DSP-based Research On Moving Object Real-time Detection And Tracking

The research of target detection and tracking has always been a hot field of computer vision, but also a difficulty. There are static and dynamic background of target detection and tracking, the research under static background have already been more mature, while there are many problems to be solved for the research on moving target detection and tracking under the dynamic background. This paper studies how to detect moving target in a dynamic background, and then tarck the target on this basis. The processor used in this paper is C6000 series TMS320DM642 produced by TI (Texas Instruments) company. The camera mounted on a PTZ that fixed on an unmanned aerial vehicles(UAV), and the target could be any moving object on the ground. The main work in this paper is as follows:1. The purpose of compensation for dynamic background is to get a camera or other image sensor's motion model parameters, and then to obtain the position transformation relationship of corresponding pixels between adjacent frames, and eventually compensate the relative movement between the camera and the target, making the background "static". Firstly,Harris corner detection algorithm based on adaptive threshold is used to extracte the feature points in a image, then, match these points with the similar function, and improves the matching rate with RASUAC algorithm, finally we obtain the affine model, used to realize the background compensation.2. The background subtraction is used in this paper to detecte the moving target. Since the background always changing, the methods for static background modeling is no longer applicable, in this paper, a new method for modeling dynamic background is proposed, which fully considered the dynamic background compensation. The key to this algorithm is the updating part of the background model, which both considering the current frame and the image of previous frame after the compensation. Ultimately we can detect the moving target after the result of background subtraction processed with morphological filtering.3. There are many methods and features for moving target tracking. Taking the application of this research and the superiority of particle filter algorithm's nonlinear and non-Gaussian property into account, we choose particle filter algorithm and texture histogram for target tracking, and the tracking feature is the color histogram. In this paper, we focuses on the specific application of the particle filter in visual tracking, including the establishment of sports models,weight update and objective positioning.4. The implementation platform for moving target detection and tracking is DSP. This section focuses on how the detection and tracking algorithm described above programmed and run in DM642 with C language and DM642's own API. Firstly, configure various objects and their properties in the real-time embedded systems DSP / BIOS according to the actual needs. Then, achieve the synchronization and communication of task threads under the reference frame RF5. There are four main task threads in this paper, which is video capture task, video processing task, video display task and control task. Also there are video capture and display program and their hardware device driver settings. Finally, the experiment results show that DSP-based moving target detection and tracking algorithm can track the moving object well under dynamic background, and the tracking result based on fusion feature have better performance than just based on a single color feature, and an auxiliary particle filter's tracking effective is better than the effect of particle filter.