Detecting pedestrians from a video by stereo cameras

The technologies for safe driving assistance are of particular importance to escape pedestrians from traffic accidents. We have been doing research on detecting pedestrians by use of stereo cameras on an automobile and giving them priorities according to their degree of danger by analyzing their motions/behaviors. In this particular paper, we present a technique for detecting pedestrians employing stereo camera images. From a disparity map, each pixel on the image is voted on a depth image employing a 2-D Gaussian distribution and the region having a peak value of the vote is chosen as an object region. This region is represented by HOG features and judged if it is a human by SVM. Experimental results show satisfactory performance of the technique.