Dynamic Target Detecting And Tracking Based On DSP

Targets detecting and tracking are widely used in intelligent transportation,battlefield targets strike,security monitoring,human-computer interaction and other fields.Effective detecting and tracking of the target is one of the hot topics in the field of computer vision.Because the detecting and tracking system is used in different scenarios.Usually it is necessary to transplant the algorithm to the lighter,smaller,more stability hardware platform such as DSP,FPGA,and so on.The algorithm which is stability,accuracy and real-time is demanded.With the continuous development of computer hardware,many efficient but highly complex algorithms can be applied in practice.For example,due to the rapid improvement of GPU in recent years,deep learning has been developed rapidly.Many achievements have been achieved in the field of target recognition and artificial intelligence.In this paper,TI's DSP TMS320DM6437 platform is used to realize the detecting and tracking of small UAV.Firstly,the performance of several different target detecting algorithms and the data processing capability of DM6437 platform are taken into consideration.Then the three-frame difference method and the mixed Gaussian background modeling are selected to detect the UAV.It is found that the real-time of three-frame difference method is much better than mixed Gaussian background modeling,what's more the anti-noise ability of three-frame difference method is very strong,so we choose three-frame difference method for our target detection algorithm in DM6437 platform.Then the detected target image is morphological processed and then the sliding windows are used to initialize the position of the target.Then,we are going to track the target.Then we compared and analyzed the commonly used target tracking algorithm,centroid tracking,template matching and mean-shift tracking algorithm.Because the large computation of the Mean-shift algorithm,so first of all we realized the Mean-shift in the Intel Core(TM)i5-4430 CPU 3.00 GHZ processor for UAV tracking.The results of the experiments show that for 50x50 target region,the processing speed is about 31.5F/s,so we know that mean-shift algorithm is very difficult to achieve real-time on DM6437 with the frequency of 600 M.So we intend to use centroid tracking and template matching method,for template matching,we realize the mean normalized product correlation and the improvement of the mean normalized product correlation algorithm in the reference [51] on DM6437,and we find that both of them can achieve better tracking effect with their own characteristics.Finally,because the target detecting and tracking algorithm described above is difficult to get the ideal tracking effect in the complicated environment,this paper studies the application of deep learning in target detecting and tracking.We use NVIDIA TITAN X(Pascal)GPU to accelerate the processing on a computer with Linux system,we realize YOLO v2 to detect and track the UAV.We realize the real-time and accurate detecting and tracking of UAV in complicated environment,the speed reaches 40F/s.