Research And Implementation Of Pedestrian Detection In Video Images

Because of the rapid growth of the quality of vehicles, traffic accident frequently happen, which is threatening our society, thereby, how to reduce the traffic accidents and decrease the loss caused by it are attracting the attention of us naturally. Therefore, give immediately warning before possible collision between pedestrian and ego vehicle to offer sufficient protection to pedestrian has already became the main topic of the recent research.In the thesis, by making use of monocular vision detector as the main mean of catching outside environmental information, a pedestrian detection system including segmenting of ROI (Regions of Interesting) and recognizing detection system was built. This study includ ing the following two parts.In the ROI segmentation process, based on the particular characteristic of pedestrian, bring forward pedestrian segmenting method grounding on vertical edge and the symmetry property of it, and then accurately locate and segment the pedestrian from the video image. The segmenting algorithm in this thesis is the supplement and extension of the algorithm based on the moving pedestrian, mainly aiming at two kinds of occasion:1st. the motion direction of pedestrian is parallel to the motion direction of car-carried camera or the angle of them is very mall; 2nd, the pedestrian is static or his velocity is very small.In the recognition process, a linear support vector machine is used for pedestrian recognition. This process includes two stages:offline training and online recognizing. In the offline training stage, according to the characteristics of the local appearance and the shape of the object, this characteristic usually can be well showed by local gray gradient or edge direction distribution, HOG features are used for training linear SVM classifier. In the online recognizing stage, the SVM classifier trained in the last stage is used to recognize pedestrian candidate from the ROI.For the pedestrian detecting algorithm in this study, a large number of tests at different kinds of weather and scenes are carried out. Experimental results showed an effective performance of pedestrian detection and the detection rate was relatively high.