Moving Human Tracking In Video Image Sequence

In recent years, public security issues are becoming increasingly prominent in the world. Therefore, automatic video surveillance technology has been widely researched and applied in both military and civil fields, such as national defense, aviation, marine, medical and sensitive locations, and so on. Object tracking in video is one important research hot point of automatic video surveillance and it can provide video content understanding, such as object classification and behavior analyses, with important information. However, due to dynamic changes of the background image,such as weather, light, shadow, target gesture nonlinear deformation, and occlusion between objects and background, the multi-target tracking in complex environment is still facing difficulties, and the designing of multi-target tracking algorithm is still the challenges.In this paper, the research is focused on the multi-target tracking using data-driven the probability hypothesis density (PHD) particle filter algorithm with important function of particle filter updated using target detection cancroids. Firstly, classical target detection methods are reviewed; the Lehigh Omni-directional Tracking System (LOTS) and HOG detection algorithm are introduced and validated by simulation experiments. Second, we study the multi-target tracking using data-driven the probability hypothesis density particle filter algorithm, and the HOG and LOTS detection algorithm respectively are combined with PHD particle filter to track multi-target video. Experiment results show that the tracking algorithm performance is over dependent on the detection results. In order to solve this issue above, an improved algorithm is presented, with the intermediate output of detector and the combination of the KNN classifier online trained and the frame difference, to reduce the effect of detection on tracking. Performance analyses are given by the VEPER, a performance evaluation platform for object detection and tracking. The results show that the presented method can track the more human targets, has a higher tracking accuracy, and can keep each target's track well under the condition with occlusion, illumination changes significantly.