Long-term Target Tracking Algorithm Research Based On CMT Framework

Visual target tracking technology has become a world wide hot spot as it is an important branch of computer vision,in recent years,and numerious tracing algorithms have reported.However,the difficulties namely as illumination instability,image shadows,target occlusion and disappearance reproduction persist a chanllenge to the common algorithms the real-time performance of which is needed to improve.This thesis selects the CMT target tracking algorithm,and starts from the above-mentioned key issues,starting from the feature point threshold,global matching,target position prediction and occlusion reemergence,and proposes a long-term target tracking algorithm based on CMT framework.The experiments exhibits the long-term tracing algorithm presented in this thesis is an effective method to deal with the foregoing issues(the illumination changes,image shading,target occlusion,disappearance,reproduction).The algorithm real time performance has been improved siginificantly by compared with the original CMT algorithm.(1)The Parzen-window nonparametric probability density estimation method and the image optimal threshold solving algorithm based on KSW entropy method are studied.The two functions are combined to construct the objective function based on Parzen-window and entropy.Finally,a threshold adaptive FAST feature point detection algorithm is proposed.The algorithm replaces the parameterized probability density estimation in the original KSW entropy method by using the coordinate space information of the image pixel.It does not need to artificially assume the distribution of the problem,so it can better adapt to the complex image feature points detection in practical application.Then,the global preliminary extraction and local dynamic adjustment of the FAST feature points are performed by using the solved dynamic global threshold and the dynamic local threshold.The algorithm can overcome the interference of illumination change and image shadow.Compared with the original algorithm,the feature point extraction effect is more stable and robust.(2)In order to solve the problem of large number of mismatches in global matching of feature points in CMT algorithm,a new measure of similarity matching is adopted in this thesis.This method is used to measure the similarity of feature points before global matching,which can remove some feature points that will mismatch.The experimental results show that,compared with the original algorithm,the accuracy of global matching of feature points is effectively improved by adding the proposed similarity matching measure.(3)In the face of complex images,the increasing number of feature points extracted by the original CMT algorithm leads to the inadequate real-time performance of the algorithm.In this thesis,a cascaded Kalman filter local detection algorithm is used to improve the tracking speed.Then,the proposed threshold adaptive FAST feature point detection algorithm is used to detect feature points in the predicted region of the target.Finally,it achieves faster feature point extraction in the face of complex images.The actual test proves that the tracking speed of the cascaded Kalman filter is 6.43 times faster than that of the original algorithm.(4)Aiming at the problem of target occlusion,this thesis designs a re-detection module based on the matching rate of foreground points,and combines the above three improvement measures,finally proposes a long-term target tracking algorithm based on CMT framework.The long-term target tracking algorithm based on CMT framework uses foreground matching rate to real-time monitor whether occlusion occurs and occlusion degree.Combining with the tracking snapshots collected by neighboring frames,the background information of the target is updated when the severe occlusion is about to disappear,and the re-detection module is started to re-initialize the tracking algorithm for continuous tracking.Compared with TLD,KCF and the original CMT algorithm,the proposed algorithm has the ability to re-detect the target in abnormal situations such as long-term occlusion or disappearance and recurrence.Compared with other long-term tracking algorithms,the overall performance of this algorithm is better in long-term tracking.