Research On Extended Target Tracking In Complex Background

Target(object) tracking is a basic problem in computer vision. Extended target tracking in complex background, which has been applied widely, is a key technique in optic-electronic observatory tracking system. Tracking system may drift from and even lose targets completely because of change of background, variation of targets, and occlusion. Though many classical tracking algorithms have been proposed, there are still a lot of problems in practical application. To solve these problems, this essay makes an overall research on target tracking in complex background in depth. First, we analyze components of basic tracking system and its shortcomings. Then, referring to this basic system and state-of-the-art algorithms, we put forward some new and effective algorithms to solve problems in projects. Generally speaking, research of this essay mainly includes the following aspects.First, this essay research extraction, evaluation and improvement of image features, which are key problems in target tracking. In computer vision, quality of image features determines performance of algorithm. So this essay summarizes common image features and local invariant features based on HVS. Then we propose three performance parameters to evaluate image features:robustness, dissimilarity and real-time. To track cylindrical aerocrafts in visible light, this essay evaluates a lot of image features using proposed parameters and then improves these features and their description. Finally, local binary pattern combining with Hu invariants is adopted to track part of this kind of targets accurately. To track targets of which geometric attitude changes severely in cluttering background, this essay puts forward a correlation-based tracking algorithm based on adaptive optimal clustering. We research physical meaning and generating method of entropy feature of image under the condition of unsupervised clustering. Then we adopt classification criterion in pattern recognition to propose a new image feature: extended multi-scale entropy(EMSE), which fuses intensity distribution with geometry distribution of image pixels. Reliable and accurate tracking of this kind of targets is realized using this new feature.Secondly, this essay studies fragment-based tracking algorithms, which achieve good performance in practical application. After analyzing advantages and disadvantages of basic fragment-based algorithms, we try to improve and extend these algorithms to increase their performance under condition of occlusion, rapid change of both background and targets. Often there are both background fragments and target fragment in bounding box, so we adopt forward-backward error, which is based on time reversibility and proposed in TLD algorithm, to reject background fragments and occluded target fragments and select unoccluded target fragments. Then we only use these unoccluded fragments to track the target via correlation-based method accurately. At the same time, size of the target may change during tracking process. We use purified SIFT features to calculate average distance ratio between frames, and use the ratio to adjust the size of bounding box and fragments. By doing so, we can track the target of which size changes. To update template accurately and timely, our algorithm use SAD correlation-based matching to pick out stable target fragments, then distinguish unstable target fragments from background fragments via anti-projection of major gray components and histogram. All the target fragments are updated to capture change of the target, while background fragments are not updated. We compare our algorithm with state-of-the-art algorithms through enough and strict experiments. It is obvious that our algorithm is superior to existing algorithms under condition of rapid change of background and occlusion.Finally, this essay studies application of machine learning theory in extended target tracking. Machine learning and computer vision both belong to artificial intelligence so there are natural link between them. One of core issues in machine learning is extraction and selection of features via supervised or unsupervised learning. While this process is used widely in other fields such as object detection and recognition, it is too slow to be used for real-time target tracking. This essay make research on sparse representation theory(SR) of images, which is based on human visual system. Then we exploit application of dictionary learning method of SR in extended target tracking. In the view of target tracking, constructing of dictionary can be understood as extraction of multiple templates of the target, seeking sparse coefficients can be understood as matching of multiple templates, and dictionary update can be understood as template update. In our project, we use principle components analysis method to construct the dictionary. And we adapt Lasso equation to increase performance of our algorithm under condition of occlusion and distraction of background. We adopt incremental learning method to update the dictionary. Combining this new dictionary learning method with particle filter framework, the essay proposes a new observation equation to solve occlusion. We select samples for incremental learning according to degree of occlusion and then update the dictionary. Particle filter method requires too many samples and sampling process is not completely precise. To solve this problem, we calculate affine parameters between frames and use it to improve sampling method. This improved method can sample more precisely while require fewer samples. At last, we compare our algorithm with state-of-the-art algorithms through strict and enough experiments.