Intelligent Video Analysis System Based On DM6437 Platform

Intelligent video analysis techniques mean using computers to process and analyze video source, and extract key information to control the machine to complete the surveillance tasks. Intelligent video analysis technique significantly overcomed the shortcomings in traditional video monitoring technologies, like low efficiency and reliability, need a large number of people to maintain in recent years, and obtained a rapid development. At this stage of intelligent video analysis technique has achieved to detect suspicious vehicle and pedestrians which cross exclusionary area or barriers, detect suspicious articles, and detect converged people and vehicle congestion, etc. In recent years, with the continuous increase of the embedded processors'performance, more and more intelligent video analysis systems began to make their implementation in the front end of embedded systems.Due to the video analysis algorithm in DSP platform needs to consider many factors such as the real-time and space needs. This paper includes the following aspects of research:(1) Introduce the research progress of background detection algorithm, and present the basic principles of the classic background detection algorithms such us basic background subtraction (BBS), W4 method, Single Gaussian model(SGM) and Multiple Gaussian model(MGM) .On this basis, introduce the principle of the project by the actual code book background detection algorithm in details. Then the code book algorithm is simplified and made some comparative experiments.(2) Considering the real-time performance of tracking, the linear forecast method is used to design the moving objects tracking module, which predict the objects'position by their historical trajectory and then judge the movement direction of the objects. In motion analysis module, we solved the problem of interlacing judgment by constructing a correspondence matrix and judging the coincidence values between the separated rectangular frames and enclosing interlacing rectangular frame. Then we use the current position and movement direction of the objects getted in the tracking module to implement the alert zone, alert line and retrogradation detection capabilities.(3) Implementing video analysis algorithm task in single thread will cause frame jumping problems, which will affect the display effect of our system. So this thesis uses lock, semaphores, message queues function modules provided in TI's DSP/BIOS operating system to achieve the video analysis algorithm's multi-threading implementation, which solves the problem of displaying continuous algorithm results. On this basis, we introduce several common optimization methods on the DSP platform, and then give some optimization schemes which are adopted to the intelligent video analysis program presented in this thesis. Finally, we give the optimization's experimental results.