Research On Moving Target Detection Method Of Intelligent Video Analysis Technology

With the development of the concepts such as safe city and smart city, intelligent video surveillance(IVS) is becoming a leading subject of much concern, without manual intervention, intelligent video surveillance processes and analyses video sequences in real time automatically by the steps including moving target detection, target identification, target tracking and behavior analysis, thus it can get the understanding and interpretation of video content. Then it can guide the appropriate warning and action planning. In recent years, more and more intelligent video surveillance systems are applying to various kinds of areas like schools, finance, transportation and so on.This dissertation studies the moving target detection of intelligent video surveillance. Firstly, a variety existing problems of moving target detection are summarized, such as real-time detection and illumination change. Various kinds of methods of moving target detection are reviewed by introducing the classical methods and improved methods. Then, it analyzes the advantages and disadvantages of various types of methods. On this basis, because of its strong temporal continuity of the video frames and the highly structured spatial relationship of the intra-frame, two kinds of improved algorithms are proposed which integrate the method of background subtraction and the method of inter-frame difference:one is based on the adaptive blocking and p-Hausdorff distance, the other is based on adaptive blocking and LBP texture features.Both of the two methods are based on image adaptive blocking, making full use of the relationship between neighbor pixels, avoiding operating on single pixel, reducing the interference with single noise and a large amount of computation. Among them, in order to avoid the influence of noise and improve the detection speed, the first method references the thinking of image matching, by computing the improved Hausdorff distance of corresponding blocks of adjacent frame, it obtains the similarity of corresponding blocks, and the similarity is used to distinguish the foreground from background; then in order to cope with the light mutations, the second method references the texture distribution operator named Local Binary Patterns, the texture histogram similarity is calculated between the corresponding block, and the similarity is used to judge the foreground or background. Experimental results demonstrate that the improved methods have better performance and better efficiency than traditional methods. The first method is based on adaptive image blocking and partial Hausdorff distance, it improves the speed of the target detection, and meets the requirement of real-time processing; because of LBP texture feature’s gray invariance, the second method can achieve preferable results in handling illumination change.