Research On Object Tracking Algorithms Based On Spatial And Temporal Regularization

Object tracking is one of important topics in the field of computer vision,and it has wide applications in many fields such as virtual reality,behavior recognition,video surveillance,and traffic detection.In the process tracking,the object variations(e.g.deformation)and the interferences from the external environment(e.g.occlusion)often make the tracking results unsatisfactory.Therefore,it is a key problem for object tracking to design a robust tracking algorithm.Focusing on the object variations and the environment interferences,this thesis has designed some effective object tracking algorithms to improve the accuracy and robustness of the tracking results in the framework of particle filter,deep learning,and correlation filter.The contents mainly include an object tracking algorithm with a temporal sparse collaborative appearance model,an object tracking algorithm based on spatial and long-short temporal attention correlation filter,and object tracking algorithm based on verifying networks.The concrete contents are shown as follows.1.Object tracking algorithm with a temporal sparse collaborative appearance model.Firstly,under the assumption that the overall feature information of the target can better distinguish the target and the background,a discriminant sparse similarity model with time constraints is designed.In the meanwhile,under the assumption that the local feature information of the target can handle the target's appearance variations well,a sparse generation model called a similarity measurement model is designed.Secondly,the discriminative score model and the similarity measurement model are combined to generate a sparse joint appearance model to obtain the discriminant score for each candidate target.Finally,an update mechanism of the target sample is designed for the discriminant model and the sample's histogram in the similarity measurement model.Proposed object tracking algorithm considers both the global feature information and local feature information of the target.Experimental results show that the tracking algorithm can achieve better results in the accuracy and robustness.2.Object tracking algorithm based on spatial and long-short temporal attention correlation filter.Firstly,considering that fast motion or motion blur aggravates the interference of boundary effects,a new spatial weighted map is designed to reduce the impact of fast motion or motion blur.Secondly,considering that frame-to-frame information transfer and inheritance can improve the stability,a regularization term with a short-term memory is proposed.In the meanwhile,considering that the deformation or occlusion of the target will cause the target drift,a regularization term called a long-term memory is proposed and it is related to the initial frame.Experiments on some data sets show that the accuracy and robustness of the tracking algorithm have been significantly improved.3.Object tracking algorithm based on verifying networks.Considering the idea of correction on the deep learning based tracking algorithms,we have designed a new neural network structure model.The proposed algorithm consists of two networks: tracking network and verifying network.In the tracking network,considering the fusion of deep features and shallow edge features,a multi-input residual network is designed to learn the relationship between the target and its corresponding Gaussian response map to obtain the position information of the target.In the meanwhile,in the verifying network,a shallow chain discriminate network is designed to verify the target obtained by the tracking network,and the tracking network is updated according to the verified results.Experiments show that the accuracy and robustness of the tracking results of the tracking algorithm have been significantly improved.