Research On Object Tracking Algorithm Based On Correlation Filter And Deep Neural Network

Visual object tracking is an important research task in the field of computer vision,which is realized by modeling the object according to the initial state of the given object,then evaluating the state of the object in the subsequent video sequence,and finally giving the location and scale information of the object.As one of the key techniques in image processing,object tracking is of great significance in the field of target detection,recognition,segmentation,image analysis and understanding.With the rapid development of artificial intelligence technology,many object tracking algorithms are proposed in recent years,and the research of object tracking technology has made remarkable achievements.However,due to the complex background clutter,occlusion,deformation and low frame rate video problems often occur in the process of object tracking,it is easy to result in wrong or incomplete feature information in the process of object apparent modeling and updating,which weakens the discriminative ability of the tracker and leads to object tracking failure.Therefore,object tracking remains a challenging and practical problem in the real scene application.Based on the existing researches,this dissertation carries out in-depth research on visual object tracking,and the main content and innovative points are summarized as follows:1.A saliency detection algorithm based on region contrast and guided filter optimization is proposed for the problems of occlusion,target disappearance and reappearance in the process of target tracking.The proposed algorithm can detect the salient object from the image quickly and accurately,preserve the edge and texture information of the object,highlight the salient region and weaken the non-salient region.Then,combined with the correlation filtering method of multi-feature fusion,the correlation filter tracking based on saliency detection is proposed.The framework of the tracking algorithm mainly includes target feature extraction,target feature response,reliability evaluation,re-detection module and feature fusion response module.The reliability evaluation module can effectively detect whether the tracking result is reliable,if it is reliable,the target state will be given directly,otherwise,it will start the re-detection module,so the object position can be obtained again.In this dissertation,several tracking datasets are used for evaluation and test,and the results of experiments show the effectiveness of the proposed algorithm.2.Aiming at the problems of low frame rate videos,such as motion blur and large appearance change between adjacent frames,a new convolutional neural network framework based on joint training of frame interpolation sub-network and twin symmetric object tracking sub-network is proposed,and an end-to-end framework is formed.The frame interpolation network can effectively alleviate the problem of target deformation or blur caused by improving the video frame rate,while the tracking subnetwork will track efficiently during the frame interpolation.The end-to-end framework not only avoids the problem of inconsistent targets and mutual influence between multiple models,but also reduces the complexity of the project.In addition,the proposed loss function and the strategy of joint training interpolation network and tracking network make the model achieve the optimal performance.Lastly,the proposed algorithm in this dissertation is tested on multiple datasets,and the results of experiments show the proposed algorithm can effectively improve the tracking performance in low frame rate videos.3.The circulant shift of the correlation filtering tracking algorithm will lead to the undesired boundary effect,which will produce some non-real samples,especially when the object is affected by background clutters and fast motion,the boundary effect will have a greater impact.To solve the above problems,an adaptive spatially regularization and interframe constraint correlation filter tracking method are proposed in this thesis.Firstly,the proposed algorithm learns a more robust correlation filter algorithm by introducing the spatial regularization and interframe constraint term into the objective function,which makes the correlation filter model more discriminative,and also reduces the influence of boundary effect.Then,three sub-problems with closed-form solution are solved by using the alternating direction multiplier method,which helps solve the problem of large amount of calculation.Lastly,the experimental results evaluated on OTB and VOT datasets demonstrate the robustness and efficiency of the proposed algorithm.Besides,the proposed method of this dissertation also obtains a good results on long-term La SOT datasets.