Selected Publications

A model used for velocity control during car following was proposed based on deep reinforcement learning (RL). A total of 1,341 car-following events extracted from the public Next Generation Simulation (NGSIM) dataset were used to train the model. Results show that the model demonstrated the capability of safe, efficient, and comfortable velocity control in that it 1) has small percentages (8%) of dangerous minimum time to collision values (< 5s) than human drivers in the NGSIM data (35%); 2) can maintain efficient and safe headways in the range of 1s to 2s; and 3) can follow the lead vehicle comfortably with smooth acceleration.
IEEE Transactions on Intelligent Transportation Systems, under review, 2019.

This study uses the technique of deep reinforcement learning to model drivers’ car following behavior. Two thousand car-following periods extracted from the Shanghai Naturalistic Driving Study were used to train the proposed model and compare its performance with that of four traditional car-following models. The proposed model can reproduce human-like car-following behavior with significantly higher accuracy than traditional car-following models, especially in terms of speed replicating. Moreover, the model demonstrates good capability of generalization to different driving situations and can adapt to different drivers by continuously learning.
Transportation Research Part C: Emerging Technologies, 2018.

Five representative car-following models were calibrated and validated with 2,100 urban-expressway car-following periods extracted from the 161,055 km of driving data collected in the Shanghai Naturalistic Driving Study (SH-NDS). The results showed that 1) the full velocity difference model, with a validation error of 24%, performed best in modeling Chinese drivers’ behavior; 2) according to the Intelligent Driver Model (IDM), drivers from the SH-NDS adopted a desired time headway which is only a third of that adopted by drivers from the VTTI 100-Car Study in the US; 3) based on the fundamental diagrams of the IDM model, Chinese drivers adopted shorter following gaps than US drivers, which leads to a higher expressway capacity.
Oral Presentation at the 96th TRB Annual Meeting, Transportation Research Part C, 2017.

A simulator is used to test drivers’ collision avoidance behaviors under different initial headways and different lead vehicle deceleration rates. As situational urgency increase, drivers release the accelerator faster, brake to full braking with less time and brake harder. Transition time between initial throttle release and brake initiation is not affected by initial headway or LV deceleration rate. At low situational urgency, multi-stage braking behavior leads to longer delays from brake initiation to full braking.
Transportation Research Part C: Emerging Technologies, Vol. 71, pp. 419-433, 2016.

A total of 111 brake-only non-collision events were presented in the Tongji University Driving Simulator and drivers’ braking behaviors were used to model their Expected Response Decelerations (ERDs). We found ERDs depended on the interaction of LV deceleration and relative speed. In response to this finding, a non-linear function with an interaction term was combined with a linear function into a piecewise function that accommodated both higher and lower LV deceleration conditions. The applicable domain of the warning onset range was then computed for a wide range of kinematic conditions. Results showed the piecewise function to be a better predictor of ERD than the linear function, and to result in fewer driver rejections of the forward collision warnings.
IEEE Transactions on Intelligent Transportation Systems, Vol. 17, No. 9, pp. 2583-2591, 2016.

This study aimed to investigate the impact of Forward Collision Warning (FCW) systems on drivers’ car-following behaviors. Five data collecting vehicles are equipped with Mobileye® systems, which include an FCW function. Participants drive the instrumented vehicles for two months, with the Mobileye® system not activated for the first month, but activated for the second month. The results of this study show that (1) drivers tended to maintain a longer headway when FCW activated; and (2) the FCW resulted in a 0.13s decrease of reaction time in daytime driving, and a 0.09s decrease when a following vehicle had higher speed than the lead vehicle. Moreover, this study further confirms that the reaction time is affected by relative distance, lead vehicle acceleration, and traffic density.
Proceedings of the 14th World Conference on Transport Research, 2016.

Publications

For a most recent publication list, please refer to my CV.

. Velocity Control for Autonomous Driving with Multi-Objectives: Safety, Efficiency, and Comfort. IEEE Transactions on Intelligent Transportation Systems, under review, 2019.

Project

. Human-Like Autonomous Car-Following Planning by Deep Reinforcement Learning. Transportation Research Part C: Emerging Technologies, 2018.

Project Source Document

. Human-like Autonomous Car-Following Planning by Deep Reinforcement Learning. Presentation at the 97th TRB Annual Meeting, 2017.

Project Project

. An Exploration of Cut-In Behavior and Gap Acceptance Using Shanghai Naturalistic Driving Data. Accepted for Presentation at the 97th TRB Annual Meeting, 2017.

Project

. Impact on Car Following Behavior of A Forward Collision Warning System with Headway Monitoring. Transportation Research Part C: Emerging Technologies, under the first round review, 2017.

Project

. Modeling Car-Following Behavior on Urban Freeways in Shanghai: A Naturalistic Driving Study. Transportation Research Part C: Emerging Technologies, acceptance contingent on minor revision, 2017.

Project

. Calibrating Car-Following Models on Urban Expressways for Chinese Drivers Using Naturalistic Driving Data. Oral Presentation at the 96th TRB Annual Meeting, Transportation Research Part C, 2017.

Project Slides Source Document

. Dimension Reduction and Multivariate Analysis of Variance for Drivers’ Forward Collision Avoidance Behavior Characteristics. Journal of Tongji University, Vol. 44, No. 12, pp. 1858-1866, 2016.

PDF Project

. Impacts of Collision Warning System on Car following Behavior Based on Naturalistic Driving Data. Journal of Tongji University, Vol. 44, No. 7, pp. 1045-1051, 2016.

PDF Project

. Impacts of Situational Urgency on Drivers’ Collision Avoidance Behaviors. Journal of Tongji University, Vol. 44, No. 6, pp. 876-883, 2016.

PDF Project

. Drivers’ Rear End Collision Avoidance Behaviors under Different Levels of Situational Urgency. Transportation Research Part C: Emerging Technologies, Vol. 71, pp. 419-433, 2016.

PDF Project Source Document

. Car-following Headways in Different Driving Situations: A Naturalistic Driving Study in China. Proceedings of the 5th International Symposium on Naturalistic Driving Research, 2016.

Project

. Development of a kinematic-based forward collision warning algorithm using an advanced driving simulator. IEEE Transactions on Intelligent Transportation Systems, Vol. 17, No. 9, pp. 2583-2591, 2016.

PDF Project Source Document

. Impact of a Forward Collision Warning System on Headway and Reaction Time during Car Following. Proceedings of the 14th World Conference on Transport Research, 2016.

PDF Project

. Car-following Headways in Different Driving Situations: A Naturalistic Driving Study. Oral Presentation at the 16th COTA International Conference of Transportation Professionals, 2016.

PDF Project

Projects

Shanghai Naturalistic Driving Study

The first large-scale naturalistic driving study in China. A total of 161,055 km of real-world detailed driving data were collected from 60 Chinese drivers, providing an unprecedented opportunity for investigating driving behavior in China.

AlphaDrive: Mastering Driving with Real-World Driving Scenarios

AlphaDrive is a platform equipped with real-world driving scenarios where intelligent agents can learn to drive by trial and error, trials and errors that number in the billions.

Driving Behavior Research for Intelligent Collision Avoidance Technology

A driving simulator study jointly conducted by Tongji University and China First Automobile Work Corporation. Tongji University’s eight-degree-freedom driving simulator (its fidelity ranking top 3 in the world) was used. Its primary purpose is to improve the understanding of drivers’ collision avoidance behavior under different rear-end scenarios and to develop an effective forward collision warning strategy.

Udacity Self-Driving Car Engineer Nanodegree

Udacity’s Self-Driving Car Engineer Nanodegree Program offered 12 autonomous-driving-related projects and covered topics including deep learning, computer vision, sensor fusion, localization, controllers, vehicle kinematics, automotive hardware.

Autonomous Car Following by Deep Reinforcement Learning

This on-going project aims to develop autonomous car-following strategies based on deep reinforcement learning.The demo animation shows the car-following learning process of the RL agents (cars). In early training episodes, cars become red frequently, indicating penalties caused by bad performances in car following, e.g., rear-end crashes. As training converges, the RL agents maintain steady car-following headways and receive fewer penalties.

End-to-End Learning for Steering Control

This on-going project is inspired by NVIDIA’s research (1) and MIT 6.S094’s DeepTesla project (2). It aims to learn human drivers’ strategy in steering control. The learned model can map raw pixels from a single front-facing camera directly to steering commands.

Human-Like Driving by Inverse Reinforcement Learning

This on-going project is inspired by Levine et al. (2012). It aims to imitate human driving patterns. The assumption is that drivers act based on a utility (reward) function, which represents the preference of the driver and elicits the driving behavior. The role of inverse reinforcement learning is to infer or discover the latent utility function from driver (expert) demonstrations and thus generalize the driving policy to unobserved situations.

Motion Segmentation Based on Subspace Clustering

This project aimed to segment different moving objects in a video sequence, using subspace clustering methods.

Traffic Parameter Analysis Platform based on Unmanned Aerial Vehicle

This project aimed to extract real-time traffic flow parameters, including volume, density, and intersection delay, etc., based on videos captured by an unmanned aerial vehicle. The featured animation demonstrates the extraction of traffic-flow parameters by setting virtual loop detectors.

Traffic Simulation based on Cellular Automaton Method

This project aimed to model the keep-right-except-to-pass rule using Cellular Automaton. The cellular automaton system is a discrete dynamic system consisting of a regular grid of cells, each in one of a finite number of states. And the state of each cell is updated every single discrete time step according to the updating rules, which are determined by the current state of the cell and the states of its neighbor cells.

Optimized Design for Combined Road Alignment

Evaluating the safety performance of combined horizontal and vertical alignments in mountainous freeways, to guide the design of safer mountainous freeways.

Research

Table of Contents

Driving Behavior

Normal Driving Behavior

Shanghai Naturalistic Driving Study collected 60 Chinese drivers’ daily driving data, with a total mileage of 161,055 km. These data provide an unprecedented opportunity for investigating driving behavior in China.

Car-Following Behavior

Extracted 108,933 car-following events, and investigated the distribution of drivers’ car-following gaps and headways.

Lane-Change Behavior

Extracted 17,309 lane-change events, and investigated the frequency, motivations, gazing behavior, and turn-signal usages for lane-change behavior.

  • Shanghai drivers averagely performed 0.645 lane changes per kilometer, higher than US drivers (0.224 lane changes per kilometer) (Olsen, 2004).
  • Lane-change behavior occurred most frequently (0.82 lane changes per kilometer) on urban arterial roads, compared with urban expressways, freeways, and suburban roads.

Cut-In Behavior

Extracted 7,845 cut-in events, and analyzed cut-in purposes, turn signal usages, cut-in duration and urgency. (Paper: An Exploration of Cut-In Behavior and Gap Acceptance Using Shanghai Naturalistic Driving Data)

  • Cut-in behavior is relatively dangerous and risky with smaller time to collision than normal lane change, and more than 50% of cut-ins are motivated by a slow preceding vehicle.
  • Almost half of Chinese drivers did not use a turn signal when cutting-in, which is indicative of poor driving habits and an aggressive driving style.
  • Unlike a typical lane change, cut-ins have a shorter duration as well as a smaller lag gap.
  • Road type, relative speed, and following vehicle’s acceleration are important factors that might influence drivers’ lag gap acceptance.

Collision Avoidance Behavior

In Driving Behavior Research for Intelligent Collision Avoidance Technology, a high fidelity driving simulator was used to examine the effects of differing levels of situational urgency on drivers’ collision avoidance behaviors. (Paper: Drivers’ Rear End Collision Avoidance Behaviors under Different Levels of Situational Urgency)

  • As situational urgency increased, drivers released the accelerator and braked to maximum more quickly.
  • The transition time between initial throttle release and brake initiation was not affected by situational urgency.
  • At low situational urgency, multi-stage braking behavior led to longer delays from brake initiation to full braking.
  • These findings show that effects of situational urgency on drivers’ response times, braking delays, and braking intensity should be considered when developing forward collision warnings systems.

Distracted Driving

Based on the Shanghai Naturalistic Driving Study data, drivers’ phone-use characteristics and the influence of phone use on driving performance were examined.

  • A total of 943 phone-use events were extracted from 53 drivers’ data, and 3,865 eyes-off-road cases and 3,131 sub-tasks were extracted from those events.
  • A hierarchical coding structure for phone-use event was built, and each change in every driver’s visual behavior, manual behavior or subtask was recorded.
  • The influences of phone use on longitudinal and lateral driving performance were studied especially the speed adaptation behavior.

Vehicle Active Safety

Developing Forward Collision Warning System

Evaluating Forward Collision Warning System

The impact of the FCW system on drivers’ car-following behaviors was investigated. (Paper: Impact of a Forward Collision Warning System on Headway and Reaction Time during Car Following)

  • In Shanghai Naturalistic Driving Study, five data-collecting vehicles were equipped with Mobileye® systems, which include an FCW function.
  • Participants drove the instrumented vehicles for two months, with the Mobileye® system not activated for the first month, but activated for the second month.
  • Drivers tended to maintain a longer headway when FCW activated.
  • The FCW resulted in a 0.13s decrease of reaction time in daytime driving, and a 0.09s decrease when the following vehicle had higher speed than the lead vehicle.
  • The reaction time during car following was affected by relative distance, lead-vehicle acceleration, and traffic density.

Optimized Road Design

In Optimized Design for Combined Road Alignment, the effects of combined alignments on driving behavior (e.g., speed and lateral acceleration) and safety performance (crash surrogate measures like lane departure events and emergency braking events) were examined using driving simulator and field data. The results will assist engineers to design safer mountainous freeways.

Traffic Flow Modeling and Simulation

Car-Following Model

The performances of existing car-following models applied to Chinese drivers were evaluated. (Paper: Calibrating and Validating Car-Following Models on Urban Expressways for Chinese Drivers Using Naturalistic Driving Data)

  • Five representative car-following models (Gaxis-Herman-Rothery, Gipps, intelligent driver, full velocity difference, and Wiedemann) were calibrated and validated with 2,100 urban-expressway car-following periods extracted from the 161,055 km of driving data collected in the Shanghai Naturalistic Driving Study.
  • The full velocity difference (FVD) model had the lowest error term on validation dataset (21%), and the smallest standard deviation of errors.
  • Compared to the Wiedemann model used by VISSIM®, the FVD model is more easily calibrated and demonstrates a higher and more robust performance, justify its suitability to be applied for microscopic traffic simulations in China.

Traffic Simulation Based on Cellular Automata

To evaluate the performance of the keep-right-except-to-pass rule in traffic systems, the cellular automaton method was applied to simulate traffic flows in one-lane, two-lane and three-lane freeways. The results showed that the model could replicate volume-density relationship well.

Autonomous Driving

Autonomous Car-Following

In Autonomous Car Following by Deep Reinforcement Learning, two types of autonomous car-following strategies were developed based on deep reinforcement learning (RL).

Human-Like Car Following

Historical driving data are fed into a simulation environment, where an RL agent learns from trying and interaction, with a reward function signaling how much the agent deviates from the empirical data. Results showed that this new model can reproduce human-like car-following behavior with significantly higher accuracy than traditional car-following models, especially regarding speed replicating. (Paper: Human-like Autonomous Car-Following Planning by Deep Reinforcement Learning)

  • The model has a validation error of 18% on spacing and 5% on speed, which is 15 and 30 percentage points less, respectively, than that of traditional car-following models.
  • The model demonstrates good capability of generalization to different driving situations and can adapt to different drivers by continuously learning.

Safe, Efficient, and Comfortable Car Following

To fulfill the multi-objectives of car following, a reward function reflecting driving safety, efficiency, and comfort was constructed. With this reward function, the RL agent learns to control vehicle speed in a fashion that maximizes cumulative rewards, through trials and errors in the simulation environment. Results showed that the model demonstrated the capability of safe, efficient, and comfortable velocity control. (Paper: Reinforcement Learning based Velocity Control for Autonomous Driving with Multi-Objectives: Safety, Efficiency, and Comfort)

  • The model had small percentages (8%) of dangerous minimum time to collision values (< 5s) than human drivers in the NGSIM data (35%).
  • The model can maintain efficient and safe headways in the range of 1s to 2s and can follow the lead vehicle comfortably with smooth acceleration.

Autonomous Driving in Mixed Traffic Environments

Current autonomous driving decision-making algorithms mainly focus on interactions between passenger cars, while the interactions between large vehicles, non-motor vehicles, and pedestrians are rarely taken into account. However, the mixture of different transport modes is just the critical feature of Chinese urban traffic. Therefore, this on-going research aims to develop and test autonomous driving algorithms for mixed traffic environments. The following steps will be followed:

  • Extract typical mixed-traffic driving scenarios from accident database and naturalistic driving database.
  • Develop autonomous driving decision-making algorithms that are suitable for mixed-traffic driving scenarios.
  • Evaluate the developed algorithms based on a high-fidelity driving simulator.

Recent Posts

Poster presentation on human-like autonomous car following at the 97th TRB annual meeting.

CONTINUE READING

During September 25 and 26, 2017, the Fifth International Symposium on Transportation Safety was held at Tongji University. Over 200 scholars and experts from research institutions and government agencies attended the symposium. Zhu served as a volunteer for the meeting organizing.

CONTINUE READING

Oral presentation on car-following model calibration at the 96th TRB annual meeting.

CONTINUE READING

Attending the top international symposium on naturalistic driving study, one of my major research interests.

CONTINUE READING

Coming across one of the world’s most well-known transportation scientists, Carlos F. Daganzo, at the 16th COTA International Conference of Transportation Professionas.

CONTINUE READING

Awards

Slected Awards and Scholarships

  • National Graduate Scholarship (twice), Ministry of Education, China. Oct 2017, Oct 2016
  • Outstanding Student Award, Tongji University. Oct 2016
  • China Post-Graduate Mathematical Contest in Modeling, Second Prize. Sep 2016
  • “Inspirational Star”, Tongji University. Jun 2015
  • Volvo Group Scholarship, Volvo Group. Dec 2014
  • Excellent Student Scholarship, Tongji University. Nov 2014
  • National Competition of Transport Science and Technology for Students, Second Prize. May 2014
  • Mathematical Contest in Modeling, Honorable Mention. Jan 2014
  • National Endeavor Fellowship (twice), Ministry of Education, China. Nov 2012, Nov 2013
  • The Second Prize of East China Mathematical Modeling Contest. Oct 2013
  • China Undergraduate Mathematical Contest in Modeling, Second Prize. Sep 2013
  • The Grand Prize of Tongji University Mathematical Modeling Contest. Jun 2013
  • China National Undergraduate Physics Contest, Shanghai Division, Third Prize. Dec 2012
  • First Class Excellent Student Scholarship, Tongji University. Nov 2012

Contact

  • meixin92@uw.edu
  • +12066987121
  • Department of Civil and Environmental Engineering, University of Washington More Hall 101, Box 352700, Seattle, WA 98195

Website Stats