Research On Object Tracking Algorithm Based On Correlation Filtering And Deep Learning

Nowadays,computer vision technology has been widely used in face recognition,driverless,intelligent monitoring and other fields.One of the key technologies in these applications is object tracking.At present,there are two main technical routes of object tracking algorithm,which are based on correlation filtering and depth learning,based on these two technologies,many excellent tracking algorithms have been developed.Scenarios of object tracking are often complex and changeable,in complex scenes such as object occlusion,lighting changes,and scale changes,objects are easily lost during the tracking process,the tracking tasks in these scenarios are a great challenge for object tracking algorithms,so it is of great significance to develop more robust object tracking algorithms.The thesis will optimize and improve the tracking algorithm based on correlation filtering and deep learning to deal with tracking tasks in complex scenarios.The object tracking algorithm based on correlation filter has great advantages in tracking speed and tracking accuracy,and has strong robustness,at present,the object tracking algorithm based on correlation filter emerges in endlessly,which has also become a hot research direction.Therefore,this thesis will improves the object tracking algorithm based on correlation filtering,to solve the problem of the model drift that inaccuracy of the tracking algorithm in the case of occlusion and interference will lead to the positioning failure of the object in the next frame,so that the follow-up tracking algorithm cannot continue to complete the tracking task.The thesis uses multiple features of the response map fusion to filter the interference factors in a single feature to enhance the representation of the response map,and then adds scale estimation to find the best scale of the object.At the same time,in the case of tracking failure,in order to ensure the tracking effect of subsequent video frames,object re-detection is needed to relocate the object.After experimental comparison,the improved algorithm has improved accuracy and success rate.Siamese network structure is the main research direction of tracking algorithm based on deep learning.The thesis proposes a tracking algorithm based on multi-layer feature fusion that based on the Siam FC algorithm used siamese network structure.The feature extraction network of object tracking algorithm based on siamese network is improved,and then the final feature has the advantages of different convolution layer features by combining shallow features and deep features.Finally,the template is updated by update strategy to prevent model drift,which improves the robustness of object tracking algorithm.For low-resolution object tracking,a siamese network tracking algorithm based on attention mechanism is proposed.In the feature extraction network of siamese network,attention mechanism model is integrated,combined with the idea of multi feature fusion,depth feature and hog feature are fused to improve the ability of feature representation.The experimental results show that the performance of the improved algorithm has been further improved,and good results have been achieved in the task of low-resolution object tracking.