Research On Moving Object Detection And Tracking Algorithm In Video Surveillance

With the development of the modern technology, the intelligent video surveillance technology based on computer vision represents the main development trend of the video surveillance system, where the key technology problems including moving object detection, object tracking and object behavior understanding and analysis should be solved. The moving object detection and tracking technology is located in the deepest bottom of the video surveillance system which is the basis of the object behavior understanding and analysis. Therefore it is of great significance that we conduct a study on moving object detection and tracking of the video surveillance in this thesis.In this thesis, the moving detection and tracking algorithm for a single object under static scenes is developed. Based on the analysis of existing research results, the improved algorithms, which are combined with the GM(1,1) prediction model into the traditional algorithms, are proposed. The main contributions of this thesis are as follows:As for moving object detection, we adopt the algorithm based on the background difference and edge detection. An improved edge detection algorithm is put forward based on GM(1,1) prediction model. The grey value of a background difference image often presents multimodality. While the edge of background differencing image is detected with this method, the uncertainty due to the threshold segmentation is avoided and the face object can be detected out effectively.As for moving object tracking, the principles of the kernel density estimation method and Mean Shift algorithm are analyzed in this thesis, on the basis of which two algorithms are proposed: one is the object tracking algorithm based on Mean Shift algorithm and GM(1,1) prediction model and the other is the Mean Shift tracking algorithm based on multi-feature space and grey prediction model. Both algorithms introduce the judgment and processing mechanism of the general and similar occlusion. Moreover, they can enhance adaptability to object tracking under partial and total occlusion. The GM(1,1) prediction model is first used to predict the possible position of object,and then Mean Shift algorithm is used to search the real position near the possible position. With this method, the number of iterations can be reduced and occlusion occurred in a short time can be handled well. The first algorithm's feasibility is verified by face tracking experiment in this thesis and experiment results demonstrate that this algorithm obviously is more superior to the traditional Mean Shift algorithm. The second algorithm is used for moving human tracking under the complex background, where the object model is described by the weighted values between the color histogram and the gradient histogram, overcoming the limitations of a single color histogram and making full use of object's spatial information. Meanwhile, the optimized background value of GM(1,1) prediction model is adopted to improve the prediction precision of this algorithm. Experiment results show that this improved algorithm can well adapt to moving object tracking under complex background.