Research On Moving Target Detection And Tracking In Video Sequence

Computer vision is to make computer perceive and understand environment like a man. Dynamic object detection and tracking are the important issues of computer vision and the hotspot of the certain research domain. In recent years, the target detection and object tracking have attracted many researchers’attention. Although many effective visual object detection and tracking methods have been proposed, there are still a lot of difficulties in these tracking algorithms due to the challenging complex scenarios. Therefore, the researching on moving target detection and object tracking under complicated background has both important theory significance and application value. The main research contents and contributions of this paper can be summarized as follows:(1) For the moving target detection, firstly, three common algorithms in motion detection are introduced and compared, including the optical flow calculation method, frame difference method, background difference method. Then two main background modeling methods are introduced, single Gaussian Model method Gaussian Mixture Model method. On the basis of the comparison of these algorithms, the moving target could be detected by the combination of Gaussian Mixture Model method and the background difference method. During detection of the moving target, first of all, transform the original video images to be grayscale images and use median filter to restrain noise. Then background model is established with Gaussian Mixture Model method to update the background image, which should then be applied to the background difference method. Finally morphological filter on the target area detected by this method is carried on. It is shown in experiments that the algorithm can extract the moving target area completely and quickly.(2)When using the traditional Mean-shift algorithm for tracking a detected moving target, deviation and even failure would be caused if tracking a target with growing si ze in the video sequence. In order to solve this problem, a Mean-shift tracking algorith m with an adaptive bandwidth has been proposed. In this algorithm, first of all, registration of the target centroid is realized by backward tracking, which can compensate the positioning deviation. After that, the feature points which in the tracking window of adjacent frames are normalized to the coordinate system that treat the target centroid as origin. Finally, use the amplitude to update the window width. Experiments show that this algorithm could track the size bigger target exactly.(3) Deviation or even failure would be caused when the tracking moving target encounter interference from the object of similar color and occlusion. To solve this problem, a tracking algorithm based on filtering is introduced. In this algorithm, random drift model is used to characterize the state transition of the particles and the weights of the particles are measured by the color distribution of the target. In addition, re-sampling is proceeded on the particles according to the weights, which would have a better effect on the tracking for a single moving target in complex background. The key to track multiple moving targets is to eliminate the jamming of the false target and distinguish each target effectively, which also called data correlation. In this article, the observation data are associated with the sampling particles of the moving targets by the joint probability data when tracking for multi-target. The Gating technology is used to exclude the impossible data correlation to avoid data correlation in the whole observation space and reduces unnecessary computation, which could improve the efficiency of the correlation.