Understanding drivers' risk behaviors from dashcam videos

Abstract

Dangerous driving causes preventable injuries and deaths. To promote safe driving practices, a forward-facing dashboard camera or dashcam can provide visual context of the environment in which a driver exhibits dangerous driving behaviors. In this thesis, I designed and implemented a system that identifies and analyzes dangerous driving events using monocular videos from dashcams and odometry data from smartphone sensors. To extract useful information such as the position of the ego-vehicle in the lane, the following distance to the next vehicle, and the position and velocity of other vehicles in real-world coordinates, the system performs multiple computer vision tasks, including 2D and 3D object detection and tracking, lane detection, and camera calibration. Then, the information from perception tasks are used to identify risky events, including tailgating and running stop signs, analyze the speed of traffic flow, and classify the cause for hard braking events, such as reacting to the next car in the lane, stopping for a pedestrian, or being inattentive. Using over 8 hours of dashcam videos in multiple driving scenarios, experiments were performed to demonstrate the system's capabilities and limitations.