Deep representation and stereo vision based vehicle detection

Abstract

Vision based vehicle detection is the key component of intelligent transportation technology. Monocular vision based technology is with the shortage of high false detection rate while stereo vision based technology is with the shortcomings of long time-consuming for depth map calculation. Focus on this issue, a monocular and binocular vision based vehicle detection and tracking algorithm is proposed. Firstly, a Deep Convolutional Neural Networks (DCNN) is trained to search for the whole area of image so that vehicle hypothesis can be generated in a short time. Then the dense disparity map and UV disparity map only in the area containing vehicle hypothesis are calculated with binocular vision. By analyzing in the UV disparity map, false detection is eliminated and accurate vehicle position in image coordinate as well as world coordinate is maintained. Experiment results demonstrate that the proposed vehicle detection algorithm is with the merits of both monocular and stereo vision based method and is with high application value.