Research On Multiple Target Detection And Tracking Based On The Dynamic Background

Multi targets detecting and tracking is a practical and tough topic in the field videotracking, especially the one under dynamic background. Currently, only a small numberof tracking algorithms deal with dynamic background, and each of them has its ownadvantages and disadvantages. In practice, such tracking under dynamic background hasa very wide application, which not only refers to image processing, video processingand pattern recognition, but also has a great importance of research. For example, multitargets detecting and tracking can be used in civilian, commercial, and public facilitiesmonitoring, intelligent traffic monitoring, military weapon guidance, vision navigationand so on.In this paper, we proposed an efficient multi target detecting and tracking algorithmunder dynamic background. First, for the detecting part, we compute the optimalparameters and apply them in affine transformation of the current video frame, whichcompensates the motion of background, so that the question is changed into multitargets tracking under static background. Then, we use frame difference to detect themoving areas in the current frame, and thus obtain the targets set with numbers andrectangles, which work as the input of the next tracking algorithm. The experimentalresults verified the correctness and rationality of the proposed algorithm.For the tracking part, we first introduce a single target tracking algorithm based oncolor feature and Kalman filter, Here we convert color feature space into HSV colorfeature space, then extract the color feature of the moving target, and then combinewith the Mean Shift algorithm, using the similarity function determine whether thetarget occurred shelter, If the target blocked we use Kalman filter to target positionprediction, if the target does not occur to block we use the Mean Shift algorithm to trackthe target, and verify the feasibility of it during experiments.At last we extend this algorithm to the multi targets situation. For the targetsobtained from the detecting part, we extract the global feature and forecast the roughposition of each target by using feature matching method. Then Mean Shift iterationsare applied at such position, while Kalman filter will be applied in case of occlusion,therefore the precise position of each target is obtained. Experimental results shows thatthis algorithm has a good performance on real-time tracking, and can handle generaltracking problems under dynamic background.