Real-time Moving Target Detection And Recognition In Video Tracking System

Moving target detection, recognition and tracking is an important research task in applied visual field. It has been widely used in military filed such as weapon control and guiding system. This thesis studied the key technic of the video tracking system —automatic moving target detection and recognition and their real-time dealing methods from two angles which are analyzing algorithm and making experiments. On the premise of meeting the requirement of real-time processing, this thesis solved the problems of the precision of object detection and segmentation, as well as the accuracy of object recognition. With the algorithm proposed by this thesis, we can track the given target continuously and steadily. In short, this thesis consists of image precondition, obtainment of the target's moving information, image segmentation and object recognition. First of all, median filter is employed in image precondition to improve the quality of image.Then, neighborhood comparison based on the difference between image sequences is used to obtain the moving information of the target. This method is simple and easy to be realized. To some extent, it can eliminate most background noise, and enhance the ability of anti-disturbance of the tracking system.To improve the quality of segmenting, this thesis provided a two-step target segmentation strategy, the cursory one and the following accurate one. In the second step, it proposed a new segmentation algorithm based on the technique of one-dimensional maximal between-class variance and region growth. The technique of region growth is improved in the algorithm by proposing a growing rule in a restrictive condition and a new growth mode. So we can reduce the complexity of the algorithm , cut down the calculation and overcome those fail segmentation caused by figure blur or uneven target grey. Therefore, the image of moving target can be segmented quickly and accurately in course of tracking.Moreover, an object recognition algorithm by the primary way of feature matching and the secondary way of template matching was put up in this thesis according to different working state. We also constructed a cubic non-linear recognition reliability function with which target of each working state can be recognized accordingly. In this way, the tracking system will work with high recognition rate and low false alarm probability. Furthermore, in order to enhance the target recognition precision and the tracking stability, many effective strategy are used in the system, such as similar objects distinguishing, feature parameter model and gray model refreshing automatically, adaptive tracking window.Results of the experiments indicate that these target detecting, segmenting and recognizing algorithms are steady and practical to segment and recognize the missile correctly within a certain range of error.