Research On Anti-occlusion Object Tracking Algorithm Based On Multiple Instance Learning

Object tracking is widely used in human life,so it has become the popular topic in the field of computer vision.In recent decades,researchers have proposed many excellent tracking algorithms to deal with various problems in the real scenario.However,some problems still affect the performance of tracking algorithm,such as occlusion.Therefore,designing a robust anti-occlusion object tracking algorithm is a main research emphasis in the field of object tracking.By circulant matrices and kernel trick,the kernelized correlation filter has the high tracking accuracy while keeping the fast tracking speed.However,occlusion can seriously affect the object model and the location of the kernelized correlation filter.Therefore,the improved algorithms based on the kernelized correlation filter are proposed to deal with occlusion.The main researches of this thesis are as follows:A object tracking algorithm based on multiple instance learning and the kernelized correlation filter is designed to deal with occlusion during target tracking.Firstly,the occlusion detection mechanism is established by combining the classification ability of multiple instance learning and the distribution feature of the maximum response value of the kernelized correlation filter.Then,when the occlusion detection mechanism determines the target is occluded,the re-detection mechanism is activated to search the target and updating the model is suspended to prevent the model from being interfered by occlusion.Finally,a scale correlation filter is constructed to determine the scale of the target.The experimental results show that the proposed algorithm can effectively deal with occlusion.In order to improve the tracking speed while keeping the high tracking accuracy of the tracking algorithm in the occlusion scene,an object tracking algorithm based on the dual occlusion detection is proposed.Firstly,the initial occlusion detection mechanism is established based on the average peak-to-correlation energy of the kernelized correlation filter.Then the local binary patterns feature and the cosine similarity are used to design the occlusion fine detection mechanism.If the initial occlusion detection mechanism determines the target is occluded,the occlusion fine detection mechanism is used to further determine whether the target is occluded,which can not only guarantee the occlusion detection function of the tracking algorithm,but also reduce thecomplexity of the tracking algorithm.When the target is occluded,the weighted re-detection mechanism is activated to search the target and updating the model is suspended.The algorithm performs well in anti-occlusion and possesses the fast tracking speed.