Learning Spatio-Temporal Features via 3D CNNs to Forecast Time-to-Accident

Taif Anjum, Louis Chirade, Beiyu Lin, Apurva Narayan

Abstract: Globally, traffic accidents are one of the leading causes of death. Collision avoidance systems can play a critical role in preventing accidents or minimizing their severity. Time-to-accident (TTA) is considered the principal parameter for collision avoidance systems allowing for decision-making in traffic, dynamic path planning, and accident mitigation. Despite the importance of TTA, the literature has insufficient research on TTA estimation for traffic scenarios. The majority of recent work focuses on accident anticipation by providing a probabilistic measure of an immediate or future collision. We propose a novel approach of time-to-accident forecasting by predicting the exact time of the accident with a prediction horizon of 3-6 seconds. Leveraging the Spatio-temporal features from traffic accident videos, we can recognize accident and non-accident scenes while forecasting the TTA. Our method is solely image-based, using video data from inexpensive dashboard cameras allowing for