Research Of Pedestrian Detection And Tracking Based On Sequence Images

Moving object detection and tracking as a hot topic in the field of computer vision and pattern recognition is widely used in video surveillance, human motion analysis, robot navigation, anomaly detection, video conferencing, intelligent transportation, etc. However, the way to achieve a stable, accurate and rapid visual tracking is still a challenge because of the impact factors such as target itself, algorithms and environmental interference, etc. This paper discusses the zenithal pedestrian detection, tracking and counting algorithm based on the specific application with binocular vision sensor Kinect, namely using kinect to obtain the color images and the depth images of same scene simultaneously. It can detect the crowed target in intense scene after a comprehensive analysis of the color information and depth information, finally the tracking and counting of the zenithal pedestrian detection can be realized. The main research work is as follows:First, images are preprocessed, including color image correction, depth image inpainting, moving targets extraction and morphological processing. For the mismatch problem between color images and depth images in the space position, three-to-three affine transformation is used to compensate for the differences caused by the different spatial distribution of kinects’ three lens. Aiming at the problem of the presence of many uncertain pixels in depth image, the improved joint bilateral filter is applied to fix the depth image, which is to get close to the true accurate depth image. Then the background subtraction algorithm is devoted to extract the targets and the morphological processing is used to ensure the accuracy of moving target detection and tracking.Second, according to the characteristics of zenithal pedestrian in depth images, the non-maxima suppression algorithm(NMS), random hough transform(RHT) and fast radial symmetry transform(RST) are devoted to detect the zenithal pedestrian respectively. In the meanwhile, an approch of heads segmentation based on local depth image histogram is presented, and for the head, its geometry center that is treated as the center of the circle is transformed according to its expanded figure based on the radius. In this paper, an innovative discriminant arithmetic of combining with the HOG feature descriptors of the radius unfold transform diagram and SVM classifier is put forward. Finally, the accurate head position is obtained.Then, a traditional data correlation matching algorithm based on Kalman Filter and K-Nearest Neighbor is introduced. And on this basis, according to the detection results of zenithal pedestrian, an improved data correlation matching algorithm based on multi-feature fusion and a pedestrian counting algorithm based on state space transformation are presented, thus realizing the zenithal pedestrian tracking and counting. Based on the above tracking results, an improved discretecontinuous optimization algorithm for multi-target tracking is put forward, which can significantly improve the tracking results and obtain more accurate trajectories.Finally, a simple tracking demonstration system based on microsoft visual studio 2013 and open source computer vision library Opencv2.4.9 is designed, including pedestrian detection module, pedestrian tracking and counting module, which further verifies the feasibility of our algorithm. Then, the deficiency of our algorithm and the future research direction is analyzed.This algorithm is not affected by the problem such as illumination change, shadow interference, and avoid mutual occlusion by taking zenithal sample, so the drawbacks of traditional multi-target detection method can be overcome. Experimental results show that using the above methods to proceed with zenithal pedestrian tracking and counting, accurate results have been achieved, and in solving the inherent problems such as error detection, residual and false match, the effect is obvious. As a result, this algorithm can effectively improve tracking accuracy, stability and credibility.