Study Of Target Tracking Techniques Based On Vision

Visual target tracking, which aims to estimate target trajectories in video scenes, is an important research direction in computer vision. The research involves related knowledge of many fields such as computer image processing, pattern recognition, artificial intelligence and automatic control, and so on. This technique has a promising application prospect and research value.Although many existing trackers have been proposed after many years of research, target tracking is still challenging because of occlusion,change of a target appearance,frequent interactions between targets and illumination variation.This dissertation thoroughly explored the key technology of target tracking by investigating the existing work about single target tracking and multiple target tracking:multiple cues adaptive integration, template updating, online learning target model and tracklets optimal association, and proposed innovative solutions to deal with interference. The main work of this thesis includes:●A GMM background extraction method based on gradient statistical information is proposed. Aimed at the defect of classical GMM background modeling which is sensitive to quick illumination changes and cannot detect a slowly moving and stationary object, the gradient statistical information of local image regions is integrated into the GMM model, which enhances robustness to illumination changes. Then a method of estimating parameters of GMM based on split EM is presented, which improves the performance of the model. An adaptive learning rate concept is used to handle the problem of detection for slowly moving and stationary object. The experiments show that the proposed method can effectively improve the detection rate and robustness of the model.●Researching the key technology of single target tracking, we propose a method of adaptive multi-cue integration tracking based on fragments. It divides the template and candidate targets into multiple fragments with each fragment being described by color and CS-LBP cues and cues are integrated adaptively. The greater the weight of a cue the stronger the ability to prescription. Thus the contribution of each cue to the joint tracking result is different under different conditions. Our algorithm also can update the fragment template online during the tracking process. In the small fragments the method detects the possible fragments where appearance changes or occlusions occur, and then takes corresponding update strategies according to the matching of the small fragments. The whole algorithm provides a good solution for the problem of tracking object in the complex background of illumination, occlusion, and the appearance change.An approach based on online learning target model and hierarchical association of tracklets for multiple target tracking is proposed. It first detects targets in each frame and each target is described by position, size and color features. The first batch of target reliable tracklets is generated by conservative dual-threshold strategy that only links detection responses in consecutive frames, then the discriminative features of color and LDB and the non-linear motion map are learned online simultaneously. Finally, a global association tracking procedure is carried out within a temporal sliding window in4lays. Good results on occluded objects tracking is obtained through this method.