Research On Correlation Filter Based Object Tracking

As one of the most challenging problems in computer vision,the application of visual tracking can be found in a wide range of tasks including video surveillance,navigation,military,human-computer interaction,virtual reality,intelligent robot,automatic driving and so on.With about thirty years of effort,a large number of classic and excellent object tracking algorithms have been proposed.However,due to the complexity of realistic environment and target movement,the current tracking algorithms are still far from enough to meet the needs of practical applications in the aspects of accuracy,robustness and real-time performance.The accurate,robust and efficient tracking algorithms are still in highly demand.Since the correlation tracking proposed,it has attracted the attention of many researchers due to its promising performance and computational efficiency.The correlation filter convert filter operation to the frequency domain by flourier transformation,which greatly improves the running speed of the algorithm.Thus correlation filter can quickly detect the center position of the target and resampled for updating filter parameters,these ensure the accuracy and real-time performance of the algorithm.This thesis has carried on deep research on the object tracking method based correlation filter,and improved the feature fusion,scale estimation and filter update strategy.On the basis of this,correlation filter combined with state judgment and cascade object detection to achieve stable long-term tracking.The main contents and innovations are as follows:(1)This thesis first introduced the research background and significance,development status and technical challenges of object tracking,and summarized the current mainstream framework for object tracking algorithm,then the basic concept of correlation filter and its application principle in tracking is concluded.(2)In order to improve the precision and success rate of correlation tracking algorithm,the scale and learning-rate adaptive tracking algorithm based correlation filter is proposed.Firstly,the efficient feature extraction method is used as the appearance of the target sample for filter input,and aiming at the limitation that the correlation filter can't cope with the change of the target scale,based on optical flow tracking,this thesis estimates scale by the displacement of the key points between adjacent frames.Moreover,adaptive learning-rate method improved the filter updating strategy.The comprehensive utilization of efficient feature extraction,fast scale estimation and adaptive learning-rate method improve the accuracy and stability of tracker,meanwhile,this can save the computation quantity and ensure the real-time performance of the tracker.To testify the effectiveness of the algorithm,we perform comparative experiment,component analysis and qualitative evaluation on the Object Tracking Benchmark.(3)Aiming at the challenges in the long-term tracking,an object tracking algorithm based on correlation filter and cascade detection is proposed.Firstly,the improved algorithm based on correlation tracking is used as the base tracker,and it combines the tracking target state judgment and cascade detection to constitute the algorithm framework of long-term tracking.The cascade detectors include local detector based on color model,Nearest-Neighbor detector and Fine-tuning and the efficient state-judgment method of tracking target is the key to be able to start the detector in time.The cascade detection in local area can screens search samples gradually and retrieves the missing tracking target,it can improve the stability of the algorithm in long-term tracking.