Video Target Tracking Based On Deep Contourlet Network And Salient Attention

In computer vision,object tracking is a fundamental task with various applications,including video surveillance,human-computer interaction,military attack,etc.In recent years,many methods have been developed under different scenarios and tested on different large datasets.Even so,due to the complexity and variance in real applications,it remains a challenge in computer vision with many unsolved problems.There are several branches of visual object tracking.In this thesis,we focus on the problem of single object tracking.Single target tracking can be divided into feature extraction and target recognition,in which target recognition includes target location and state estimation.In this thesis,non-subsampled contourlet transform and salient attention mechanism are used to enhance the feature expression based on the two main frameworks of current tracking algorithms,so as to provide better discriminant basis in different subtasks of target recognition.The main contributions of this thesis are as follows.First,the appearance of object may change a lot in consecutive frames due to the unexpected transformations,including view angle changes,illumination variation,rotation,etc.In this thesis,a target tracking algorithm based on the deep contourlet network is proposed.Non-subsampled contourlet transform is used to extract multi-scale and multi-direction image information.The weighted fusion method is used to enhance the feature expression of the target effectively.On this basis,a kind of deep contourlet network is designed,which have the generality of extracting features.The position response of the extracted feature is obtained by the correlation filter.The experimental results show that the proposed method is more effective than other correlation filter algorithms,not only in dealing with the moving change,but also in background interference.Compared to the baseline algorithm,the success rate for the OTB2013 increased by 1.2%,and the expect average overlap for the VOT2018 increased by 0.8%.Second,above method using multi-scale search strategy to obtain inaccurate bounding boxes.We propose a target tracking algorithm based on salient attention and siamese network,which extracts the features of candidate regions of interest by training the deep structure on large-scale datasets.In this thesis,a cris-cross attention method,which optimizes the computational load of spatial attention,is used to weight cross-correlation operations to improve discriminate performance.Experimental results show that the proposed method can estimate the target state more accurately than other siamese network algorithms,and the visual effects are more intuitive.Compared to the baseline algorithm,the success rate for the OTB2015 increased by 1.9%,and the expect average overlap for the VOT2018 increased by 1.66%.Third,a target tracking algorithm based on deep contourlet network and salient attention is proposed.Based on the above two improvements,we analysis the problems of siamese networks and present a network structure which can update the template to adapt to the change of the target appearance.The network structure has two branches,which use residual attention to enhance the classification features,and use the deep contourlet network to extract features from the main network.Finally,estimate the object on the maximal intersectionover-union prediction network.Experiments on OTB2013,OTB2015 and VOT2018 verify the accuracy and stability of this method.The success rate on OTB2013 reached 0.674,the success rate on OTB2015 reached 0.669,and the expect average overlap rate on VOT2018 reached 0.3863.