PUBLICATIONS

 The near-miss events involving vulnerable road users can lead to serious accidents. Safe and careful expert drivers perform a hazard-anticipatory driving and they will naturally seek to reduce the uncertainty by attempting to fit their current driving context into a pre-existing category they have already developed, that is, predicting what can happen. In this study, our target situation consists of a cyclist attempting a road crossing at a blind spot. This study aims at developing a context-aware driver model for determining the recommended driving speed at blind intersections based on the analysis of near-miss-incidence database, which includes the data on driver behavior and road environmental factors just before the near-miss. First, we extracted the drive-recorder data using the management tool provided in the database. Second, risk, which is defined as the time margin for drivers to perform evasive actions to avoid a crash, was quantified for the extracted data using the safety-cushion time. The safety-cushion time can be observed as a result of the driver’s adjustment to the vehicle velocity depending on the given road environment. One of the key aspects in developing the context-aware driver model is to categorize the extracted near-miss data into two levels based on the risk quantifications: low- and high-risk events. The low- and high-risk events were regarded as a result of the driver’s appropriate adjustment of, and inability or failure to adjust the vehicle velocity depending on the given road environment, respectively.

 Third, based on a multiple linear regression analysis with low-risk event dataset, we constructed a context-aware driver model to produce the recommended vehicle speed depending on the given road environment. The road environment variables, determined by stepwise regression, were identified as factors that reduced or increased the vehicle velocity at blind intersections, and were incorporated into the model as predictors. Furthermore, we quantitatively visualized drivers setting the baseline for speed adjustment and increasing or decreasing the speed according to the given road environment context. Fourth, the model validation demonstrated a coefficient of determination (R2) of 0.20, and a mean absolute error (MAE) of 6.54 km/h on average in the 5-fold cross- validation. Finally, to investigate the effectiveness of the constructed driver model on safety performance, we used the dataset of high-risk events as test data. Theoretically, the constructed driver model guided the drivers to drive the vehicle at the recommended speed, and thus convert more than half of the high-risk events into low-risk events. These results indicate that the context-aware driver model is feasible to be used to adjust the approaching speed at blind intersections in accordance with the road environment factors.

 Near-miss events or accidents involved pedestrians and cyclists are the result of conflict between a driver behavior and the road user behavior. When expert drivers with more driving experience are facing uncertainty, they naturally seek to reduce the uncertainty by attempting to fit their current driving context into a pre-existing category based on knowledge-based decision making. Our study goal is to develop DAS to attain “a hazard-anticipatory driving” depending on driving contexts, through both the driver behavior analysis and the clarification of cause-and-effect chain in accident mechanism. The near-miss incident database has been constructed and managed by the Smart Mobility Research Center (SMRC [3]) of Tokyo University of Agriculture and Technology in JAPAN since 2004. The main contribution of this paper is to propose a context-sensitive driver model to determine the recommended speed in intersection scenarios. Based on investigations with the near-miss event database, this paper describes a method for determining the recommended safe speed based on the annotated information by applying the statistical processing and the machine learning techniques, and then this study explores the mechanism of adjusting the vehicle velocity according to the given road environment context.

 The driving safety performance of autonomous driving vehicles must be ensured before on-road implementation. Because it is not realistic to evaluate every single test condition in real-world traffic, computer simulation methods can be used. The driving safety performance can be evaluated by simulating various driving scenarios and calculating surrogate indicators representing dangerous collision risk. This study used a near-miss database and introduced a surrogate indicator that represents a potential risk in the driving scenarios for rear-end and cut-in collisions. The near-miss video database includes several driving scenarios experienced by human drivers, such as dangerous situations that lead to accidents, potentially dangerous situations that have a risk probability to escalate into dangerous situations, and normal driving situations. A skilled and attentive human driver foresees dangerous situations while driving and avoids them. Therefore, autonomous driving vehicles, which should be safer than human driving, must avoid potentially dangerous situations, as well as overtly dangerous ones. Level 3 autonomous driving vehicles must be safely operated to prevent potentially dangerous situations for rear-end collisions and cut-in collisions, which are the most frequent danger cases on highways. A calculation method of surrogate indicators to predict the severity of driving scenarios for rear-end and cut-in collisions was developed. The near-miss video database was used to validate that these indicators can illustrate risk probabilities and help assess dangerous situations. Thus, dangerous situations and potentially dangerous situations in the driving scenarios for rear-end and cut-in collisions were quantified using the surrogate indicators, and the driving safety performance of autonomous driving vehicles could be evaluated.

 本研究は，死角からの自転車道路横断の潜在危険を対象に，運転文脈へのドライバの応答において，上手く適応できた成功データから学ぶことによって推奨速度を決定するコンテキストアウェアドライバモデルを提案する．

 This study proposes a context-sensitive hazard anticipation for predicting the criticality of situations depending on driving context and driving behavior state. The data of 901 near-miss events were extracted to analyze human error as well as cause-and-effect chain studies of accidents, and the annotations that describe the driving context were investigated to find their influence on the criticality of the recorded incidents. The results can be used to develop next generation ADAS or to improve algorithms for autonomous driving technology, increasing their safety performance as well as the driver acceptance.

 著者らは交通ヒヤリハット場面を収集および分類したデータベースの構築を行っている． 本研究では，既存のデータベースシステムに生じている課題を解決するために，Ontologyに基づく新たなアノテーション方式の提案，および，柔軟なデータ入出力を実現するためのインターフェース設計について報告する．

 This study proposes a potential risk estimation method for predicting the criticality of driving situations with respect to driving context and driving behavior state. The data of 901 cases were extracted to conduct human error analysis as well as cause-and-effect chain studies of accidents, and the annotation parameters that describe the driving context were investigated to find their influence on the criticality of the recorded incidents. The obtained results can be used to develop next generation ADAS or to improve algorithms for autonomous driving, increasing their safety as well as the driver acceptance.

 This paper introduces a method for predicting the criticality of situations in traffic depending on the driving context and the driving behavior. The corresponding models are based on crash-relevant data recorded in Japan since more than ten years. The handling of the big amount of data, especially finding the relevant data considering the objective of analysis was one of the major challenges of this work. It has been achieved by an intelligent filter that is based on objectified driving behavior of experienced drivers with a sense of responsibility in traffic. The results of this work can be used to develop next generation ADAS or to improve algorithms for autonomous driving, increasing their safety as well as the driver acceptance.

 Image and vehicle data recorded by drive recorders are analyzed to clarify the difference of rear-end near-miss incidents between the low-speed (20km/h or lower) and the high speed (40 km/h or higher) regions. The major understandings to be drawn from the analysis are as follows. There are significantly fewer relevant factors involved in the low-speed incidents than high-speed incidents. In the cases of low speed, small overlap cases account for a comparatively high percentage in the number of incidents whereas the actual accidents have different tendency. In the case of high speed, rear-end incidents after another car cuts in front of a subject car accounts for a considerably higher percentage compared with the low-speed cases.

 Data recorded using drive recorders (DRs) and stored in a “Near-miss incident database” are analyzed to assess the difference between low- and high-speed rear-end near-miss incidents. The following conclusions are obtained: (1) the higher is the car’s speed, the greater is the potential for rear-end near-miss incidents in a single lane. (2) Rear ending after another car cuts in front of a car with a DR (RACC) accounts for a higher percentage of high-speed cases compared with low-speed cases. (3) For high-speed RACC, most cars approach at relatively high speeds in left-side cases, whereas right-side cases present various situations.

 東京農工大学では、自動車技術会の活動を引き継ぎ、ドライブレコーダによるヒヤリハットデータを収集分析するドラレコＤＢセンターを設置した。現在は都内のタクシーをはじめ全国数か所のヒヤリハットデータベースをおよそ６．５万件登録している。本解説では、その活用の現状と今後の可能性を述べる。

 This paper describes the assessment of a collision avoidance system with active braking assistance by using a driving simulator with reconstructed near-miss incident scenario relevant to vulnerable road users. In the first half of the paper, near-miss incident data collected by image-captured drive recorders, which are relevant to pedestrian near-miss, are analyzed to get the basic parameters of pedestrian relevant crash scenario model. A representative situation in unsignalized intersection is focused in this paper. In the latter half, a hazard-anticipatory driving assistance system to prevent crashes in unsignalized intersections is proposed and implemented in the driving simulator. The collision avoidance performance evaluation by human-in-the-loop experiments are conducted.

 This paper describes the pedestrian and bicycle behavior model while crossing the road at unsignalized traffic intersections from the stored near-miss incident database. In the first half of the paper, the near-miss incident database analysis is shown by classifying the near-miss incidents into various categories depending on location, time, dash-out position and moving direction. In addition, the reaction behavior of the pedestrians or bicycles when facing the vehicles are also analyzed. In the latter half, the numerical parameters which are the basis of driver assistance system design such as the velocity of the pedestrian or bicycle crossing the road and Time-To-Conflict-Point in near-miss incidents are discussed.

 To reduce traffic accident against pedestrian/bicycle, it is necessary to develop not only the passive safety system or the pre-crash safety system, but also the active safety system. The system is evaluated by using accident data. But it is sometimes hard to estimate the scenario of the accident with the data. The effectiveness of Collision Avoidance Brake System is analyzed by using Hiyari-Hatto (near-miss incident) data recorded by drive recorders equipped on taxis in Tokyo. This paper proposes the evaluation method of pedestrian/bicycle accident avoidance system based on analysis of near-miss incident database.