Research On Features Extraction Method Of Moving Target In Video

The paper mainly researched into the method of moving target feature extraction in complex scenes video. In military use, the background of video scene is often very complex, for example, the background of the wild jungle, desert, urban neighborhoods, sea and sky background in bad weather conditions and so on. In these complex environments, the traditional methods of detecting and tracking moving targets are difficult to obtain good results, mainly because the moving process of objects in video is a non-linear non-Gaussian dynamic one. In this case, only the efficiently extracting and making use of target features, and may implemented detecting and tracking targets.This paper proposed a feature extraction method of the moving target in a video with complex background. The main idea of this method is that the the moving target is regarded as a set of sub-regions which denote portions of the target separately, and the moving features of the moving target are expressed in terms of geometric and demographic features of these sub-regions, such as geometric location, size, direction and probability distribution are obtained via statistical method. Thus, the demographic features of a moving target are composed of jointing the moving relationship between sub-regions. The feature extraction method is divided into five steps: firstly, using Gaussian mixture model (GMM) method to preselect areas of the moving target; secondly, using the information bottleneck (IB) method to cluster the preselect regions of moving; thirdly, using the method of unsented kernel-based Bayesian filtering(UKBF) to track the moving target; fourthly, clustering the sub-regions of moving target; fifthly, extracting the statistical features of sub-regions of moving target, as their own features, such as the width-height ratio of the circum-rectangle belongs to the target sub-region, the variance of the velocities of each sub-region, the number of the target sub-regions and the probability density of their state variable. Finally, experiments demonstrate the effectiveness of this method, distinguish the moving target in complex video and judge target belong to human or car.