names Recognition Using Multi-Feature CRF from Online Reviews

In cuisine recommender service, online user review is an important data source avoiding a cold-start problem. Cuisine-domain named entity recognition(NER) can be used as an entrance to comprehend the semantic information of reviews. This paper describes a supervised approach recognizing Japanese dish name entity (DNE) from online reviews of Japanese cuisine website. In the first stage, this work adopts tweets as the data source to construct the dictionary of dish name elements through semantic rules and use Bayesian posterior to remove noise. Next stage, we maps first-stage dictionary as a non-local feature into Conditional Random Field (CRF) to recognize the dish name. This method can automatically add new dish name elements into the non-local dictionary by iteration during the recognition proceeding. By using 10-fold validation, experimental results show our method can reach 84.38% in F1 score and outperform the two baselines using the dictionary or CRF with term feature separately.

Design and Implementation of Event Information Summarization System

In this research, we have designed and implemented an Event Information Summarization System (EISS) for collecting Event Info as a web-service. EISS collects mass event data from several non-uniform event website APIs and data sources. The Collected event data is visualized by some user-friendly user interfaces for consumer. EISS can summarize the event data in locational info and temporal info automatically and visualizes them to consumer on online maps. The Event Info is not showed to the consumer by a single list any longer. The consumer will experience the Event Info that is shown by locational online maps. Consumers also can set the query conditions or categories of events to filter out the events info that they need. We also designed and implemented a machine-learning algorithm to estimate the categories of event. EISS results in F1-Score to 0.47 by simple feature. We mentioned that some features are strong and positive correlate with categories expressly.

A location estimation method using mobile BLE tags with tandem scanners

We have developed an indoor location estimation method using mobile Bluetooth Low Energy (BLE) tags carried by people and BLE scanners fixed to a building. By using the method, we can analyze the behavior of the attendees at some large-scale exhibition, such as the order of the visited booth and the duration of the stay. Using mobile BLE tags has some advantages: to collect a large amount of data easily, to provide location estimation without smart-phones. However, in a real environment, the BLE signal is unstable due to many people and obstacles. Fingerprinting is difficult because arranging booths finishes few hours before exhibition starts. Our previous trilateration method resulted in 10--30 meter accuracy with the data collected at Geospatial EXPO 2015 because of packet loss. We, to decrease packet loss, have developed tandem scanner that is equipped with multiple Bluetooth adapters. This paper presents the result of the improved method with the data collected by tandem scanners at Geospatial EXPO 2016. We also examined what improves accuracy: shorter advertising interval of the BLE tag and a larger number of tandem scanners available adapters. Average location estimation accuracy was 4.51 meters when using best settings.