Research On Visual Target Tracking Algorithm Based On Correlation Filter

Visual target tracking technology is a vital part of computer vision.The traditional correlation filtering algorithm running with high precision and speed is widely used in target tracking.There exist numerous applications in real world,such as tracking of target characters in various sports competition scenes,vehicle and pedestrian tracking problems at traffic intersections and target tracking issues under rotation camera and drone camera.Because the occlusion attribute and the rotation attribute(in-plane rotation and out-of-plane rotation)of the video sequences have the most weight in the tracking database,based on the correlation filtering algorithm,several improved algorithms are derived to resolve the target occlusion challenge and the target rotation challenge.(1)For the target tracking problem involving occlusion attribute video sequences,in term of color image,an occlusion evaluation criterion is proposed to optimize the correlation filter tracking algorithm.First of all,by virtue of the spatial regularization model,the target output response is obtained,then,corresponding the average peak correlation energy is derived,also,combing the target response value with the corresponding target position yields the mass response function;In addition,the occlusion evaluation criterion is generated by the weight coefficient to determine whether the target is occluded;Finally,according to the occlusion condition,the model is adaptively updated.The accuracy and success rate of the algorithm compared with STRCF algorithm in the dataset OTB2013 are increased by 0.7566% and 0.7599%,respectively.Moreover,the accuracy and success rate of the occlusion attribute are increased by 0.5200% and 0.9539%,respectively.(2)For the target tracking problem involving occlusion attribute video sequences,in term of depth image,the concept of overlap rate is proposed to optimize the correlation filter tracking algorithm.Firstly,a multi-template algorithm using an iterative operation to select an optimal classifier is derived;Then,the concept of overlap rate is given to determine whether the target is occluded in the depth map;Finally,according to the occlusion condition,the model is adaptively updated.The accuracy and success rate of the algorithm compared with DSKCF algorithm in the Princeton dataset are increased by 4.67% and 7.04%,respectively.(3)For the target tracking problem involving rotating attribute video sequences,in term of color image,the fusion background information and multi-peak detection strategy are proposed to optimize the correlation filter tracking algorithm.First of all,the target loss function of the traditional correlation filtering is optimized by virtue of the background information around the target and the target response that obeys the Gaussian distribution;Then,multi-peak detection strategy is derived to select candidate target peaks from the target output response map;Finally,the target response under different scale weights is gained,and the maximum response value is selected to obtain the final target position and target scale.The accuracy of the algorithm compared with SAMF algorithm in the dataset OTB2013 is increased by 7.9561%,the success rate of the algorithm compared with CACF algorithm is increased by 6.8259%.Besides,the accuracy and success rate of the in-plane rotation attribute compared with DSST algorithm are increased by 5.2326% and 2.1467%,respectively.Furthermore,the accuracy and success rate of the out-of-plane rotation attribute compared with SAMF algorithm are increased by 8.4270% and 7.6377%,respectively.