Incrementally perceiving hazards in driving

Abstract

Abstract Perceiving hazards on road is significantly important because hazards have large tendency to cause vehicle crash. For this purpose, the feedbacks of more than one hundred drivers with different experience for safe driving are gathered. The obtained feedbacks indicate that the irregular motion behaviour, such as crossing or overtaking of traffic participants, and low illumination condition are highly threatening to drivers. Motivated by that, this paper fulfills the hazards detection by involving motion, color, near-infrared, and depth clues of traffic scene. Specifically, an incremental motion consistency measurement model is firstly built to infer the irregular motion behaviours, which is achieved by incremental graph regularized least soft-threshold squares (GRLSS) incorporating the better Laplacian distribution of the noise estimation in optical flow into the motion modeling. Second, multi-source cues are adaptively weighted and fused by a saliency based Bayesian integrated model for arousing driver’s attention when potential hazards appears, which can better reflect the video content and select the better band(s) for hazards prediction in different illumination conditions. Finally, the superiority of the proposed method relating to other competitors is verified by testing on twelve difficult video clips captured by ourselves, which contain color, near-infrared and recovered depth simultaneously and no registration or frame alignment is needed.