Research On The Implementation Of Real-time Moving Detection Base On GPU

Moving target detection and tracking is one of the most challenging issues in the electric and communication area. Application of the visual detection and tracking are numerous and they span a wide range of applications including surveillance system, vehicle tracking and aerospace application, to name a few. Due to the increase of the amount of information and improvement of the algorithm accuracy, the implementation of moving target detection algorithm on CPU has occupied more and more time.Thus it gradually unable to meet the requirements of real-time image processing algorithm efficiency. When the 32 bit floating-point vertex and pixel shader processor GPU arose, the image processing on GPU has possibility. At the same time, with the GPU support of the increasing number and complexity of flow control, the implementation of complex real-time image processing algorithm on GPU has been possible.Aiming at the problem a method to improve the speed of the implementation of image processing algorithm was proposed, which used the GPGPU to implement the efficiency real-time image processing. In order to verify the correctness of the method proposed by the issue. The pyramid optical flow algorithm was used which has large computation implements the moving target detection. Then the results were analyzed to verify the correctness of the target of the issue. The issue mainly contains the researching of basis theory, the study of technology, the implementation of the system and the analysis of the results. The moving target detection was implemented from the theory to practice, then the target of the issue was verified in the results analyzing.First a research on the base theory of the system was completed.The base theory of the system has two targets includes GPU general computation and programming and the research in target detection algorithm. In the aspect of GPU general computation and programming, to understand the working principle of GPU shader programming, a research on the programmable vertex shader and programmable pixel shader processor on GPU was completed. To understand the advantage of the GPU stream parallel computation model in numerical computation, a study on the GPU general numerical computation was occurred. At the same time, a comparison between GPU programming and CPU programming was researched to determine the programming idea of GPU algorithm. In the aspect of target detection algorithm, the general target detection algorithm was studied to analyze the advantage and defect of different algorithms at the beginning. The result shows that the implementation of the optical flow algorithm depends on the hardware computation speed. On the basis researching on the basic optical flow algorithm, the pyramid optical flow algorithm was studied to solve the problem that the optical flow algorithm can’t solve the defect of target wide range movement. Finally base on the study of algorithm iteration, the pseudo code of the pyramid optical flow algorithm was summed up.Then the technologies of the implementation of the system required were studied, which mainly includes video capture, video rendering and the implementation of video processing algorithm. Through the study of Direct3D basis and its rendering pipeline, understands the whole working principle of GPU system and knows how transmit data from CPU to GPU and how the data processed in GPU transmit to CPU. Also understands the relationship and the means of the matrix uses to conversion between the location coordinate, view coordinate and projection coordinate. Based on the study of DirectShow video capture system framework, knows how to use the data flow control to capture and redirect the video data stream and how to implement video capture and video format configure or acquisition of different video hardware driver. Then create the basic video capture system framework. Through the study of Direct3D effect framework and HLSL programming, knows the working principle of GPU programmable pipeline and the basic grammar and semantic binding in GPU programming. Then use theory to complete the GPU programming framework.In the end, based on the study of basis theory and technology, a basic video capturing and processing system was created to implement the pyramid optical flow algorithm and display the results. The system can be divided into two modules which were image capturing module and image processing module. The image capturing module use the Microsoft DirectShow video capture framework to implement the initialization of the vide capturing system, the selection of different capturing devices, the initialization and setting of video preview window and the redirection of the video data stream. It is a remarkable fact that the HOOK technology was used to redirect the data stream which redirecting the data by changing the pointer of the video stream receiving function. Image processing module uses the Microsoft Direct3D interface and effect framework to implement the initialization of image processing module, the connection between GPGPU algorithm and application and the rendering of processed image. The initialization of the Direct3D interface includes registering the Direct3D interface, detecting the GPU supportable, registering and creating the IDirectDevice interface, the initialization of view coordinate and projection coordinate transforming matrix, creating texture and the initialization of texture UV coordinate, compiling shader program and obtaining the variable handler. Base on the pseudo code on the study of basis theory, the implementation of pyramid optical flow algorithm was implemented. After all of this, a simple window was created to integrating the image capturing module and image processing module to complete the verification platform for verifying the feasibility and efficiency of the implementation of pyramid optical flow algorithm on GPU.Based on the basic theory, the system software was implemented. The software was tested whose results shown that it runs well and can preview the image. The software was tested in four situations, the high speed movement in the direction of up and down, low speed movement in the direction of up and down, high speed movement in the direction of fore-and-aft, low speed movement in the direction of fore-and-aft situations. The results was recorded. After analyzed the results of the image processing system, the feasibility of the implementation of GPGPU pyramid optical flow algorithm was verified. Based on the literature and the results of the experimental, the feasibility of GPU using to real-time image process was fully proved. The comparison of the computation time of implementation of pyramid optical flow algorithm between CPU and GPU validated the efficient of implementation of image processing algorithm on GPU. The research of this issue, it not only has a certain reference value in the implementation of efficient image processing algorithm on GPGPU, has a certain academic significance to research on GPU general computation at home and abroad.