Research On Moving Target Classification,Matching And Tracking In Visual Surveillance

With the extensive use of video surveillance systems, the associated video image processing based on pattern recognition technology has been widely studied and applied. The moving target detection, target classification, occlusion handling and multi-target tracking technology in visual surveillance had become a hot research field of computer vision. It has broad application prospects.In this thesis, target detection, target classification and identification with multiple classes, the object real-time patch matching, target tracking under occlusion, multi-target tracking and other issues were studied. The purpose is to develop algorithms for practical application of specific goals and specific scenarios of intelligent image processing. The main goal is to improve the robustness and real-time performance in intelligent visual surveillance. The research is focused on these questions:How to classify object in the video sequence with multiple classes? How to conduct real-time patch matching online? How to track object under occlusion? How to track multi-target in video sequence. The main contribution of this paper is as follows:1. To solve the problem of multiple object classification in visual surveillance, we propose an algorithm which based on random forest. Targets training and recognition methods usually only deal with a specific target in the image. But in intelligent visual surveillance, There are often difference types of targets in video sequence. Multiple type classification algorithm is required for higher level image understanding. In this paper, efficient feature extraction method and random forest classification are used in the learning stage and recognition stage. It can perform real-time recognition in visual surveillance with pedestrians and vehicles with higher performance.2. To solve the patch matching problem in video sequence, we propose a efficient image patch descriptor with improved affine space quantization. Patch matching is the base of image database retrieval, texture recognition and image registration. We calculate the mean patch in different affine transformation pose, which can remove the influence of perspective transform, scale transform and lighting change. We use discrete cosine transform, which can be used to create a mean image descriptor, instead of the traditional principal component analysis. In order to reduce the dimension of affine space, we introduce an optimal affine parameter quantification method. For different spherical latitude, longitude using different quantization step. It reduces the dimension of described operator and further the computation and storage.3. To solve the problem of occlusion problem in tracking, we propose a feature fusion algorithm with improved particle filter. In visual tracking, interactive movement often exists between different object and the path of different objects often overlap, which may lead to tracking failure. We use color and shape feature to represent the target, and the fusion feature can be used in the observation model under the improved particle filter framework. It has a good performance in occlusion handling.4. To solve the problem of multi-target tracking, We propose a algorithm based on the probability hypothesis density filter. Multi-target visual tracking consists of two parts:the estimation of the state space set and the identity association of the target. The difficulty of multi-target tracking lies in tracking number and position estimation from uncertain observations. The computation cost grows with the number of the state space in an exponential manner, which make the multi-target tracking challenges. In this paper, we treat the state of the target set as a random finite set. The complexity of tracking is reduced under the particle implementation of the probability hypothesis density filter, which can filter the clutter and false alarm. We also use a data association method to get the trajectory of the target.The result shows the algorithm can improve the robustness and accuracy in multi-target tracking.