Research And Implementation Of Moving Pedestrian Detection Algorithm Based On Monocular Vision

Because of the rapid growth of quantities of vehicles and road accidents, how to reduce road accidents has become the focus of attention. As main participants of road traffics, pedestrians also are the direct victims of traffic accidents. Therefore, giving immediately warning before possible collisions between pedestrians and ego vehicle to offer sufficient protection to pedestrians has received considerable attention in the field of Driver Assistance Systems. Pedestrian is non-rigid body, it may appears in a variety of positions in the road scene, together with the variety of clothes texture and weather and light conditions, vision-based pedestrian detection is very challenging.A moving pedestrian detection system including ROI (Regions of Interest) segmenting and recognizing based on monocular vision form the view of practicability is designed in this thesis. In the ROI segmentation process, an ego-motion compensation model is firstly generated based on features matching in sequential frames, and the model is used for eliminating camera moving to generate difference image; then the difference image's vertical projection of intensity is used to detect motion; finally, motion information, together with horizontal edge information are used for ROI locating.In the recognition process, a cascade AdaBoost classifier is generated for pedestrian recognition. This process includes two stages:offline training and online recognizing. In the offline training stage, because of pedestrians'characteristics such as vertical edge and symmetry, Haar-like features are used for training AdaBoost classifier; in the online recognition stage, a cascade structure of AdaBoost classifier is used for recognizing pedestrians from ROI. Finally, recognition results of sequential frames are associated to enhance the stability of the pedestrian detection system based on the continuity of pedestrians in image sequence.A large number of tests are carried out on the test set including difference weather and illumination conditions. Experimental results have shown the effective performance behaved on moving pedestrians.