The Research Of Kernel Density Estimation And Particle Filter Moving Object Detection And Tracking

Moving target detection and tracking, which has had many years of research history, is one of the foundation tasks as well as it is one of the key technologies in computer vision. It combines the frontier technologies in many areas, such as image processing, pattern recognition, artificial intelligence ect. And it has been applied widely in video surveillance, video conferencing and human computer interaction. This thesis tries to get insights on some key issues of real-time detection and tracking in complex scenes, including the model changes, light effects, background color similar, and target scale variations and so on. The main research contents are as follows:1) This thesis proposed a method of kernel density estimation using clustering difference image for foreground object detection for non-parametric kernel density estimation information redundancy and repetition computation in the training stage estimate error and large amount of computation time in the estimated phase. Firstly, the algorithm obtain the typical sample sets using sample clustering mechanism to; then classify the typical movement pixels and atypical movement pixels based on an adaptive global threshold frame difference and background difference method to, while updating sampling sample set and the reference background image in real time. Experimental results demonstrate that the proposed algorithms eliminate the typical non-movement noise points by estimated error and improve the real-time capability.2) For the lack of space distribution information and target scale variations in the traditional Color feature based particle filter tracking, an efficient tracking algorithm was proposed. This algorithm fuse object feature and spatial information sin adaptive scale particle filter framework. And the scale size of the tracking window adjusts using the update strategy of scale adaptive changes. At the same time shared memory parallel computing OpenMP was used for the acceleration of particle filter tracking. Experiments show that the algorithm can improve accuracy and speed object tracking in the applications of complex object and background. Meanwhile, the adaptive scale update strategy can better adapt the target scale changes in the video scene.3) A prototype system formed by the existing detection and tracking algorithms and the proposed target detection and tracking algorithm in the thesis. First, under different parameters of experiments, the system can be used in the analysis of common target detection and tracking algorithm of different video scene. Second this system also can be modified, and then apply to of the target detection and tracking in actual video scenes.