Realization Of Moving Object Detection And Tracking Under Moving Ground Background Based On DSP

Object detection and tracking under moving ground background is the main task of the airborne electro-optical targeting system. By using DSP, the embedded image processing core unit, which has strong computing performance and low power consumption, object detection and tracking algorithm under moving ground background is realized efficiently based on the study of C64x+ DSP structures, program optimization step and strategies.Background's motion estimation is the basis of the moving object detection under moving background. Background's motion estimation based on the KLT optical flow algorithm is studied. Feature point extraction and tracking is the algorithm's most complex part, which is studied and optimized to realize the background's motion estimation on DSP. To reduce background's motion estimation's time, based on the systematic study of the methods to accelerate block match, a background's motion estimation algorithm based on the fast block matching is put forward. It uses sobel operator to extract the feature blocks which have the most texture information to reduce the amount of the blocks, and hexagon search strategy to reduce the search points to accelerate block match. It is realized based on DSP.Parameters obtained from the background's motion estimation are used to compensate image background motion. By interframe difference binarization, morphological filtering and connected domain labeling with filter, object is detected, noise and false objects jamming is eliminated. For detection results by using temporal information solely maybe inaccurate, the salient detection method is deeply studied. A moving object detection algorithm based on fusion of regional pixels salient information and temporal information is put forward, and it is realized based on DSP.To solve complex problems of appearance change and occlusion in ground object tracking, the compressive tracking algorithm is improved. Each frame's classification score and the MCD distance between object and model are studied. Classification score is used to judge occlusion coarsely and MCD distance is used to judge occlusion finely. Kalman Filter is used to predict the position in next frame, reduce searching range and improve the accuracy. When occlusion happens, the object's position is predicted, and motion trajectory is estimated. Real-time object tracking is realized by using improved CT algorithm based on DSP.