Towards Semantic Understanding of Surrounding Vehicular Maneuvers: A Panoramic Vision-Based Framework for Real-World Highway Studies

Abstract

This paper proposes the use of multiple low-cost visual sensors to obtain a surround view of the ego-vehicle for semantic understanding. A multi-perspective view will assist the analysis of naturalistic driving studies (NDS), by automating the task of data reduction of the observed sequences into events. A user-centric vision-based framework is presented using a vehicle detector and tracker in each separate perspective. Multi-perspective trajectories are estimated and analyzed to extract 14 different events, including potential dangerous behaviors such as overtakes and cut-ins. The system is tested on ten sequences of real-world data collected on U. S. highways. The results show the potential use of multiple low-cost visual sensors for semantic understanding around the ego-vehicle.