Research On Visual Target Tracking Algorithm Based On Sparse Coding

In recent years,sparse representation theory has drawn wide attention of scholars,and the visual tracking algorithm based on sparse coding achieves good performance.We focused on the sparse coding in visual object tracking(VOT)process and systematic study,which achieved the following research results:1.When the target is disturbed by the illumination changes and background clutter,the tracking results of the tracking algorithm with sparse coding exist the drift phenomenon.We therefore proposed a novel visual tracking algorithm based on dictionary learning.The dictionary was constructed with a set of target templates and background templates,and the low dimensional dictionary and linear classifier were trained simultaneously by using the label consistent K-SVD dictionary learning mechanism.In the update stage of particle filter,the likelihood function of each particle was calculated by using the classification results and the sparse coding histogram.Finally,the dictionary,linear classifier and target template histogram were updated by the dictionary learning method.Numerous experiments on various challenging videos show that the proposed algorithm has better tracking performance than some benchmark methods in the scenarios with the interference of occlusion,illumination change,background clutter.2.Taking into account the inconsistency between fuzzy measurement and particle filter caused by error of sensor measurement and signal processing in target tracking field.We propose an adaptive sparse coding visual tracking algorithm based on box particle filter.The proposed algorithm uses the box particle filter to approximate the posterior probability density of the target,which reduces the computational cost of the tracking algorithm.At the same time,Gaussian kernel distance measure is adopted for the sparse coding sampled patch,which considers the spatial distribution structure of the patch.In the stage of template update,the target template is adaptively updated according to the occlusion status of the current tracking result,which avoids the drift phenomenon caused by over-update of the template.The experimental results show that the proposed algorithm has good anti-interference and robustness against complex scenes such as background clutter,occlusion and illumination changes,and the template update mechanism is more adaptable to complex tracking scenarios than using fixed parameters.3.In order to solve the problem of target number change,target occlusion and scale change in visual multi-target tracking,a dictionary learning sparse coding algorithm based on multi-Bernoulli filter is proposed.First,the label-consistent K-SVD(LC-KSVD)method is introduced to train the low-dimensional dictionary and the background classifier of the targets.Then the multi-Bernoulli(MB)random finite set theory is employed to approximate the posterior probability density of the visual multi-targets,and the particle filter(PF)technique is implemented to extract the states of the multi-targets.Where the likelihood model of the particle is constructed by combining the classification results and the sparse coding histogram,which can improve the ability of visual multi-target tracking.Experimental results show that the proposed algorithm can effectively achieve the visual multi-target tracking with thenumber change of the targets,and has good adaptability to target occlusion and scale change.