Research On Correlation Filter Tracking Algorithm Via High Confidence Strategy

Object tracking has been a hot research direction in the Computer vision field,and it has broad application prospects in real life.Correlation filtering is a mainstream framework in the field of target tracking.Many excellent algorithms have been proposed by many scholars under this framework.It solves the basic problem of target tracking and also has better accuracy.However,in practical applications,they still face many challenges,such as illumination variation,occlusion,scale variation.In view of these challenges,this thesis mainly improves the tracker from the two modules of feature extraction and template update on the basis of previous work.The main work of the thesis is as follows:(1)In the stage of feature extraction.Aiming at the failure of some correlated filter trackers via a fixed weight feature fusion in a specific scenario,the correlation filter tracker via an adaptive feature fusion with a confidence region of response map is proposed for enhancing the arithmetic robustness.In the method,according to the peak state of the response map,the confidence region is the region in which each response value is higher than the expectation of the response map,and the fusing weighs of a HOG response map and a color histogram response map at every frame can be calculated by the confidence region of the HOG response map.This method achieves the adaptive fusion of two features.In particular,it performs well in illumination variation and scale variation(2)In the stage of template update.Aiming at the failure of some trackers via template drift is caused by the continuous updating of the template after the object is occluded,the correlation filter tracker via a high confidence template update strategy is proposed for enhancing the arithmetic robustness.The method discriminates the confidence of the current tracking result according to the fluctuation of the rate of change of the maximum response value sequence.In this method,the confidence of the current tracking results is obtained according to the fluctuation of the change rate of the sequence of maximum response values.When the random error of the rate of change is more than 3?,the target is considered to be deformed or occluded.At this time,the template update is stopped to prevent the classifier from learning the negative sample.When the random error of the rate of change is greater than 4?,it is considered that the template is contaminated at this time,and the current template needs to be traced back to the confidence template to ensure the reliability of the template.According to the above method,a high confidence template update strategy is implemented,which overcomes the interference of short-time occlusion.For the above algorithm,this thesis conducts experimental analysis based on 100 videos in the OTB-100 benchmark dataset.The experimental results show that the adaptive feature fusion algorithm is superior to other tracking algorithms in the scene of illumination change and scale variation and in the occlusion environment,the high confidence update strategy algorithm improves the performance of the tracker.