Research On Correlation Filter Based Object Tracking Algorithm

Object tracking is a significant and challenging research topic in computer vision.Object tracking plays an important role in many fields including security,intelligent transportation,military,human-machine interaction.The high variety on object appearance caused by external or internal factors makes the tracking problem challenging.External factors come from environment changes such as illumination and angle,and the internal factors of appearance variations come from object scale changing and occlusion.In recent years,the correlation filter based tracking is hot topic in the filed of tracking.The correlation filter based tracking transforms the tracking problem into the problem of separating the target from the background,and the maximum response is considered to be the new position of the target.The tracking speed is improved due to the correlation filter based tracking uses fast Fourier transform to convert the calculation to the frequency domain.Besides,the correlation filter based tracking updates the classifier and template in real time to handle the changes of the object's appearance and the scene.This thesis carries on the deeply analysis to the correlation filter based tracking method,and propose a series of improvements.The main results as follows:(1)Aiming at the single feature has poor performance in tracking,the thesis proposes a method that fuse color feature and Histogram of Oriented Gradient feature in decision layer.First,the color feature and Histogram of Oriented Gradient feature are extracted to train ridge regression classifiers,and multiple tracking results are obtained by using different classifiers,then fused the different results by the weights assigned according to the peak to sidelobe ratio.This method can improve the accuracy of the algorithm.(2)To address the traditional correlation filter based tracking method cannot handle the object scale change,this thesis introduces information measure method into correlation filter.The object scale is predicted by comparing the information of two consecutive frames.This method can adjust the size of tracking frame according to thesize of object,and complete the correlation filter based tracking method scale adaptive improvement.(3)In order to improve the robustness of the correlated filter based tracking method,a strategy of adaptively updating the learning rate is proposed.The similarity of two adjacent target regions is used to analyze the appearance change of the target,and the learning rate is adaptively adjusted according to the appearance change.This method can improve the performance of the correlation filter tracking method in occlusion and target deformation scenarios.The proposed algorithm is evaluated with visual tracker benchmark.The experimental results show that the tracking algorithm performs best in terms of tracking accuracy and success rate compared with the traditional correlation filter tracking algorithm and shows good robustness at the same time.