Research On Object Tracking Algorithm Based On Kernel Correlation Filtering

Object tracking has always been worth studying in the field of image processing.A qualified object tracking algorithm requires that the targets of other frame sequences be accurately located based on the object information of the first frame of the video sequence.The application of object tracking can automate and intelligentize the self-management of the intelligent monitoring system.And in patients with medical treatment of lesions can be tracked.It also can make the dynamic unknown environment more transparent and controllable in Aerospace.And as the cost of hardware declines year by year,target tracking will definitely have greater potential for development in more fields.The main contents of this paper include:First of all,this paper introduces some basic knowledge of target tracking,including the significance of target tracking and the status of development at the present stage,and combs the mainstream algorithms of target tracking.Make a simple introduction from the type of object tracking,and make a brief introduction to the mainstream evaluation system.Secondly,based on the understanding of several classic object tracking algorithms,this paper studies an object tracking algorithm based on kernel correlation filtering.The main work is to build a tracking system framework.On the basis of the kernel correlation filter tracking framework,multichannel features are used to deal with the changes of illumination.Aiming at the difficulties in algorithm rotation,deformation,scale adaptation and target adaptation,we track the target in video sequence by integrating spatio temporal context information.The spatial relationship is inherent in the target and the target neighborhood,and the time is obtained by the weighted spatial relationship of accumulated historical frame,establish the statistical target and the surrounding area in the low level features through the Bayesian framework,so it is good in does not affect the tracking speed based on solving the problem of adaptive target change,and can handle partial occlusion and the degree of deformation of a small.Finally,when the tracking system framework is completed,in order to increase the ability and universality of the algorithm to deal with complex scenes.Based on the depth convolution model trained by the detection dataset,use the video sequence samples from the tracking standard library to transfer learning and tuning of the model.The features learned by the deep convolutional network are directly applied to the tracking framework of the correlation filter.The depth features of the object are used as the input of the correlation filter tracker to calculate the response value,which is the object position.And in the actual tracking tasks,the multi-target tracking.And track multiple object in the actual tracking tasks.The Experimental results show that our algorithm can cope with most challenges such as occlusion,rotation and illumination changes,and achieve fast tracking.