Skip Navigation

Mid-America Transportation Center

Crash Prediction and Avoidance by Identifying and Evaluating Risk Factors from Onboard Cameras



University

Missouri University of Science & Technology

Principal Investigator
Ruwen Qin
PI Contact Information
qinr@mst.edu
Funding Source(s) and Amounts Provided
USDOT: $75,000
Dept. of Engineering Management and Systems Engineering, MS&T: $15,973
Computer Science Dept., MS&T: $36,411
MS&T: $22,616
Total Project Cost
$ 150,000
Agency ID or Contract Number
69A3551747107
Start Date
01/01/2019
End Date
09/30/2020
Brief Description of Research Project
Motor vehicle crashes are a huge concern of roadway transportation safety, resulting in over 37,000 fatalities and $800 million losses annually. In recent years, the number of fatalities is growing. Risk factors that have been traditionally used no longer fully explain causes of the recent increase in fatalities. Human beings have bounded abilities in vision, cognition, making judgment, and simultaneously handling multiple tasks, particularly in complex, dynamic environments or in response to suddenly occurring situations. Therefore, assisting them in the cognition of risks and making the right decisions in a near real-time manner is a particular need in order to advance transportation toward zero fatalities. This project is motivated to developing a data-driven, computer-vision empowered, verifiable system that can predict crashes, and thereby improves drivers’ ability to avoid them. Pursuing a systematic approach, this project seamlessly integrates data analytics, deep learning, and computer vision technology to achieve the goal. Specifically, the project creates a crash report dataset trimmed from FARS, and analyzes the data to identify a set of risk factors that contribute to crashes and assess their significance levels. Provided with these, a Convolutional Neural Network (CNN) for scene segmentation and vehicle detection, and a Multi-tag Classification Network, are trained using the public KITTI dataset without crash accidents and the new dataset of crashes collected from YouTube. With the trained neural network models, videos captured from cameras mounted in vehicles can be analyzed in a near real-time manner, which infers the risk factor values for crash prediction and avoidance. For the purpose of crash prediction, a Long Short Term Memory (LSTM) model is developed to analyze the time-series data of risk factors in all frames along with their corresponding significance levels. The developed system, and the underlying technology and methods, are new capabilities for addressing motor vehicle crashes.
Describe Implementation of Research Outcomes
Impacts/Benefits of Implementation
Web Links
Modal Orientation
  • Safety and Human Performance
  • Systems