Research On Block Target Tracking Technology Based On Kernel Correlation Filter And Deep Learning

Target tracking is one of the most important topics in computer vision and artificial intellegence.One of the research goals is to detect and track the state of the target in the following image frame when the target position and size of the initial image frame of a video sequence are specified.But in the dynamic situation,the target itself and its moving scene are changing all the time,and are interfered by different challenge factors at the same time,which leads to the failure of tracking.So in this paper,some occlusion,scale transformation and long-term tracking problems encountered in the process of target tracking are studied in depth,and corresponding research technology are proposed respectively.The main research results are as follows:(1)The method of block target location based on convolution feature and multi correlation filter is studied.In order to reduce the accuracy of target tracking caused by the challenge factor of partial occlusion,firstly,the target frame of initial frame is partitioned based on the condition number,then the different convolution features of VGGNet are analyzed,the first and fourth convolution features are selected,and the center positions of different partitioned targets are calculated by kernel correlation filter,and the reliability of different blocks is measured by the Barker coefficient Finally,the overall target response graph and the overall target location center are obtained based on the reliability measurement.In the whole research technology,the target block is used to optimize the situation that the tracking accuracy decreases when the target is partially occluded.(2)The target scale adaptive tracking method based on affine transformation is studied.In view of the interference factor of tracking error caused by the change of target frame scale,firstly,a series of affine transformation matrices including scaling and rotation information are estimated based on the estimated target centers of each block,and a series of target frames are obtained from the initial frame image block through affine matrix mapping;Secondly,in the image block of frame t,the target frame size obtained by affine transformation matrix is taken as the target candidate frame size of frame t with the estimated overall target position as the center,and it is taken as the target candidate frame set of frame t;Finally,the average of the tracking results of the first five frames is taken as the baseline sample,and the candidate frame closest to the baseline sample in the candidate frame set is the optimal candidate frame.This research technology uses affine matrix to realize target scale adaption.(3)The re-detection method based on long-term memory filter is studied.In view of the challenge factor of the target leaving the field of vision under the condition of long-time tracking,firstly,the complementary correlation filter is learned through the appearance features of the target to connect the long-term memory of the target leaving the field of vision again,and the confidence score value is used to determine whether the target tracking fails;assuming that the target tracking fails,in order to improve the tracking efficiency,the online support vector machine is used as the detector,By drawing intensive training samples around the estimated location and scale changes,and assigning binary tags to these samples according to their overlapping rate,SVM classifiers are trained incrementally to generate classification detection model and detect the target to be tracked again.In this research,SVM classifier is used to achieve long-term continuous target tracking.For the above research technology,simulation experiments are carried out on the data set published on OTB,and qualitative and quantitative experiments are compared with current mainstream algorithms.The experimental results show that the block target location method based on multi convolution feature and correlation filter can effectively improve the accuracy of target location under partial occlusion;the target scale adaptive tracking method based on affine transformation can effectively achieve the target scale adaptive;the re-detection method based on long-term memory filter can effectively solve the tracking loss caused by leaving the field of vision Failure,to achieve continuous tracking.