Localization and tracking of same color vehicle under occlusion problem

Abstract

This paper proposes a vision based multiple vehicle detection and tracking system. Vehicle tail light information is used to localize vehicle potential region, then each candidate is verified by a back propagation neural network (BPNN) trained by Gabor feature set. In the multiple vehicle tracking stage, multiple scale vehicle tracking, same color vehicle occlusion and observation model updating problem are investigated. Mean shift algorithm is main part of tracking sub-system; each detected vehicles are tracked by a mean shift tracker in parallel. Vehicle tail light pairs which are determined in vehicle detection step are used to adjust tracking windows size. Color information is observation model in tracking algorithm, which is insensitive to different color vehicle occlusion problem; mean while, only color information based tracking algorithm can't deal with same color vehicle occlusion case. To overcome this problem, local edge image projection technique is implemented. In experiments, the result shows the proposed system have good performance in multiple vehicle detection and tracking system in the daytime, the results shows 84% accuracy detection rate, 5254 frames image sequences are texted for multip le vehicle tracking.