Research On Correlation Filtering Target Tracking Algorithm Based On Convolutional Layer Inference

Target tracking comes from the common gaze behavior in human life.It is one of the most basic and important tasks in computer vision.It is widely used in military,security,humancomputer interaction,medical clinical diagnosis and other fields.The current hot research tracking algorithms are mainly correlation filtering tracking algorithm and deep learning tracking algorithm.In the CSK correlation filtering tracking algorithm,regularized least squares is used to calculate the template,and the cyclic displacement operation is used to expand the training set.The nonlinear kernel is introduced by defining the kernel space in the training image of cyclic displacement.Because of its high speed and accuracy,it has become a typical model in the field of target tracking.However,as a correlation filtering algorithm,there is a coupling between the size of tracking area and template in the CSK tracking algorithm.This coupling causes only a small part of the tracking image to be tracked in the CSK tracking algorithm.In actual scenes,due to factors such as rapid movement of objects,dropped frames or poor accuracy,the target will be shifted out of the tracking area,which will cause the process of tracking to fail completely in the CSK tracker.Aiming at this shortcoming of CSK tracker,an improved CSK target tracking algorithm is proposed.The convolutional layer inference method in deep learning is used to decouple the size of tracking area and template.An improved CSK algorithm in which the tracking area can be freely selected is obtained,and three template calculation methods(CSK template cropping,Alternating direction multiplier method,gradient descent method)are introduced.On the OTB100 data set,the mean precision of improved CSK tracking algorithm under the same configuration is 10.8% higher than that of the original CSK tracking algorithm,verifying the effectiveness of the algorithm.In addition,in order to speed up the convolutional layer inference process,an efficient convolutional layer inference algorithm based on fast Fourier transform is also proposed,the speed of convolutional layer inference is significantly improved in the case of large convolution kernel size.In the case of a convolution kernel size of 50,the speed of the fast Fourier transform convolutional layer inference algorithm is more than 100 times faster than that of the ordinary convolutional layer inference algorithm.