A New Generation Of Intelligent Video Analysis System Object Extraction Algorithm

With the improvement of living standards, the requirement of improving safety precautions becomes more and more critical nowadays. Intelligent video analysis system has being developed rapidly among numerous security products due to its advantages. Wide application prospect and great business needs make it a hot research topic at home and abroad.As a product of the Internet age, the new generation of intelligent video analysis system is free from geographical restriction. It provides users with useful information in a faster and more intelligent way. Video analysis algorithm plays an extremely important role in such systems, and it determines the intelligent level of the very product. Typically, intelligent video analysis includes many aspects such as moving object extraction, moving object recognition, moving object tracking and behavior recognition. Moving object extraction, which affects the performance of the follow-up steps, is the cornerstone of the whole system. Since the environment varies from time to time, moving object extraction has always been a very challenging task.At present, most of the intelligent video analysis systems apply background subtraction to extracting objects, and Mixture of Gaussians (MOG) has been widely used in applications. This method is adapted to slow changes of environments, and it can reliably tackle various troubles, such as leaves sway, flags flying and so on. However, the spatial-domain information is as important as the time-domain information for moving object extraction. As a result, some image processing methods (Graph-cuts, Seeded region growing, etc) which based on spatial-domain information are introduced to intelligent video analysis procedure. These methods could be adapted to different situations. Experiment results show that using adaptive threshold greatly improved the performance of MOG. It can not only cope with the problem that the moving objects are segmented to small fragments(over-segmentation), but also obtain a good compromise between efficiency and performance.The implementation and optimization of the algorithms is also very important. An algorithm's optimization is much more complex than its implementation process. There are so many optimization techniques that could facilitate this step. Usually, they are classified to four different levels: the physical level, the algorithmic level, the program level and the code level. However, good optimization strategies always rely on the characteristics of the hardware resources and the algorithm. So, it is necessary to get familiar with the architecture of the DSP, the principle of the algorithms, the usage of the tools, etc.