Algorithm Research Of Binocular Vision Moving Target Detection And Tracking Based On GPU

At present, vision system has been widely applied in the field of robot and has become an important indicator to measure the performance of robots. Target detection and tracking, as one of the key technology of robot vision, has always been a research hot spot. Traditional target detection and tracking technology generally relies on monocular vision, provides less information and is unable to access to the location information of objects; binocular vision matching technology can obtain depth information from a pair of matching images of co-polar, but it is not suitable for real time operation due to calculation complexity. In recent years, with the advent of CUDA, people gradually pay more attention to the parallel data processing ability of GPU. Therefore, the project applied binocular vision and GPU in target detection and tracking and concluded a kind of binocular vision moving target detection and tracking algorithm based on GPU.First of all, the target detection method based on binocular stereo matching was researched. Disparity map background difference method is put forward based on SIFT. SIFT is a kind of relatively stable feature operator. Due to the limited number of feature points, regional growth is needed for SIFT so as to obtain dense disparity map sequence. Then, background model is built through the mean shift algorithm; the between-cluster variance method was combined for binarization threshold segmentation. Finally, the morphology processing was conducted on the binarization image and complete target image was segmented.Secondly, the parallelization method of target detection was researched. In order to ensure the real-time detection, the parallel computing ability GPU needed to be used; CUDA programming model and storage model needed further study. In addition, the image preprocessing algorithm and more time-consuming SIFT stereo matching algorithm were accelerated on the CPU+GPU heterogeneous platforms. Besides, detailed division was conducted on the parallel tasks of algorithm and implemented by using CUDA platform programming.Thirdly, the improved Cam Shift target tracking algorithm based on Kalman filter was researched. Cam Shift is a kind of tracking algorithm based on color feature. In order to ensure the Cam Shift tracking precision, the color space model of the image was analyzed. Since Cam Shift algorithm lack of effective motion prediction module and target update mechanism, the research efficiently combined Kalman filter with Cam Shift algorithm and designed Kalman filter update strategy under occlusion to solve the problem that the targets were easily lost.Finally, binocular vision platform was built to conduct experiments on the project algorithm. Experiment analysis was carried out on the algorithm performance under the condition of different scenarios; the feasibility of the algorithm was validated. The experimental results showed that the algorithm could accurately detect and track moving objects and realize three-dimensional positioning in a short distance; its positioning accuracy could satisfy a certain performance index and achieved the desired effect of the experiment.