Research On Object Tracking Algorithm Based On Correlation Filtering

Moving object tracking has always been a key research direction in the field of computer vision at home and abroad,and plays a key role in the fields of intelligent security,smart city,medical image and other video information processing.From the early generative object tracking algorithm to the recently high-profile discriminant object tracking algorithm,each object tracking algorithm is accompanied by the emergence of new technology in the field of object tracking,and the technological innovation continuously promotes the breakthrough in the performance of object tracking algorithm.However,in the actual tracking scene,the tracking object and the tracking environment often have unpredictable changes,such as illumination variation,object rotation,and fast motion,which will have a great impact on the performance of the tracker.From the perspective of feature fusion and feature selection,this thesis proposes two improvement strategies to improve the algorithm as a whole.(1)In view of the poor adaptability of the color histogram features in the Staple object tracking algorithm to background interference and illumination variation,fixed feature fusion weights and model drift caused by model updates,this thesis proposes a feature self-control under the control of a confidence mechanism adaptive fusion object tracking algorithm.Firstly,CN color feature and HOG feature,which have stronger color expression ability,are used to represent the object,and CN color feature has certain invariance to illumination variation and background interference in the tracking process.Secondly,ridge regression was used to calculate the filter response output of the two kinds of features,and the two confidence indexes of smoothing constraint and average correlation peak energy of each feature were calculated,and the fusion weight of the features was adaptively assigned according to the confidence index.Finally,the confidence index of APCE is used to judge the reliability of HOG and CN features,and the adaptive model update is realized.(2)On the basis of the adaptive feature fusion target tracking algorithm,this thesis introduces the perception area,and directly combines the global context perception information into the training of the correlation filter.On the basis of almost no loss of the real-time performance of the algorithm,the correlation filter is improved for the negative Sample information discrimination ability.At the same time,the feature descriptor is pre-trained for the sensitive tracking environment changes of CN color feature and HOG feature,so as to improve the ability of the two features to deal with the tracking environment changes.Combined with the variation coefficient,the feature fusion coefficient is calculated so that the features can be fused adaptively in the complex and changeable tracking environment.In this thesis,two standard video sequence sets,OTB-2013 and OTB-2015,the algorithm in this thesis and the current object tracking algorithm with better tracking effect are compared with simulation experiments,and the tracking performance of the algorithm is evaluated by two indexes,range accuracy and tracking success rate.Experiments show that the tracking success rate of this algorithm is more than 7% higher than that of the Staple object tracking algorithm.At the same time,when the challenging tracking environment(Illumination Variation,Object Rotation)appears,the tracking success rate of the Staple object tracking algorithm is increased by about 10%.The algorithm has good robustness to various tracking interferences such as illumination changes,object rotation,fast motion,etc.