Intersection-related traffic crashes and fatalities are major concerns for road safety. This project aimed to understand the major causes of conflicts at intersections by studying the intricate interplay between roadway agents. The approach involved using the current traffic camera systems to automatically process traffic video data. As manual annotation of video datasets is a very labor-intensive and costly process, this research leveraged modern computer vision algorithms to automatically process these videos and retrieve kinematic behavior of the traffic actors. Results demonstrated how traffic actors and road segments can be modeled independently via graphs and how they can be integrated into a framework that can model traffic systems. The team used a graph neural network to model (a) the interaction of all the roadway agents at any given instance and (b) their role in road safety, both individually and as a composite system. The model reports a near-real-time risk score for a traffic scene. The study concludes with a presentation of a new drone-based trajectory dataset to accelerate research in intersection safety.
- The project demonstrated the usability of the traffic intersection cameras to automated traffic monitoring and real time safety analysis using advanced computer vision and graph based learning.
- The project used graphs to represent traffic scene. We demonstrated how traffic dynamics, roadway structure, relative position of traffic actors in a traffic scene can be captured in a unified setting. This simplifies analysis of traffic in any given roadway architecture.
- A graph based formulation helped us analyze traffic in node level (traffic actors), edge level (interaction between actors), and graph level (a traffic scene). We demonstrated the usefulness of graph based formulation first through graph structure analysis. This helps us to study vulnerability of each actors including vulnerable road users.
- We used graph neural network (GNN) to determine the risk of each traffic scene. GNNs help to summarize information from all traffic actors including their relative locations, headings, kinematics. We used information for any given instance and demonstrated that real time risk prediction is feasible. We used standard safety metrics including time to collision (TTC) and post encroachment time (PET).
- The code has been released (link)
EWD & T2 Products
Results from this project was used to build a course module in the Virginia Tech course, BMES 5234: Adv Vehicle Safety Systems. The module presents on “Computer Vision and Machine learning in transportation research”. Course modules are available here.
Two Chapters in the master’s thesis of Akash Sonth have been inspired by this project.
The code for the research can be found: (link)
Student Impact Statement(pdf): Three students received funding under this project (Hirva Bhaga, a Master’s of Science student from Virginia Tech, Akash Sonth, a Masters’s of Science student from Virginia Tech, and Sparsh Jain, a Doctor of Philosophy student from Virginia Tech). This file contains statements from all three students as to the impact this project had on education and workforce development.
Additionally, the research team collaborated with Tsinghua University who have already released similar dataset for intersection safety research. The data collected by Virginia Tech will be merged with the dataset (available at https://github.com/SOTIF-AVLab/SinD). This dataset will be first of its kind intercontinental drone dataset for intersection safety analysis. The team along with Tsinghua university is planning for an international competition that will use this dataset. It is expected that the competition will be conducted during IEEE IV 2025.
Sarkar, A. & Sonth, A. (2023, October 24). Real Time Risk Prediction at Signalized Intersection Using Graph Neural Network. Webinar SafeD Virtual Webinar is available here. PDF of PowerPoint slides from Webinar available here.
Sarkar, A. (2023, December). Artificial intelligence in transportation: Paving the way for the future of mobility. Indian Symposium on Machine Learning, Mumbai, India. https://indoml.in/#talk-as
The final datasets for this project are located in the Safe-D Collection on the VTTI Dataverse; DOI: 10.15787/VTT1/BBJGFE.
Research Investigators (PI*)
Start Date: 2022-05-01
End Date: 2023-06-30
Grant Number: 69A3551747115
Total Funding: 320,000
Source Organization: Safe-D National UTC
Project Number: 06-012
Safe-D Theme Areas
Safe-D Application Areas
Operations and Design
Planning for Safety
Office of the Assistant Secretary for Research and Technology
University Transportation Centers Program
Department of Transportation
Washington, DC 20590 United States
Virginia Polytechnic Institute and State University
Virginia Tech Transportation Institute
3500 Transportation Research Plaza
Blacksburg, Virginia 24061