Research On The Technology Of Multi-pedestrian Detection And Tracking

Intelligent video surveillance is a hot issue in computer vision domain in recent years,which combines computer vision and networked video surveillance to accomplish the taskof target detection, classification, tracking, behavior recognition and so on. Multi-pedestrian detection and tracking is a difficult problem in Intelligent video surveillance. Asnon-rigid bodies, pedestrians’ movements are more flexible than the vehicle and theirvariable contour features are difficult to extract, which has brought many challenges to thetracking accuracy and computational complexity of the algorithm. This paper focuses onthe multi-pedestrian detection and tracking in the field of view of a fixed camera.In terms of moving target detection, the existing moving target detection algorithms,especially the background model methods are analyzed and implemented. Then, based onbackground substraction method, this paper proposed an detection algorithm which is acombination of edge detection and region merging. Experimental results show that thisalgorithm extracts more complete and accurate target regions than traditional backgroundsubstraction and has certain robustness to shadow under low light circumstance.In target tracking, we group the target state into four categories, involving appearanceof a new target, disappearance of an old target, target under occlusion and target match.According to the nonrigid characteristics of pedestrians, center distance, weighted colorhistogram and the overlap ratio of targets are adopted to build dynamic association matrixto acquire target association between frames. When occlusion occurs, we make use ofkalman filter to predict and estimate the location of occluded targets in current frame. Thismethod can attain good performance while pedestrians’ movements are uniform and steady.However, if the whole occlusion process lasts for a long time, predicting and estimating the location of occluded targets during the whole process by using kalman filter will lead to anunignorable cumulative error and tracking failure. In this paper, we proposed an improvedalgorithm to handle occlusion. According to occlusion factor, the occlusion is separatedinto severe occlusion and partial occlusion. If targets are determined as under severeocclusion, we still predict and estimate their locations with kalman filter. Otherwise, MeanShift algorithm is used to track the occluded targets separately when they are under partialocclusion. Experiments demonstrate that the improved algorithm can achieve moreaccurate tracking result when two targets are under occlusion. Adding the count function tothe algorithm and implementing it based on OpenCV open source platform can achieveaccurate result and faster speed.