Research On Video Tracking Algorithm Based On Correlation Filter

Video target tracking,a research hot spot and difficult point in the field of computer vision,is widely applied in intelligent video monitoring,military guidance,man-machine interaction,intelligent transportation,video retrieval,etc.Video target tracking algorithm based on the correlation filter of good tracking accuracy and tracking speed has become a main stream algorithm in recent years.However,it still needs further study due to the interference of illumination variation,occlusion,border effect,etc.This paper proposes two improved algorithms on the basis of the correlation filter algorithm.(1)First,this paper proposes a kind of wide area search tracking algorithm integrated with multi-featured descriptor on the basis of the ECO tracking algorithm frame,thus further improving the real time feather and precision of tracking.The main work is as follows.First,it integrates the DAT descriptor and HOG+CN descriptor,which increases the accuracy and robustness of feature extraction and reduces the border effect of target by using background perception algorithm.Second,it constitutes the final wave filter with the original filter linearity by using the border response map and training the new filter through the BACF frame.Third,it optimizes the model update and sample model weight selection method.The result through the comparison with the current mainstream algorithm shows that the average DP is improved by 1.6%and average OP is improved by 2.1%compared with the ECO-HC algorithm before improvement.(2)Second,this paper proposes a kind of empirical background perception filter and re-detection tracking algorithm in terms of the object occlusion.First,it judges the target occlusion degree according to the manual features of the correlation filter extraction,the maximum response value and average peak value in the target search area.Second,it classifies the target occlusion degree according to judgment result.The SiamFC algorithm is used for the re-detection of the target in case of target occlusion and the target model has self adaption and upgrade according to matching confidence.The frame target model update operation can be discarded if obstructed.The algorithm is tested for its performance in OTB-50 video set and compared with other mainstream target tracking algorithm.The result shows that the average DP is improved by 4.3%and the average OP is improved by 1.1%compared with BACF algorithm before improvement.