Study On Pedestrain Detection Based On Depth Information

In the intelligent monitoring system, the pedestrians in the video are the main objects of monitoring, so pedestrian detection technology plays an important role in intelligent transportation, security and other fields. Most of traditional pedestrian detection methods based color image detect and recognize human by the means of extracting some related features from color images. Usually a human perception of the world is three-dimensional, but the color image lost depth information, so it is difficult to segments moving objects directly from the color image. Although many existed pedestrian detection methods extracted relevant features from the color information in the image, and achieved good results, but these methods are difficult to overcome the interference caused by illumination change and the similar color between target and background.In this thesis,two methods based depth image were presented for pedestrian detection, and the two methods are in vertical downward control perspective. The main idea of the first method is to identify the target by detecting pedestrian head. First,extract the outline of the target from the depth image,then use the Hough transform to detect the circles in the outline image. These circles are regarded as head-like region,so some priori knowledge were needed to distinguish the circles which are not heads and then remove it. Upon detection of pedestrians, using Camshift algorithm can realize precise tracking of pedestrians according to clustering characteristics of pedestrian head depth information. Experiments show that, compared to the pedestrian detection method using color feature, this method has better detection accuracy. The second method takes use of the feature of head-shoulder region for pedestrian detection according to the fact that the depth of head-shoulder region in the depth image shows a peak valley feature. In this paper,the first step is to capture enough samples and then extract Haar feature of head-shoulder,finally use the Adaboost algorithm to train the head-shoulder classifier. Experiments show that, this method can not only overcome the interference of similar color, but also can accurately distinguish pedestrian and non-pedestrian moving target.