Research And Implementation Of Moving Object Contour Extraction Algorithm Based On GPU

Intelligent video surveillance is the use of digital image processing and computer vision to solve traditionally required human participation in monitoring mode, instead,computer itself processes the captured images from the front and to achieve the purpose of detecting, tracking and identification moving target. On the basis of prior result, the analysis and judgment of the moving target can be completed and gives the description of the moving target’s behavior and actions so that implements the unattended case of abnormal carried identification and automatic alarm, and then constraints and guide the action of people.The research and application of intelligent video surveillance is a hot and difficult problem at home and abroad, many scholars concentrated study this filed and get a lot of promising results. Based on the results of previous studies of intelligent video surveillance system, this paper makes a key step research and analysis on moving target detection and contour extraction.Now, at home and broad the major technologies in dealing with the problem of moving target detection include background subtraction method, the continuous inter-frame difference method, which consists of two frames, three frames and even the multi-frame difference, and the optical flow method. In this paper, a static background video file uses frame difference method for the detection of moving targets, and then binarized the extracted image. After comparing the pros and cons of several mathematical morphology algorithms, finally, morphological dilation and erosion operations are chosen for contour extraction, the experimental results are consistent with prior theoretical analysis of the moving target.However, due to the mathematical morphology in dealing with the large amount of data is very difficult to achieve the needs of real-time in video surveillance.Because the dilation and erosion are parallel processing operations, in order to improve the speed of the mathematical morphology operations, the related concepts of image processor GPU (Graphic processing unit) is proposed. Graphics processor hardware structure and how CUDA (Computer UnifiedDevice Architecture) platform parallelly processes the morphology algorithm on graphics processor are analyzed. Compare the capabilities of serial processing on CPU and parallel processing on GPU. Finally, the detailed implementation process of dilation algorithm on GPU parallel computing is given and described some specific problem in the programming process in order to take full advantage of GPU resources. The experimental results show that the speed of parallel processing of mathematical morphology algorithm on GPU can be enhanced to reach a multiple of speed than serial processing on CPU.