Human Visual Intelligence Motivated Robust Object Tracking

Visual object tracking plays an important role in the field of computer vision.It has a wide range of applications in military and civilian fields,and has important research value.In the past decade,object tracking has been widely studied with the rapid development of computer technology,and lots of object tracking methods have been proposed.However,compared with human vision,there is still a lack of robustness and intelligence about existing methods.Currently,how to deal with the the influence of uncertain aspects that appear in complex environment,to achieve more robust and reliable tracking,is the most difficult problem about object tracking.In this thesis,we try to improve tracking robustness by incorporating the intelligent characteristics of human vision that undergos complex envoriment changings.Based on traditional tracking methods,we propose three robust object tracking algorithms motivated by human visual intelligence,including,discriminative feature selection based tracking,salient subregion based tracking,and context learning based tracking.The main works are as follows:(1)We review the problem of robust tracking,and conclude key intelligent characteristics about human visual perception,which are selecting salient features,focusing on attentional regions,learning context information for better perception,and learning and associating features from prior knowledge.Then,we discuss how to incorporate human visual intelligence with traditional tracking methods to construct more robust tracking methods.(2)We imitate the character that human pay attention to object salient features,and propose a method that try to track object by online selecting salient color features,which are measured through the contrast between foreground and background region.The proposed method performs more robust to background distractors.(3)We propose a tracking method from the motivation that human pay attention to salient regions.The poposed method try to track object by online selecting salient subregions.The discriminative subregions are detected and selected according to the center-surround difference and difference to background.Then we track each subregion and analyze its temporal consistency,and only stable subregions in spatio-temporal are used to determine target location.The proposed salient subregion based tracking method performs more adaptively to partial occlusions and background clutters.(4)We propose a tracking method based on global estimation and background context learning,from the perspective that human also pay attention to clutters in surrounding background to assist tracking.We estimate the similarity between object and background in global manner to find similar distractors,which may affect tracking.Then,we sample negative samples from the similarity distribution to learn a discriminative context model which distinguish target tracking results from similar background.The proposed tracking method combines the generative model with the discriminative model to compensate for limitations of feature representation in traditional methods,and by evaluating the reliability of tracking results,incorrect target predictions to background are rejected,which leading to less tracking drift and more tracking robustness.