Research On Target Tracking Based On Video Image Sequence

In recent years,with the development of computer vision technology and expanding demands in practice, moving target tracking has broadly applied in safety surveillance, robot navigation,intelligent transportation, unmanned and other fields, and then it becomes the focus of domestic and foreign researchers. However,technical issues such as the complex of the scenarios, the difficulty of feature selection, the occlusion problem and so on, still need to be resolved, so this subject research has important theoretical significance and pratical value.This thesis mainly aims to study algorithm of tracking moving target in video image sequences. First the tracking method based on mean shift is studied, considering the existing problems, improved algorithm that increases tracking stability is presented, and then Scale-Invariant Feature Transform (SIFT) feature points matching is studied, considering the existing problems when the SIFT feature points matching is applied in moving target tracking, improved algorithms are presented. The major works of this thesis are summarized as follows.The tracking method based on meanshift is studied, which has good adaptability to target round and part-occlusion, its calculation is simple, so this method can be used in real time.However, it could not adjust the size of the tracking window adaptively.A method based on meanshift and the Kalman filter is used when there are serious disturbances which may lead to failure to target track.According to different disturbances circumstances, different weight factors are adopted to combine Kalman filter predition result with meanshift tracking result. This method makes good use of space position of the target, so increases the reliability of the tracking. When the scale of the target changes, tracking based on meanshift could not adjust the size of the tracking window, and this may leads to lose of the target.Continuously adaptive mean shift is then studied (Camshift), this algorithm can adjust scale with object during tracking process. However, when there are disturbance from objects of similar color, it may consider the disturbance as the target mistakenly. An improved Camshift is proposed, which uses the ratio of long axis and short axis of the target to determine if serious disturbance happens. When the disturbance happens, the algorithm switchs ( tracking based on meanshift.In the study of SIFT feature points matching, to eliminate the error matching points that are generated during the matching process, this thesis adoptes the max-min distance means to classify the slopes that are generated between each pair of matching points. When SIFT feature points matching is applied to track target, the target may have major changes, tracking use the same initial template would fail. A new method that uses meanshift to extract the matching area in order to update template in real-time, and when there is occlusion from other target of similar color, the method based on meanshift and Kalman filter is used to update the template. When target scale changes, Camshift is used to update the template. The computation of the SIFT feature points matching is large, so it can not be used in real time. This thesis uses Kalman filter to predict the positon that the target may be presented, and then uses a appropriate image rectangle around the predicted position that contains the target to be tracked to match with the template, the experiments shows that the SIFT'featrue points matching is improved in real-time.