Pedestrian detection using stereo night vision

A method is presented for pedestrian detection using a stereo night-vision system installed on the vehicle. Much of the work in this area makes use of shape information. The proposed method detects moving objects whose motions are not consistent with the movement of the background and is considered to be complementary to shape-based approaches. In this paper, two new techniques are introduced for this task for night vision, namely a two-stage method for stereo correspondence and motion detection without explicit ego-motion calculation. These techniques make use of characteristics of night-vision video data, in which humans appear as hotspots. This method works well in cases where the camera motion has a dominant translational motion with a small amount of rotational motion, which is suitable for the camera on the vehicle. Error analysis as well as experimental results are presented to validate our approach and comparisons have been carried out between our approach and frame-by-frame based pattern-recognition approaches.