Studies On Visual Object Tracking Algorithm

Target tracking algorithm has formed its own system during half a century of development, the study of target tracking algorithm is contending and flourishing. Traditional target tracking algorithms research tend to classify target tracking algorithm by object representation, target tracking strategies and so on, but now target tracking algorithm often combines multiple object representation or tracking strategies to achieve stable tracking in complex environment. So this thesis divides target tracking algorithm into four modules, feature extraction, tracking strategy, similarity measurement and template updating. On this basis, this thesis studies the main contents as follows:Firstly, this thesis studies target tracking algorithm based on sparse representation, and on this basis, in order to accurately determine the target occlusion, this thesis calculates reconstructing residual in block to accurately determine the extent of target occlusion; in order to improve the accuracy of the algorithm tracking under occlusion, this thesis improves the tracking model, this new model adjusts the allowed range of trivial coefficient energy by the extent of target occlusion; in order to solve the sparse coefficient of improved tracking model, this thesis introduces the proximal acceleration gradient algorithm to solve this kind of optimization problems without constraints. Experiments show that the target in scaling, occlusion conditions can achieve a good tracking.Secondly, to solve tracking drift and rotation problem in the long time tracking scenarios, this thesis takes advantage of the traditional template matching tracking algorithm and the local feature tracking algorithm, then an algorithm combining global features and local feature is proposed, its key is using the change of target posture calculated by the local feature point set to aid the tracking algorithm based in global features, In long time tracking experiments, the algorithm performs well when the target rotates and scales. In the occlusion experiments, the algorithm can track target accurately when the target is occluded and can catch the target after severe occlusion.Finally, there are many problems that any good automatic video surveillance system needs to cope with, such as object detection, tracking under occlusion, exits and entries of object from the scene and so on. To solve these problems, this thesis proposes a series of algorithm. In order to extract objects more accurately, this thesis extracts objects from the scene using a background subtraction method, then introduces shadow removal algorithm and union filter algorithm. To solve the inter-object occlusion problem, this thesis determines the target occlusion by calculating overlap of the unmatched target and all the detected regions. To solve the problem of exits and entries of objects from the scene, this thesis establishes a minimum initial observation parameter and a maximum missed observation parameter, In the true outdoor experiments, this algorithm can detect moving targets, track targets under occlusion and determine the exits and entries of targets.