A monocular-vision rear vehicle detection algorithm

Abstract

A monocular vision based detection algorithm is presented to detect rear vehicles. Our detection algorithm consist of two main steps: knowledge based hypothesis generation and appearance based hypothesis verification. In the hypothesis generation step, a shadow extraction method is proposed based on contrast sensitivity to extract regions of interest (ROI), it can effectively solve the problems caused by casting shadow and illuminations. In the hypothesis verification step, one improved wavelet feature extraction approach based on HSV space was proposed. Moreover, in order to satisfy different application requirements, a new method based on probability density function is proposed to decide the decision boundary for Support Vector Machine. The algorithm was tested under various traffic scenes at different daytime, the result illustrated good performance.