Driver Maneuvers Research Articles

Modeling of Car-to-motorcycle Overtaking Maneuver Based on Comfort Zone Boundaries

Thailand has one of the world's highest road fatality rates, mainly on motorcycles. In mixed traffic, motorcycles coexist with other vehicles. The interaction between cars and motorcycles, such as overtaking due to speed differences, can lead to accidents. This scenario also has implications for autonomous vehicles interacting with motorcycles. To increase safety in such interactions, a model was developed that simulates overtaking maneuvers of car drivers with motorcycles, using the concept of comfort zone boundary and a four-phase classification. In a driving simulator, 648 overtaking maneuvers collected from 36 Thai drivers were recorded with different lateral positions and speeds of the motorcycles. A novel graphical method using steering wheel angle and steering wheel velocity signals facilitated the identification of the phases. Time-to-collision and lateral distance characterized driver comfort zones and served as an indicator for safety measures. The lateral position of the motorcycle has proven to be the most influential factor in the model. The results suggest that overtaking vehicles exhibit non-lane-bound driving characteristics and a risk for the sideswipe accident is identified. These results provide a foundational framework for advanced driver assistance systems and motion planning of autonomous vehicles, contributing to improved road safety.

Generation of nighttime pedestrian fatal precrash scenarios at junctions in Tamil Nadu, India, using cluster correspondence analysis

Objective Modern transportation amenities and lifestyles have changed people’s behavioral patterns while using the road, specifically at nighttime. Pedestrian and driver maneuver behaviors change based on their exposure to the environment. Pedestrians are more vulnerable to fatal injuries at junctions due to increased conflict points with vehicles. Generation of precrash scenarios allows drivers and pedestrians to understand errors on the road during driver maneuvering and pedestrian walking/crossing. This study aims to generate precrash scenarios using comprehensive nighttime fatal pedestrian crashes at junctions in Tamil Nadu, India. Methods Though numerous studies were available on identifying pedestrian crash patterns, only some focused on identifying crash patterns at junctions at night. We used cluster correspondence analysis (CCA) to address this research gap to identify the patterns in nighttime pedestrian fatal crashes at junctions. Further, high-risk precrash scenarios were generated based on the positive residual means available in each cluster. This study used crash data from the Road Accident Database Management System of Tamil Nadu State in India from 2009 to 2018. Characteristics of pedestrians, drivers, vehicles, crashes, light, and roads were input to the CCA to find optimal clusters using the average silhouette width, Calinski-Harabasz measure, and objective values. Results CCA found 4 clusters with 2 dimensions as optimal clusters, with an objective value of 3.3618 and a valence criteria ratio of 80.03%. Results from the analysis distinctly clustered the pedestrian precrash behaviors: Clusters 1 and 2 on pedestrian walking behaviors and clusters 3 and 4 on crossing behaviors. Moreover, a hidden pattern was observed in cluster 4, such as transgender drivers involved in fatal pedestrian crashes at junctions at night. Conclusion The generated precrash scenarios may be used to train drivers (novice and inexperienced for nighttime driving), test scenario creation for developing advanced driver/rider assistance systems, hypothesis creation for researchers, and planning of effective strategic interventions for engineers and policymakers to change pedestrian and driver behaviors toward sustainable safety on Indian roads.

STA-Net: A Spatial–Temporal Joint Attention Network for Driver Maneuver Recognition, Based on In-Cabin and Driving Scene Monitoring

Next-generation advanced driver-assistance systems (ADASs) are a promising direction for intelligent transportation systems. To achieve intelligent security monitoring, it is imperative that vehicles possess the ability to accurately comprehend driver maneuvers amidst diverse driver behaviors and complex driving scenarios. Existing CNN-based and transformer-based driver maneuver recognition methods face challenges in effectively capturing global and local features across temporal and spatial dimensions. This paper proposes a Spatial–Temporal Joint Attention Network (STA-Net) to realize high-efficient temporal and spatial feature extractions in driver maneuver recognition. First, we introduce a two-stream architecture for a concurrent analysis of in-cabin driver behaviors and out-cabin environmental information. Second, we propose a Multi-Scale Transposed Attention (MSTA) module and Multi-Scale Feedforward Network (MSFN) to extract features at multiple scales, addressing receptive field inadequacies and combining high-level and low-level information. Third, to address the information redundancy in multi-scale features, we propose a Cross-Spatial Attention Module (CSAM) and Multi-Scale Cross-Spatial Fusion Module (MCFM) to select essential features. Additionally, we introduce an asymmetric loss function to effectively tackle the issue of sample imbalance across diverse categories of driving maneuvers. The proposed method demonstrates a remarkable accuracy of 90.97% and an F1 score of 89.37% on the Brain4Cars dataset, surpassing the performance of the methods compared. These results substantiate the fact that our approach effectively enhances driver maneuver recognition.

Driver Maneuver Interaction Identification with Anomaly-Aware Federated Learning on Heterogeneous Feature Representations

Driver maneuver interaction learning (DMIL) refers to the classification task with the goal of identifying different driver-vehicle maneuver interactions (e.g., left/right turns). Existing conventional studies largely focused on the centralized collection of sensor data from the drivers' smartphones (say, inertial measurement units or IMUs, like accelerometer and gyroscope). Such a centralized mechanism might be precluded by data regulatory constraints. Furthermore, how to enable an adaptive and accurate DMIL framework remains challenging due to (i) complexity in heterogeneous driver maneuver patterns, and (ii) impacts of anomalous driver maneuvers due to, for instance, aggressive driving styles and behaviors. To overcome the above challenges, we propose AF-DMIL, an Anomaly-aware Federated Driver Maneuver Interaction Learning system. We focus on the real-world IMU sensor datasets (e.g., collected by smartphones) for our pilot case study. In particular, we have designed three heterogeneous representations for AF-DMIL regarding spectral, time series, and statistical features that are derived from the IMU sensor readings. We have designed a novel heterogeneous representation attention network (HetRANet) based on spectral channel attention, temporal sequence attention, and statistical feature learning mechanisms, jointly capturing and identifying the complex patterns within driver maneuver behaviors. Furthermore, we have designed a densely-connected convolutional neural network in HetRANet to enable the complex feature extraction and enhance the computational efficiency of HetRANet. In addition, we have designed within AF-DMIL a novel anomaly-aware federated learning approach for decentralized DMIL in response to anomalous maneuver data. To ease extraction of the maneuver patterns and evaluation of their mutual differences, we have designed an embedding projection network that projects the high-dimensional driver maneuver features into low-dimensional space, and further derives the exemplars that represent the driver maneuver patterns for mutual comparison. Then, AF-DMIL further leverages the mutual differences of the exemplars to identify those that exhibit anomalous patterns and deviate from others, and mitigates their impacts upon the federated DMIL. We have conducted extensive driver data analytics and experimental studies on three real-world datasets (one is harvested on our own) to evaluate the prototype of AF-DMIL, demonstrating AF-DMIL's accuracy and effectiveness compared to the state-of-the-art DMIL baselines (on average by more than 13% improvement in terms of DMIL accuracy), as well as fewer communication rounds (on average 29.20% fewer than existing distributed learning mechanisms).

Deep Learning-based Motor Prognostics for Electric Power Steering Systems

The electric power steering (EPS) system is a fundamental unit for modern automotive vehicles, providing motor assistance and various control functions to aid the driver's manual maneuvering. Preventing loss of assist (LoA) in advance is critical for EPS systems to mitigate fatal accidents and maintenance costs. As part of the recent interest in intelligent software-defined vehicles (SDV), data-driven approaches have gained much attention to overcome the limitations of conventional fail-safe and maintenance strategies. However, while related works in the field have shown promising results, they are limited to proof-of-concept studies for detecting and isolating artificially injected faults using simulation and test-bed data, which lacks relevance and poses limitations for practical in-vehicle implementation. Herein, we present a deep learning (DL)-based method to predict the motor degradation level in an EPS system using in-vehicle data acquired from the controller area network (CAN) bus. Our approach proposes a novel multivariate transformer (MVT)-based neural twin model of the EPS system, which is trained using an adversarial learning strategy by utilizing only the normal data. Our method can detect degradation levels down to ten percent from normal working conditions based on an anomaly detection mechanism that outperforms baseline methods in quantitative and qualitative measures.

A Machine Learning-Based Correlation Analysis between Driver Behaviour and Vital Signs: Approach and Case Study.

Driving behaviour analysis has drawn much attention in recent years due to the dramatic increase in the number of traffic accidents and casualties, and based on many studies, there is a relationship between the driving environment or behaviour and the driver's state. To the best of our knowledge, these studies mostly investigate relationships between one vital sign and the driving circumstances either inside or outside the cabin. Hence, our paper provides an analysis of the correlation between the driver state (vital signs, eye state, and head pose) and both the vehicle maneuver actions (caused by the driver) and external events (carried out by other vehicles or pedestrians), including the proximity to other vehicles. Our methodology employs several models developed in our previous work to estimate respiratory rate, heart rate, blood pressure, oxygen saturation, head pose, eye state from in-cabin videos, and the distance to the nearest vehicle from out-cabin videos. Additionally, new models have been developed using Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) to classify the external events from out-cabin videos, as well as a Decision Tree classifier to detect the driver's maneuver using accelerometer and gyroscope sensor data. The dataset used includes synchronized in-cabin/out-cabin videos and sensor data, allowing for the estimation of the driver state, proximity to other vehicles and detection of external events, and driver maneuvers. Therefore, the correlation matrix was calculated between all variables to be analysed. The results indicate that there is a weak correlation connecting both the maneuver action and the overtaking external event on one side and the heart rate and the blood pressure (systolic and diastolic) on the other side. In addition, the findings suggest a correlation between the yaw angle of the head and the overtaking event and a negative correlation between the systolic blood pressure and the distance to the nearest vehicle. Our findings align with our initial hypotheses, particularly concerning the impact of performing a maneuver or experiencing a cautious event, such as overtaking, on heart rate and blood pressure due to the agitation and tension resulting from such events. These results can be the key to implementing a sophisticated safety system aimed at maintaining the driver's stable state when aggressive external events or maneuvers occur.

A dynamic method to predict driving risk on sharp curves using multi-source data

Traffic accidents are likely to occur on sharp curves under poor driving conditions, and the severity level of such accidents is high. Therefore, predicting the risk associated with driving on curved roadways in real time can effectively improve driving safety. This paper aims to develop a dynamic real-time method that fuses multiple data sources to predict risk when driving on sharp curves in the context of the connected vehicle environment. Six curves with three small radii (60 m, 100 m, 150 m) and two driving directions (left and right) were designed for a driving simulation experiment. Driver maneuvering data, vehicle kinematic data, and physiological data of 55 drivers were collected for this study. The data were combined and spatially and dynamically segmented. The mean value of the critical lateral acceleration of the vehicle was set as the risk assessment index. K-means clustering was used to classify the driving risk into three levels: low, medium, and high. Then, the risk level was predicted using the maneuvering data, vehicle kinematic data, and physiological data as well as road alignment characteristics as input features for the proposed model that employs the long short-term memory (LSTM) network algorithm. Models with different combinations of observation window (lookback) and interval window (delay) were compared to derive the best window combination. The algorithms selected for comparison against the LSTM algorithm are random forest, XGBoost, and LightGBM. The results show that the proposed LSTM-based method can effectively predict dangerous driving behavior on sharp curves. The optimal window combination derived using the LSTM algorithm is lookback = 20 m and delay = 20 m. The prediction performance of the proposed model is significantly better than that of the other three compared algorithms, with F1-scores of 84.8% and 86.0% for the medium and high risk categories, respectively. In addition, the proposed LSTM-based model that fuses multiple data sources is proven to outperform the model that uses only vehicle kinematics data. The dynamic prediction method proposed in this paper can contribute to the development of a real-time prediction and warning system for driving risks at vehicle terminals in the intelligent connected vehicle environment.

Evaluating overtaking and filtering maneuver of motorcyclists and car drivers using advanced trajectory data analysis

The present paper compares motorized two-wheeler (MTW) and passenger car’s interactions with the rest of the traffic in urban roads while performing overtaking and filtering maneuvers. To better understand filtering maneuvers of motorcyclists and car drivers, an attempt was made to propose a new measure, i.e. pore size ratio. Additionally, the factors affecting lateral width acceptance for motorcyclists and car drivers while overtaking and filtering were studied using advanced trajectory data. A regression model was developed to predict the significant factors affecting motorcyclist’s and car driver’s decisions to accept lateral width with the adjacent vehicle while performing overtaking and filtering maneuvers. Finally, a comparative analysis between machine learning and the probit model revealed that, in the present case, machine learning models perform better than the probit model in terms of the model’s discernment power. The findings of this study will help ameliorate the power of existing microsimulation tools.

Analysis of the injury severity of motor vehicle–pedestrian crashes at urban intersections using spatiotemporal logistic regression models

This paper conducted a comprehensive study on the injury severity of motor vehicle–pedestrian crashes at 489 urban intersections across a dense road network based on high-resolution accident data recorded by the police from 2010 to 2019 in Hong Kong. Given that accounting for the spatial and temporal correlations simultaneously among crash data can contribute to unbiased parameter estimations for exogenous variables and improved model performance, we developed spatiotemporal logistic regression models with various spatial formulations and temporal configurations. The results indicated that the model with the Leroux conditional autoregressive prior and random walk structure outperformed other alternatives in terms of goodness-of-fit and classification accuracy. According to the parameter estimates, pedestrian age, head injury, pedestrian location, pedestrian actions, driver maneuvers, vehicle type, first point of collision, and traffic congestion status significantly affected the severity of pedestrian injuries. On the basis of our analysis, a range of targeted countermeasures integrating safety education, traffic enforcement, road design, and intelligent traffic technologies were proposed to improve the safe mobility of pedestrians at urban intersections. The present study provides a rich and sound toolkit for safety analysts to deal with spatiotemporal correlations when modeling crashes aggregated at contiguous spatial units within multiple years.

DIPNet: Driver intention prediction for a safe takeover transition in autonomous vehicles

AbstractFollowing the successful development of advanced driver assistance systems (ADAS), the current research directions focus on highely automated vehicles aiming at reducing human driving tasks, and extending the operational design domain, while maintaining a higher level of safety. Currently, there are high research demands in academia and industry to predict driver intention and understating driver readiness, e.g. in response to a “take‐over request” when a transition from automated driving mode to human mode is needed. A driver intention prediction system can assess the driver's readiness for a safe takeover transition. In this study, a novel deep neural network framework is developed by adopting and adapting the DenseNet, long short‐term memory, attention, FlowNet2, and RAFT models to anticipate the diver maneuver intention. Using the public “Brain4Cars” dataset, the driver maneuver intention will be predicted up to 4 s in advance, before the commencement of the driver's action. The driver intention prediction is assessed based on 1) in‐cabin 2) out‐cabin (road) and 3) both in‐out cabin video data. Utilizing K‐fold cross‐validation, the performance of the model is evaluated using accuracy, precision, recall, and F1‐score metrics. The experiments show the proposed DIPNet model outperforms the state‐of‐the‐art in the majority of the driving scenarios.

Safety-Critical and Flexible Cooperative On-Ramp Merging Control of Connected and Automated Vehicles in Mixed Traffic

Cooperative on-ramp merging control for connected and automated vehicles (CAVs) can effectively improve traffic throughput and vehicle fuel efficiency at highway on-ramp merging bottlenecks. However, in the mixed traffic scenario where CAVs and human-driven vehicles (HDVs) coexist, the uncertain maneuvers of human drivers pose a major challenge to merging control in terms of safety and flexibility. To this end, this paper proposes a hierarchical cooperative on-ramp merging control strategy for CAVs to optimize flexible trajectories with safety guarantees in mixed traffic. First, the on-ramp merging control problem for CAVs is considered in the case of a three-vehicle coordination, resulting in an optimal control problem (OCP) coordinating on-ramp and main-lane CAVs for efficient operation while satisfying multiple safety-critical constraints. Second, a two-level hierarchical control architecture is developed to solve the OCP with mixed state-control constraints. The upper-level planner solves an unconstrained OCP with Pontryagin’s Minimum Principle to calculate an expected merging position, which is embedded in the variable time headway of safe merging constraints in the lower-level controller. Then, the controller converts the nonlinear OCP with safety-critical constraints to a quadratic programing (QP) problem by exploiting Control Barrier Functions (CBFs) and Control Lyapunov Functions (CLFs). By solving the QP efficiently, the time and energy efficient trajectory for each CAV is obtained. In addition, a receding horizon control framework is employed, which enables CAVs to determine flexible merging opportunity and tackle the disturbances caused by HDVs. Finally, comprehensive simulation results show that the proposed cooperative on-ramp merging strategy has potential in enabling merging flexibility, improving traffic efficiency and energy economy in real time.

Driver Maneuver Detection and Analysis Using Time Series Segmentation and Classification.

The current paper implements a methodology for automatically detecting vehicle maneuvers from vehicle telemetry data under naturalistic driving settings. Previous approaches have treated vehicle maneuver detection as a classification problem, although both time series segmentation and classification are required since input telemetry data are continuous. Our objective is to develop an end-to-end pipeline for the frame-by-frame annotation of naturalistic driving studies videos into various driving events including stop and lane-keeping events, lane changes, left-right turning movements, and horizontal curve maneuvers. To address the time series segmentation problem, the study developed an energy-maximization algorithm (EMA) capable of extracting driving events of varying durations and frequencies from continuous signal data. To reduce overfitting and false alarm rates, heuristic algorithms were used to classify events with highly variable patterns such as stops and lane-keeping. To classify segmented driving events, four machine-learning models were implemented, and their accuracy and transferability were assessed over multiple data sources. The duration of events extracted by EMA was comparable to actual events, with accuracies ranging from 59.30% (left lane change) to 85.60% (lane-keeping). Additionally, the overall accuracy of the 1D-convolutional neural network model was 98.99%, followed by the long-short-term-memory model at 97.75%, then the random forest model at 97.71%, and the support vector machine model at 97.65%. These model accuracies were consistent across different data sources. The study concludes that implementing a segmentation-classification pipeline significantly improves both the accuracy of driver maneuver detection and the transferability of shallow and deep ML models across diverse datasets.

Identification of Drivability Failures Using Adaline in an ECU

Abnormalities in fuel injection systems affect the drivability of vehicles, blemishing the driver's maneuvering experience or the smoothness of the response of these vehicles under different operating conditions. With the technological advances achieved by the automobile industry, objective methods for detecting anomalies in vehicle drivability have been studied over the last few years, with emphasis on methodologies using time-frequency analysis with wavelet transform. When performing an extraction of characteristics through wavelet decomposition aiming to detect abrupt transient variations, it becomes possible to improve the drivability of a vehicle identifying the occurrence of failures by using the energy of the decomposed signal. Therefore, using the concepts of continuous wavelet transform and entropy of information, this work makes a time-frequency analysis of the rotation signal of an internal combustion engine. The samples collected from the motor are standardized, the continuous wavelet transform is calculated and, finally, the entropy of the transformed signal is measured. Thus, the possibility of implementing an Adaline model capable of detecting the presence, or not, of abrupt changes in these signals is verified, and later, it can be embedded in an electronic control unit (ECU). The results show that the use of the Log Energy entropy as an input of the Adaline model is promising, granting 100% of accuracy on the dataset studied.

WFID: Driver Identity Recognition Based on Wi-Fi Signals

Driver identification is a key factor in attributing liability for car accident insurance claims and assessing driver competency. Existing driver recognition systems use mechanisms based on identity keys (e.g., car keys and identity cards) or biometric characteristics (e.g., fingerprints, voiceprints, and face recognition). However, identity keys are prone to loss or misappropriation; biometric methods are prone to driver substitution and raise issues pertaining to privacy; and neither approach is applicable to the majority of commercial applications (e.g., hiring delivery drivers and renting out vehicles). This paper presents a novel driver identity recognition system based on the channel state information (CSI) of Wi-Fi signals, which tend to vary with the user, even when performing identical tasks. CSI values corresponding to driver maneuvers (e.g., turning or going straight) are used as inputs for a deep neural network tasked with establishing a driver recognition model. The feasibility of this approach was verified through simulations in the laboratory and with a vehicle, both of which achieved average recognition accuracy of roughly 95%.

An Integrated Scheme for Coefficient Estimation of Tire–Road Friction With Mass Parameter Mismatch Under Complex Driving Scenarios

The accurate knowledge of tire–road friction coefficient (TRFC) contributes to the optimization of driver maneuvers for further improving the safety of intelligent vehicles. The performance of the existing estimation methods would decline when a vehicle performs complex driving maneuvers. In addition, the mass parameter mismatch also deteriorates the estimation accuracy of TRFC. To address these problems, in this article, an integrated scheme for TRFC estimation is proposed by combining a strong tracking unscented Kalman filter and an interactive multiple model unscented Kalman filter. Real-time experiments are implemented on a mass-produced vehicle to demonstrate the feasibility and effectiveness of the proposed method. The results show that the proposed approach has better estimation accuracy than the existing ones under various driving scenarios.

Analysis of the Structure of Driver Maneuvers in Different Road Conditions

The safety of road users is one of the priority issues taken into account in both the operation and design of vehicles. The presented work is part of a study that aims to develop a method for parametric assessment of driver behavior. The driving style of a driver depends on their skills and psychophysical characteristics, the type and performance of the vehicle used by the driver, and the type of road. This method involves the continuous measurement of the longitudinal and lateral acceleration values of a vehicle body. The paper analyzes how the type of road influences the structure of the maneuvers undertaken by the driver. The paper formulates criteria for distinguishing basic maneuvers (acceleration, braking, and turning). The structure of maneuvers was analyzed for two parameters: the extreme value of acceleration occurring during the execution of a given maneuver and the frequency of maneuvers during the passage of a given route. The analysis presented in this paper confirms that the type of road has a significant influence on the structure of the maneuvers undertaken by the driver.

A graph method of description of driving behaviour characteristics under the guidance of navigation prompt message

AbstractTo verify whether a graph is suitable for describing driver behaviour performance under the effects of navigation information, this study applies two types of prompt messages: simple and detailed. The simple messages contain only direction instructions, while the detailed messages contain distance, direction, road and lane instructions. A driving simulation experiment was designed to collect the empirical data. Two vehicle operating indicators (velocity and lateral offset), and two driver manoeuvre indicators (accelerator power and steering wheel angle) were selected, and T-test was used to compare the differences of behavioural performance. Driving behaviour graphs were constructed for the two message conditions; their characteristics and similarities were further analysed. Finally, the results of T-test of behavioural performance and similarity results of the driving behaviour graphs were compared. Results indicated that the two different types of prompt messages were associated with significant differences in driving behaviours, which implies that it is feasible to describe the characteristics of driving behaviours guided by navigation information using such graphs. This study provides a new method for systematically exploring the mechanisms affecting drivers’ response to navigation information, and presents a new perspective for the optimisation of navigation information.

Driver Maneuver Identification with Multi-Representation Learning and Meta Model Update Designs

Driver maneuver identification (DMI), i.e., the task of predicting the driver maneuver classes given sensor measurements (such as IMU sensors from mobile devices), can serve as a key enabler for many ubiquitous driver behavior analysis applications. Despite the prior studies, two major technical challenges remain before an effective DMI system can be deployed: (i) latent complex maneuver behavior feature relations for accuracy enhancement, and (ii) inconsistency and uncertainty in model adaptation given dynamic data collection settings (e.g., given different urban environments, drivers, and mobile devices). To address the aforementioned challenges, we propose MetaDMI, a novel and adaptive driver maneuver identification framework based on multi-representation learning and meta model update, with the case study on the inertial measurement unit (IMU) sensor measurements (i.e., accelerometer and gyroscope). Specifically, we first extract multiple feature representations for each driver maneuver record in the forms of graph, spectral, and time sequence. Next, MetaDMI processes them with a novel multi-representation learning network, extracting complex patterns and feature relations from the driver maneuvers. Finally, to further enhance the adaptivity of our DMI model to external impacts with dynamic data collection, we have designed a regularized meta learning-based training method to regularize the knowledge transfers across the source and target datasets (e.g., across cities/devices, and few-shot initialization) for consistent and robust identification performance. We have conducted extensive experimental studies upon our MetaDMI prototype based on two datasets (one is collected on our own) and shown that our approach outperforms other baseline approaches for DMI in terms of accuracy and adaptivity.

Learning the representation of surrogate safety measures to identify traffic conflict

Traffic conflict can be identified by the presence of evasive actions or the amount of temporal (spatial) proximity measures like time-to-collision (TTC). However, it is not enough to use only one kind of measures in some scenarios and it is hard to set a threshold for those measures. This paper proposed a method to identify traffic conflict by learning the representation of TTC and driver maneuver profiles with deep unsupervised learning and clustering the representations into traffic conflict and non-conflict clusters. We first trained a transformer encoder to encode sequences of surrogate safety measures into some latent space with unsupervised pre-training. Second, we identified informative clusters in the latent space by calculating the statistic summaries and visualizing trajectory pairs of each cluster. Some clusters are interpreted as traffic conflict clusters because they have small TTC, large deceleration rate and intertwining trajectories and they can be further interpreted as rear-end or angle conflicts. Moreover, the identified traffic conflicts contain critical conditions from the two vehicles in an interaction and one vehicle perceives them as abnormal and takes evasive action to avoid crashes.

B-GAP: Behavior-Rich Simulation and Navigation for Autonomous Driving

We address the problem of ego-vehicle navigation in dense simulated traffic environments populated by road agents with varying driver behaviors. Navigation in such environments is challenging due to unpredictability in agents’ actions caused by their heterogeneous behaviors. We present a new simulation technique consisting of enriching existing traffic simulators with behavior-rich trajectories corresponding to varying levels of aggressiveness. We generate these trajectories with the help of a driver behavior modeling algorithm. We then use the enriched simulator to train a deep reinforcement learning (DRL) policy that consists of a set of high-level vehicle control commands and use this policy at test time to perform local navigation in dense traffic. Our policy implicitly models the interactions between traffic agents and computes safe trajectories for the ego-vehicle accounting for aggressive driver maneuvers such as overtaking, over-speeding, weaving, and sudden lane changes. Our enhanced behavior-rich simulator can be used for generating datasets that consist of trajectories corresponding to diverse driver behaviors and traffic densities, and our behavior-based navigation scheme can be combined with state-of-the-art navigation algorithms.