Moving Targets Detection And Segmentation Based On GPU

Moving object detection and segmentation have always been an important researching topic in the computer vision field. In the field of computer vision, track moving targets, moving target classification and understanding of the behavior are based on the research of it. Moving target detection and segmentation have wide application prospects in intelligent traffic monitoring system, national security, and medical research and so on. In this paper, a lot of algorithms theoretical studies and experimental verification for the current problems, such as the controllability of background absorption time and background segmentation accuracy, are been done based on in-depth study about background subtraction algorithm.First, this thesis has analyzed background subtraction algorithm systematically and then discuss the interfering factors including the controllability of background absorption time and the time-consuming problem of constructing background model. A background subtraction algorithm based on codebook is presented to extract moving objects from surveillance videos. The background absorption time of this algorithm is stable and has nothing to do with the complexity of the scene. Its length can be changed by setting the algorithm parameters to meet the needs of different occasions. Compared with Gaussian mixture model and codebook model algorithm the background subtraction algorithm based on codebook needs less time in matching and updating the background model.Second, this thesis has speed up the background subtraction algorithm on the basis of GPU of CUDA. With the rapid development of hardware technology, GPU has become a general computing device with high-bandwidth, multi-core and highly parallel computing. Through rigorous experimental verification, GPU-accelerated parallel algorithm is feasible. And the GPU parallel algorithms can be optimized with storage model of the GPU and asynchronous execution techniques.To deal with the shadow of background subtraction algorithm based on codebook, this thesis proposes a multi-feature fusion background subtraction algorithm. At the pixel level, the algorithm sets up color models with background subtraction based on codebook, while at the region level, the LBP texture models are used. The texture models can ignore the effect caused by shadows. But, as the complexity of the environment and background noise and other environmental factors, the results of texture background model tend to have many loopholes and their edges are always not continuous. So it is necessary to fill the loopholes caused by texture background model before using it. Also this thesis has sped up the algorithm with the help of GPU technology. Finally, this thesis has designed a warehouse alarm monitoring software with moving target detection algorithm for real-world applications. This software is developed by using Hikvision video capture card, infrared camera, web camera and software development kit Hikvision. The software is divided into two parts, the server and the client. The server mainly provides five services, such as video capture, storage, transmission, analysis, and intrusion alarms. The client is responsible for displaying video and setting the monitoring area. The key of the software algorithm is the fast background subtraction algorithm which is proposed in this thesis. Software client and server communicate through the network. After testing, the software can basically meet the customer’s needs, and has a better extensibility and practical value.