Inverse Sparse Target Tracking Based On Deep Learning Constraints

Target tracking is a very important breakthrough in the field of computer vision,and it is also a major research direction for many researchers.At present,target tracking has been widely used in intelligent monitoring,intelligent transportation,human-computer interaction,and medical imaging and so on.The target tracking problem can be understood as estimating the trajectory of a certain target in each frame of the video sequence.However,in the target tracking process in the real scene,the target may appear to scale changes,posture changes,movement speed changes,occlusion phenomena,etc.There are also factors such as drastic lighting changes and complex background interference in the scene,which will interfere with the accuracy of target tracking and reduce the performance of the tracking algorithm.Therefore,it is still a challenging problem to design a good tracking algorithm to achieve effective tracking of the target.In this paper,through the in-depth study of the basic theory and key technologies of video target tracking,as well as analysis of the current domestic and foreign outstanding tracking algorithms.Starting from the basic modeling methods of video tracking problems,the inverse sparse representation target tracking algorithm based on deep learning constraints is proposed in combination with the current deep learning theory.The main research work of the article is summarized as follows:Firstly,a convolutional neural network structure is constructed,which is used to classify candidate samples and extract the local features of the target through an off-line pre-training model.Second,this paper analyzes the advantages and disadvantages of the discriminative model and the generative model,and proposes a target appearance model of the joint model.In the discriminative model,by using the transfer learning idea,the classification of candidate samples can be realized by online fine-tuning the convolutional neural networks.In the generative model,in order to solve the problem that it is difficult to construct an overcomplete dictionary,the theory of inverse sparserepresentation is introduced,and the L2 norm minimization also is used to solve the inverse sparse coefficient to improve the accuracy of the algorithm.Finally,an inverse sparse representation target tracking algorithm based on deep learning constraints is proposed.On the OTB(Object Tracking Benchmark)of the public test set platform,we conduct a comprehensive comparison and analysis with tracking algorithms based on different features or different classifiers or different sparse constraints.A large number of experimental results demonstrate that the proposed algorithm can track the target well in the complex environments such as illumination and scale change,partial occlusion,and fast motion,and validates the robustness and stability of the proposed algorithm.