Research On Parallel Algorithm For Moving Object Contour Extraction In Video Sequences

Video is an indispensable information carrier in current information times, and computer vision has been a hot research at home and abroad in recent years. As an important branch of computer vision, contour extraction is widely used in the field of image segmentation and target recognition. In video processing, the real-time performance is also the goal of researchers, the amount of calculation of extracting moving object contour in video sequence is often huge. So it is difficult to meet the requirements of real-time video processing. With the rapid development of heterogeneous computing, it is an effective way to solve the problem of real time by using heterogeneous computing for parallel acceleration.In this paper, we first study the theories of the parallel algorithm for moving object contour extraction, and analyze the respective characteristics in details,including the common object detection method and the common object contour extraction technology.Then, an improved algorithm of moving object contour extraction based on edge detection is proposed in this paper. Due to the large noise interference, it is difficult to extract a complete and clear contour by a traditional algorithm of moving object contour extraction based on edge detection which use Gauss mixture model and Canny edge operator. So this paper uses morphological operations to eliminate the holes in foreground image which acquired by Gauss hybrid modeling, and remove the Salt and pepper noise in foreground image through median filter, then get the edge image of the original video frame by Canny edge operator. Finally, the result of edge detection and the processed foreground image take "and" operation to get the ultimate contour of the moving object after binaryzation. The experimental results show that the algorithm proposed in this paper can greatly improve the accuracy of the moving object contour,and can extract complete and clear moving object contour.To solve the problem of real-time in video processing, this paper uses OpenCL to realize the parallel algorithm of moving object contour extraction. First of all, this paper introduces the commonly used heterogeneous computing framework OpenCL,and compares the characteristics of OpenCL and CUDA, then analyzes the reasons of the parallel implementation of the moving object contour extraction algorithm based on OpenCL. Then this paper divides the tasks of OpenCL host and device according to the characteristics of the algorithm, and introduces the implementation steps of host and device. Finally, in order to verify and analyze the performances of the parallel algorithm, this paper makes contrast experiments from three aspects: performance comparison of parallel algorithm and serial algorithm, performance comparison of parallel algorithms on different devices, performance comparison of parallel algorithms on different workgroup sizes. The experimental results show that the parallel algorithm is greatly improved operation speed compared with the serial algorithm. In addition, the accelerating ability of the parallel algorithm on different devices and workgroup sizes is different.