| Multiple pedestrian target detection and tracking technology has important research value and application value in many areas of intelligent video surveillance and modern defense. Due to the factors such as complexity of the scene and the pedestrian target variability, developing a robust video tracking algorithm faces enormous challenges in a complex background environment. Based on the previous studies,this paper developed a multi-pedestrian tracking system, solving some problems of tracking.First, for the section of pedestrian detection, this paper studied the detection algorithms based on static images and dynamic videos. For the detection based on static images, by the study of HOG features and LBP features, this paper proposed an improved kind of features of HOG-LBP. Scilicet, linking the two kinds of features together, and then classifying using SVM, by the end, greatly enhancing the probability of success. For the detection based on dynamic videos, through the study of the basic codebook algorithm and the codebook algorithm in the YUV color space, this paper established codebook background model in the YcbCr color space, in the result of running faster and improving detection rate. For the part of pedestrian recognition in binary Figure, this paper used the algorithm of vertical projection histogram that can split block targets.Then, for the part of pedestrian tracking, this paper compared the tracking results of three common kinds of tracking algorithms, what are meanshift, camshift and Kalman filter. By analyzing the results of their principles, the three algorithms all derived limits on the background environment: algorithms meanshift and camshift restrictions on the low-movement and Kalman filter algorithm was limited to linear, Gaussian case. So, the particle filter algorithm which was used more widely, had more room for improvement and better tracking results, so its improvement by this paper rised up.For particle filter, this article first introduced in detail its basic principles, processes, and the development process. Then,considering the basic particle filter algorithm's observation model was based on simple color information, and more information was on the lack of objective description, resulting low robustness of the algorithm, this paper improved the observation model of the algorithm effectively: not only considderd the color information of the target, adding the object's shape information, but also adjusted the weights of the two features adaptively by fuzzy logic, achieving real-time updates. So the target was more fully and effectively described, and the performance of improved particle filtering algorithm was greatly enhanced, solving the pedestrian target's problem of rotating and deformating. Finally,verify its validity in application by experiments.The last, according to the two above key portion, this paper proposes a multi-pedestrian tracking system. This system is divided into two parts, what are open-loop control part and Closed-loop control part. By running the shadowing operator, if no target occlusion occurs,run the open-loop part. That is, only to run track operator (particle filter) for open loop tracking. If the target occlusion occurs, run the closed-loop part, that is, to run detection operator (HOG-LBP features+SVM classifier) to correct tracking results and update the number of particles of tracking operator, achieving the purpose of feedback. The experimental results showed that multi-pedestrian-target tracking system developed in this paper can accurately and quickly achieve multiple pedestrian target tracking requirements. |
PDF Full Text Request