Study On Object Tracking Algorithm Based On Gpu Parallel Computation

In recent years, intelligent video surveillance system has been used more and more in a growing number of people’s daily life for its powerful function and practicality; it has become the leading issue of popular concern in the field of computer vision. Moving target tracking technology is not only the key technologies of video surveillance system, but also one of the research hotspots in this field. Tracking of moving object in a monitored environment means to determine the specific regional objectives, position tracking and it acts draw trajectory. In this thesis, analysis and research were conducted to the static background in video sequences of a specific target tracking algorithm.the focus of this article is how massive and fast to the tracking of moving object in the video image data.Nowadays, GPU General computing technology has been used more and mor e in the field of high performance computing, making it the absolute main power for calculates data. At the same time, Nvidia Corporation researched and opened a new unified computing architecture of CUDA, which makes GPU has a better programmability, and has a good effect of acceleration to the parallel processing a-lgorithm.As there is a high demand of real-time for target tracking in the complex scene, the complexity and calculation of the algorithm become larger. Because of the bottleneck’s appear in the massive data processing by only use CPU, it is difficult to meet the requirements of real-time target tracking. This paper focuses on the existing target tracking algorithm based on CPU, and analysis the algorithm and algorithm of processing video data in the process and time consuming part seriously. A detailed parallel operable algorithm was designed to the image processing module、the video image detection module, as well as the tracking module. A study was conducted to several algorithms of the image preprocessing module, and GPU parallel accelerated processing was conducted for algorithms individually, as well as tested these algorithms on the CUDA compute unified device.Currently, SIFT feature matching extraction method is one of the most accurate algorithm which was used to implement the image matching features and pattern recognition, and also one of the important part of the target tracking algorithm. Analysis and research of the SIFT algorithm has been focused on this article; proposal and recommendations are proposed for the using of CUDA architecture for optimization and improvement.Test results and the corresponding analysis were conducted through the comparison and verification of images with the same pixel size which were crawled from the streaming videos. The experimental data indicate that some part of the algorithm which optimized by using CUDA can improve the implementation process efficiency, which has certain practical significance.