Research Of Moving Object Tracking In Video

Moving object tracking is the process of continuously predicting state of interested object which is given in the first frame in the subsequent image sequence. Object tracking is a hot research topic in the computer vision, and it is widely used in surveillance, automatic navigation, human-computer interaction, augmented reality, medical imaging and so many other aspects. Howere, there are so many adverse factors in the real world like illumination change, occlution, rotation, posture change, background clutter, so object tracking is still a challenging and meaningful research topic. This paper focuses on studying on constructing effective object representation, dealing with fast motion, judging occlution, adjusting to scale variations, eliminating distracters and model updating strategy. The main research works are listed as follows:(1) Visual tracker via spatio-temporal context often drifts as rapid movement or occlusion happened to the target, however the confidence of real target is relatively large and target locates in side peaks or around the main peak, so the improved visual tracking via spatio-temporal context based on confidence map property is proposed. Three types of patches combinations are used to construct target representation model. Multiple peak points of the confidence map are regarded as candidate locations, then representation features of the locations are extracted to match target template to find out which is most similar with target and sequential monte-carlo method is used to obtain the best target state. Finally, the occlusion state of target patches is computed so that the learning rate of spatio-temporal context is adjusted through occlusion ratio. Simulation experiment results show that the proposed algorithm could still track the target accurately when rapid movement or occlusion happens to the target.(2) In order to address the problems of occlusion, rapid movement and scale variations, we propose multi-scale object tracking with local presentation and global presentation. At first, we use best-buddies similarity detection which is based on local presentation to obtain preliminary location. Then the exemplar-based linear discriminative analysis classifier which is based on global presentation is used to get the more accurate location. And the exemplar-based linear discriminative analysis classifier usees an offline background model as the initial background model, the background model updates online as the negatives samples collecting from object surround, only the target sample as the exemplar to train linear discriminative analysis classifier at one frame, the weight of sub classifier is depend on the similarity of the exemplar with current object state. At the same time, the CUR filter is used to do detection and we judge occlusion according to the overlapping ratio between results of detector and tracker as well as the response of target from correlation filter. Finally, the correlation filter fuse results of tracker and CUR filter to obtain location and uses multi-scale samples which are got from the location to predicate the scale of object. Simulation experiment results show that the proposed algorithm is adaptive tocale variations, and could still track the target accurately when srapid movement happens to the object or the object reappears under occlusion.(3) As color tracking is susceptible to be interferenced by the regions with similar color characteristic, we propose object tracking based on distracters elimination and multiple snapshots recover. As convolution features is complementary to the color features, when drift happens to color tracking, fusing color similarity and convolution similarity could exclude distracters effectively. The convolution feature similarity will reduce significantly when occlution happens to the target, so this can be used as a clue to determine the occlution state and adjust the update of object template of convolution feature to keep the model's accuracy. Continuously forward update may produce error to color model, the historical state is recovered according to consistency of the results of multiple color model snapshots and the final object state, so the robustness is improved. Simulation experiment results show that the proposed algorithm could exclude distracters effectively when illumination variation, occlution and deformation happen to the object, so the accurate target sate could be is obtained.