Our presenters will cover a range of interesting topics in human factors and driving including driver speed behavior, driver response time, car following behavior, driver distraction, dangerous driving behavior, bicyclist communication and bicyclist stress level, and Responsibility Sensitive Safety (RSS). The presenters will also describe methodologies and techniques used to perform this research including naturalistic driving, driving simulation, bicycle simulation, virtual reality, machine learning, neural networks and driver modeling.
Analysis of Bicyclist Communication in a Simulator Environment
Patrick Malcolm, Technical University of MunichShow Abstract
Georgios Grigoropoulos, Technische Universitaet Muenchen
Andreas Keler, Technische Universitat Munchen
Heather Kaths, Technische Universitaet Muenchen
Klaus Bogenberger, Technische Universitat Munchen
Urban roads are a dynamic environment in which formal and informal communication are crucial. In order for autonomous vehicles to operate effectively in such an environment, they must be able to deal with this complexity. In particular, an understanding of the communication patterns of bicyclists is crucial because of their vulnerability and lack of standardized indicators, which allows greater variability and subtlety in the way they communicate. In this study, participants rode a bicycle simulator through various urban traffic scenarios in which explicit and implicit communication behaviors are expected. Participants were recorded with a depth camera, and a markerless motion capture technique was used to record their movements in three dimensions. From this data, hand signal, head movement, and leaning events were extracted, analyzed, and compared. Analyses showed that even in clearly regulated scenarios, not all participants performed a hand signal, with between 80 and 95 percent of participants signaling before turning at a four-way intersection and approximately 60 percent before changing lanes. Participants were also significantly more likely to glance over their left shoulder preceding a left lane-change or left turn compared to other scenarios. The arm shape while performing a hand signal was found to be almost entirely governed by individual preference, rather than scenario. Based on these results, it is clear that cyclist communication behavior frequently does not adhere strictly to traffic regulations and varies from person to person. However, implicit cues such as head movements can be used to supplement behavior prediction models in certain situations.
Effects of Driving Environment and Driver Characteristics on Speed Adaptation Mechanisms of Drivers
Ankit Kumar Yadav, Indian Institute of Technology, BombayShow Abstract
Nagendra Velaga (firstname.lastname@example.org), Indian Institute of Technology, Bombay
Speed adaptation behaviour of drivers plays a significant role in influencing crash risks, and remains a major road safety issue. Majority of the speed adaptation behavioural studies have been conducted in the western world; relatively little knowledge is available about the speed adaptation of the drivers of developing nations. The present study aims to investigate the speed adaptation behaviour of Indian drivers in changing driving environment, and to identify the significant predictors influencing their speed adaptation. The driving scenario (consisting of rural and urban driving environments) was designed on a driving simulator where eighty-two licensed drivers completed the driving task. Driver attributes (demographics and driving characteristics) were recorded with the help of a self-reported questionnaire. Speed adaptation of drivers was estimated as the difference of driving speed from the posted speed limit of a particular driving environment, averaged over the duration of driving. A Generalized Linear Model (GLM) was developed using speed adaptation as the dependent variable along with driving environments and driver attributes as the predictor variables. Results showed that speed adaptation of drivers was higher in rural driving environment compared to urban driving environment, indicating that drivers were less able to adapt to the corresponding speed limits in urban environment. Drivers’ age showed positive association with speed adaptation. Moreover, female drivers, drivers with prior crash history, and well-educated drivers, displayed higher speed adaptations than the other drivers. Findings of the present study may assist the road safety strategies and policy interventions in reducing the speed-related crashes.
Impact of Drivers' Characteristics on Speed Choice Behavior in Adverse Weather Conditions: A Driving Simulator Study
Mehdi Zolali, Imam Khomeini International UniversityShow Abstract
Babak Mirbaha, Imam Khomeini International University
Hamidreza Behnood, Imam Khomeini International University
Speed choice behavior varies between drivers, so it is difficult to analyze and study. However, it is important to know about drivers’ speed choice behavior in terms of enhancing safety. Furthermore, due to the fact that crashes occur at high speeds, analyzing the behavior of drivers who tend to choose higher speeds compared to other drivers has more importance. previous studies have studied the drivers' behavior in other countries except for Iran. Because of the high rate of death in Iranian roads, in this study, researchers have tried to investigate Iranian drivers’ characteristics who choose speed higher than the 85th percentile speed. For this purpose, a driving simulator was used to conduct this study. Six scenarios based on the visibility condition used in the driving simulator. More than 75 participants were evaluated and their speed data had collected. Using the t-test, speed data differences in six scenarios have proved. Results showed that some of the drivers’ characteristics like age, education level, and gender have an effect on drivers’ perception of speed and speed selection. Also, results indicate that speed (higher speed selection) is one of the most important keys in accident occurrence.
Bicyclist Maneuver Type Prediction using Bidirectional Long Short-Term Memory Neural Networks
Georgios Grigoropoulos (email@example.com), Technische Universitaet MuenchenShow Abstract
Patrick Malcolm, Technical University of Munich
Andreas Keler, Technische Universitat Munchen
Heather Kaths, Technische Universitaet Muenchen
Klaus Bogenberger, Technische Universitat Munchen
Fritz Busch, Technische Universitat Munchen
In the present research, a bicycle simulator is used to study the behavior of bicyclists and to collect data that can be used for developing a Bidirectional Long Short-Term Network (B-LSTM) model that predicts a bicyclist’s intended maneuver type class (left turn, right turn, straight) at an intersection approach. First the bicyclists’ dynamic behavior and the explicit and implicit communication behavior is recorded. Features describing the bicyclists’ behavior are extracted and associated with the respective maneuver type. The B-LSTM model is then trained on the bicycle simulator dataset. Five different model cases are investigated in order to identify the optimal input feature set and evaluate the added value from the inclusion of the bicyclists’ explicit and implicit communication behavioral data in the classification task. A maximum f1-score value of 83.8% (prediction accuracy per maneuver type: 85.3% left-turn, 84% right turn and 82.5% straight) is achieved using classes of implicit and explicit communication behavior together with the dynamic behavior data for the maneuver type prediction. Possible application areas of the proposed model may include, but are not limited to, the expansion and support of existing models and functions in the field of automated driving and the improvement of traffic efficiency and safety through real-time monitoring and forecasting of bicyclist behavior at intersections.
Performance and Safety Evaluation of Responsibility-Sensitive Safety in Freeway Car-Following Scenarios Using the Intelligent Driver Model and Model Predictive Control.
Xuesong Wang, Tongji UniversityShow Abstract
Omar Hassanin, Tongji University
Xiangbin Wu, Intel Labs China
This research evaluates the Responsibility-Sensitive Safety (RSS) model and a modified RSS in freeway car-following scenarios extracted from the Shanghai Naturalistic Driving Study (SH-NDS). In this project, 6146 car-following scenarios were extracted and divided into two groups, normal scenarios (5923) and safety-critical events (SCEs) (223), to evaluate the performance and safety of the RSS strategy. RSS was proposed by Mobileye as a mathematical model that defines the real-time safety distance that the automated vehicle needs to maintain from surrounding vehicles. Some modifications were made on the RSS safe distance to reduce its conservativeness without affecting safety. The modified RSS was then embedded into the Intelligent Driver Model (IDM) and the Model Predictive Control (MPC), but these modifications caused some problems, so the IDM and MPC were also modified to match the modified RSS. Vehicle-movement characteristics and surrogate safety measurements were analyzed to evaluate performance and safety. The performance results show that the modified RSS, IDM, and MPC are better than the originals in that they have higher average speeds, smaller relative distances, and better acceleration patterns. To evaluate safety, human drivers were compared with the modified IDM and MPC. The RSS reduced the severity of 88% of the SCEs and increased mean minimum time to collision from 1.53 s and 1.65 s to 3.44 s and 4.08 s for the IDM and MPC, respectively, versus human drivers. Therefore, the RSS model can be applied as a security guarantee to ensure the AV’s response to dangerous car-following situations.
Drivers’ Safety Grade and Ecology Grade Prediction Model Based on Random Forest
Xiaohua Zhao, Beijing University of TechnologyShow Abstract
Haolin Chen, Beijing University of Technology
Yiping Wu (firstname.lastname@example.org), Beijing University of Technology
Ying Yao, Beijing University of Technology
Yuan Yan, DiDi Chuxing
Cheng Gong, DiDi Chuxing
Yang Shi, DiDi Chuxing
Safety and ecology classification is particularly vital for drivers’ training and education. The objective of this paper is to establish a prediction model of driving safety grade and driving ecology grade based on random forest with statistical characteristics of dangerous driving behavior. This paper extracts 25 statistical characteristics of dangerous driving behaviors from overspeed, fatigue driving, phoning, and playing phone. The driver’s speed safety entropy (represents safety) and 100km fuel consumption (represents ecology) were calculated with the trajectory data, and the safety and ecology were classified into 5 grades in k-means clustering. Significance analysis screens out the statistical characteristics that have a significant influence on drivers’ safety grade and ecology grade. The above data is obtained to construct the drivers’ safety grade prediction model and drivers’ ecology grade prediction model based on random forest. The prediction accuracy of the two models is 86.33% and 85.46% respectively, with high accuracy. In addition, the importance ranking of the statistical characteristics that affect safety and ecology is obtained, the indicators of which are not consistent. Among them, the frequency of playing phone at nighttime has the greatest impact on safety and ecology. This paper establishes a foundation for the classification and prediction of safety and ecology in terms of the statistical characteristics of driving behavior, which enables to adopt the accessible statistical characteristics of driving behaviors to achieve safety and ecology classification of drivers, saving data acquisition costs. This paper provides technical support for drivers’ safety and ecology classification, as well as targeted education and training.
Cross-Platform Comparison of Driver Responses during Simulated Automated Driving and Correlations with Trust
Ganesh Pai (email@example.com), University of Massachusetts, AmherstShow Abstract
Michael Knodler, University of Massachusetts, Amherst
Cole Fitzpatrick, University of Massachusetts, Amherst
Jaydeep Radadiya, University of Massachusetts, Amherst
Sarah Widrow, University of Massachusetts, Amherst
Anuj Kumar Pradhan, University of Massachusetts, Amherst
Driving Simulation is a popular method for the experimental study of driver behaviors in the context of automated vehicles. Typical driving simulation approaches use scripted events and scenarios and programmed vehicle behaviors to simulate vehicle automation. Another simulation approach, the Wizard-of-Oz method, can be used for similar studies, and may offer added or unique advantages in flexibility, experimental efficiency and customization, and therefore broader options for experimental designs. This study examines specific driver behaviors - visual gaze, and hands and feet behaviors - during automated vehicle drives conducted with the above two experimental approaches, with the aim of evaluating the viability of the latter method as an experimental paradigm. Participants experienced simulated drives with SAE Level 3 automation in an advanced driving simulator, presented in two different platforms. An evaluation of drivers’ eye movements, hand, and feet behaviors while operating the simulated automated vehicle was carried out to explore similarities in driver behaviors across two platforms. A secondary analysis examined correlations between the driver behavior metrics and self-reported trust scores. Results indicate similar driver behaviors across platforms, contributing to the evidence for the Wizard of Oz approach as a simulation paradigm with similar validity as the de facto approach. Unexpectedly, no significant correlations of these driver behaviors with self-reported measures of trust were uncovered, highlighting the need to understand these driver behaviors better in the context of driver trust.
Car-following Behavior Factors Contributing to Rear-end Crashes and Near-crashes: A Naturalistic Driving Study
Xuesong Wang (firstname.lastname@example.org), Tongji UniversityShow Abstract
Xuxin Zhang, Tongji University
Feng Guo, Virginia Polytechnic Institute and State University (Virginia Tech)
Yue Gu, China Pacific Property Insurance Co.,Ltd
Xiaohui Zhu, China Pacific Property Insurance Co.,Ltd
Rear-end crashes are one of the most common types of crashes and driving behavior preceding rear-end crashes is a key risk factor. However, the relationships between daily driving habits and rear-end crash risk have not been well evaluated. The aim of this study is to identify the most influential factors in rear-end crash and near-crash (CNC) risk from drivers’ demographic characteristics, daily mobile phone usage, and car-following behaviors. Naturalistic driving study (NDS) data were used to explore the habitual driving behavior among drivers. The One-way Analysis of Variance (ANOVA) was used to identify the effects of gender, age, and driving experience groups. The Tobit model, with both fixed effect and random effect, was developed with demographic characteristics, mobile phone usage, and car-following behavior features as independent variables. The results indicate that there were significant differences in the standard deviation of jerk among drivers’ different individual demographic. Gender, age, phone usage duration, and maximum longitudinal acceleration all had significant positive correlations with CNC rate; experience, maximum longitudinal deceleration (negative value), and standard deviation of jerk were negatively correlated with CNC. The random parameter Tobit model was found to perform better than the basic Tobit model. The results of this study can provide guidance for the design of driving behavior optimization measures.
Driver Distraction Detection Based on Vehicle Dynamics Using Naturalistic Driving Data
Xuesong Wang (email@example.com), Tongji UniversityShow Abstract
Rongjiao Xu, Tongji University
Siyang Zhang, University of Missouri, Columbia
Distractive driving is very risky as it leads to crashes easily, especially phone use while driving. Existing studies attempted to establish methods to detect different kinds of driver distraction. Real-life distraction could be comprehensive including cognitive distraction, visual distraction, auditory distraction, and physical distraction simultaneously. Previous studies isolated drivers’ facial video data from vehicle dynamic data, which would lose context and may not be realistic. This study focuses on detecting driver distraction status based on vehicle dynamics data extracted from Shanghai Naturalistic Driving Study (SH-NDS), China. The performance attributes of speed, longitudinal acceleration, lateral acceleration, lane offset, and steering wheel rate were extracted from the vehicle Controller Area Network (CAN) data. The distraction status was classified into focused and distracted, of which distracted status included phone use. A Long Short-Term Memory (LSTM) model with attention mechanism and bidirectional layer was built for driver distraction status detection. The developed model showed promising result of approximately 88.2% accuracy on testing dataset, while other machine learning models like support vector machine (SVM), k-nearest neighbor (KNN), and adaptive boosting (AdaBoost) were able to detect provide overall accuracy of 83.4%, 81.5%, and 86.8% correspondingly. These results show that vehicle dynamics attributes of speed, longitudinal acceleration, lateral acceleration, lane offset, and their standard deviation and predicted error, along with steering wheel rate predicted error were significant and effective in detecting driver distraction. The developed LSTM model could potentially be applied in an advanced driver assistant system to reduce crashes caused by distracted driving.
Visualization of Driving Scenes for Realistic Simulator Experimentation - An Efficient Framework
Tingting Zhang, University of California, BerkeleyShow Abstract
xiao zhou, Wuhan University
Pei Wang, University of California, Berkeley
Ching-Yao Chan, Lawrence Berkeley National Laboratory
Driving simulator is a widely adopted experimental platform for investigating drivers’ behaviors under various traffic situations in a safe environment. This paper presents a methodological framework of developing video-based simulation programs on a driving simulator. Specifically, this study involves the evaluation of driver perception and understanding of information display signs through driving simulation using the presented framework. Based on video footages recorded from on-road driving, real-world signs on the video footages were detected and tracked automatically. Images of newly designed signs were integrated onto the video footages, and placed onto the real-world signs locations. Then the properties of the inserted images were fused to fit into the video background to yield a natural visual effect. The motion data collected during the real-world drive-through were incorporated into the video sequence as well as the movement of the simulator. Using throttle and brake pedals, participants drove through the video sequence with control over the playback speed of the video and the movement of the simulator to achieve a comparable visualization and motion experience as real-world driving. With this methodological framework, a driving simulation program based on real-life driving scene and vehicle movement can be quickly established and used for testing a variety of traffic control devices. The simulator platform with the implementation framework was proven to be effective and efficient in executing the experiments and evaluating the engagement of driving tasks and visual behaviors of the participants.
Predicting Drivers’ Reaction Time in Unexpected Lane Departure Situations Using Brainwave Signals: Application of Machine Learning Techniques
Soheil Borhani, University of Tennessee, KnoxvilleShow Abstract
Ramin Arvin (firstname.lastname@example.org), University of Tennessee, Knoxville
Ziming Liu, University of Tennessee, Knoxville
Asad J. Khattak, University of Tennessee
Xiaopeng Zhao, University of Tennessee, Knoxville
Miao Wang, Miami University
Deaths, injuries and property damage due to transportation crashes are unacceptably high costing about $1 trillion in the US. Driving is a complex task that requires perception of the surrounding environment, decision making, and vehicle control. Drivers predict the movement of surrounding objects and decide appropriate driving reactions. As a result of these complex demands, human errors contribute to almost 90% of crashes. To improve safety, this paper develops an innovative approach to multisensory intelligent neurological decoding that entails analyzing and predicting human drivers’ behavioral and cognitive performance. Using publicly available 32-channel electroencephalography (EEG) data from 27 subjects, driver behavior and brain dynamics in immersive driving simulator experiences are analyzed. Subjects drove on a four-lane highway and were instructed to keep the vehicle cruising in the center of the lane. Randomly induced lane-departure events caused drifts towards the left or right lane. The subjects’ brainwaves were analyzed as leading predictors of their reaction times before they act in response to unexpected events. Machine learning methods are applied to the data, which has the potential to revolutionize driving through biometric analysis and driver reaction time prediction. Using noninvasive brainwaves, the results show that drivers’ reaction times can be predicted within 350 millisecond accuracy while the average reaction time in the dataset was 1.1 seconds. Brainwave signals depict real-time cognitive state and carry valuable predictive ability even prior to the actual reaction. The study provides a novel approach to understanding cognition based on a multisensory Brain-Computer Interface.
Evaluation of Brake Reaction Time and Evasive Action Performance of Motorized Two-Wheeler Riders towards Jaywalking through Mock-Up Control Studies
Pradhan Kumar Akinapalli, Indian Institute of Technology, HyderabadShow Abstract
Digvijay Pawar, Indian Institute of Technology, Hyderabad
Jaywalking (illegal pedestrian crossing) demands additional attention from motorized road users as it gives less time to react to such situations and may result into a road accident. Brake reaction times (BRTs) and evasive action maneuvers decide the crash severity of such situations. This study attempts to understand the motorized two-wheeler (MTW) rider behavior and its interaction with the pedestrians during jaywalking using mock-up control study. Four different mock-up pedestrian crossing scenarios: two surprise crossing, one unexpected crossing, and one stationary, were designed to understand the MTW rider yielding behavior. The BRTs, approach speeds, decelerations, heading, and yaw rate were analyzed for fifty MTW riders. The findings of the study revealed that, 90th percentile BRT for the unexpected and surprise events were 3.62 and 1.6 s respectively. Further, a regression models of BRT with approach speeds and time to collision (TTC) were established. The study also classified braking maneuvers into hard and soft braking. The analysis results revealed that, 32% of drivers performed hard braking in case of surprise scenario and no hard braking was observed for stationary and unexpected scenario. In addition, analysis of yaw rate indicated loss of control of the MTW at the end of the maneuver for surprise events, indicating need for training MTW riders in critical situations. The observations from the paper are useful in the development of countermeasures to improve the safety of pedestrians and MTW riders.
An optimized algorithm of dangerous driving behavior identification based on unbalanced data
Jing Wang (email@example.com), Tongji UniversityShow Abstract
Yichuan Peng, Tongji University, Jiading
Chongyi Li, Tongji University
It is of great significance to identify dangerous driving behavior by extracting vehicle trajectory through video monitoring to ensure highway traffic safety. At present, there is no suitable method to identify dangerous driving vehicles accurately based on trajectory data. This paper aims to develop detection algorithm for identifying dangerous driving behavior based on the road scene, which is mainly composed of imbalanced dangerous driver detection and labeling, extraction of driving behavior characteristics and the establishment of recognition model about dangerous drving behavior. Firstly, this paper defines the risk index of the vehicle related to five types of dangerous driving behavior: dangerous following, lateral deviation, frequent acceleration and deceleration, frequent lane change, and forced insertion. Then, a variety of methods including K-means clustering, local factor anomaly algorithm, isolation forest and OneClassSVM are used to carry out anomaly detection on the risk indicators of drivers, and the optimal method of them is proposed to identify dangerous drivers. Then the speed and acceleration of each vehicle are Fourier transformed to obtain the characteristics of the driver's driving behavior. Finally, considering the imbalanced characteristic of the analyzed dataset with a very small proportion of dangerous drivers, this paper compares a variety of imbalanced classification algorithms to optimize the recognition performance of dangerous driving behavior. The results show that the OneClassSVM detection algorithm can be effectively applied to the identification of dangerous driving behavior. The improved xgboost algorithm performs best for extremely imbalanced data of dangerous drivers.
Quantifying Bicycling Stress Level Using Virtual Reality and Electrodermal Activity Sensor
Mohsen Nazemi (firstname.lastname@example.org), Swiss Federal Institute of Technology (ETH Zurich)Show Abstract
Michael van Eggermond, University of Applied Sciences Northwestern Switzerland
The application of wearable wireless electrodermal activity (EDA) devices in virtual reality (VR) experiments has become increasingly popular and lends itself for the application in behavioral research for transport planning. The type of infrastructure and interaction with other road users can invoke different arousal levels in urban bicyclists, which can be modeled with VR applications and captured by EDA sensors. At the intersection of engineering, psychology, and physiology, this research attempted to quantify bicycling stress levels using a bicycle simulator combined with immersive 360-degree virtual reality and an EDA sensor. Overall, 150 participants rode through 5 different bicycling environments in VR while their elicited skin conductance responses (SCRs) were passively collected by an EDA sensor that was connected to the participants for the duration of the experiment. Analysis of the signal for the entire stretch of the bicycling course did not yield significant differences in SCRs between different bicycling environments. However, comparing smaller segments of the bicycling course revealed significant differences. Bicycling on the sidewalk shared with pedestrians caused higher stress levels while bicycling on the segregated bicycle path found to be the least stressful. Evidence was found of a link between self-reported perceptions of safety and SCR rates. The results of this research shows a promising path in using VR experiments to identify stressful events and locations and to quantify bicycling stress level for non-existent future facility designs.
Development of a High Fidelity Virtual Reality Cycling Simulator for Road Safety Education and Research
Fred Feng (email@example.com), University of Michigan, DearbornShow Abstract
Ayah Hamad, University of Michigan, Dearborn
Inspired by driving simulators, a virtual reality (VR) cycling simulator has potential to become a valuable educational and research tool for improving bicycling safety. This paper presents the work of developing an immersive, high-fidelity VR cycling simulator where one can ride a bicycle in a simulated virtual environment and interact with other road users (e.g., drivers). The hardware and software development were described in detail with the hope that it can be helpful for others who wish to build a similar system. Using a simulator engine (Unity®), a VR simulation was created, including a representation of a real-world urban environment with a road network. The HTC® Vive Pro Eye was used as the VR system of choice, with an appropriately matched stationary bicycle system setup for the purposes of the simulation. A use case of the simulator for driver education and training was presented. A variety of common dangerous encounters for bicyclists, such as close pass and dooring, were programmed into the VR scenarios. Drivers who have little experience in bicycling can ride the simulator and experience the encounters in a safe, virtual setting from a bicyclist’s perspective. Further research can be conducted to examine whether having such an experience would change their awareness and empathy towards bicyclists, and whether there are any behavioral changes when interacting with bicyclists on the road as drivers. In addition, the VR scenarios can also be converted into stereoscopic 360° videos and published online, so that people can experience the scenarios anywhere by playing the videos on their personal smart devices without the cycling simulator.
DISCLAIMER: All information shared in the TRB Annual Meeting Online Program is subject to change without notice. Changes, if necessary, will be updated in the Online Program and this page is the final authority on schedule information.