Research On Correlation Filtering Target Tracking Method In Video Monitoring Based On Deep Learning

Target tracking is one of the hot research directions in the field of computer vision,and it is widely used in video surveillance,drones,and autonomous driving.In recent years,under the concept of a smart city,higher requirements have been imposed on the tracking accuracy of targets in video surveillance.Although significant progress has been made in tracking algorithms based on correlation filtering,the changing environment in video surveillance makes target tracking algorithms still face tracking failure problems such as occlusion,rotation,and deformation.This paper studies the apparent modeling method in the target tracking algorithm,and discusses the enhancement of the diversity of training set samples.The main research contents of this paper are as follows.First,this paper proposes an algorithm for multi-feature adaptive fusion.In this paper,a correlation filter is used as a framework to extract the target's HOG features and use a convolutional neural network to extract high-and low-level convolutional features.A method of peak area ratio and response similarity of two adjacent frames is proposed to evaluate the contribution of different features to the tracking results.Calculate the contribution of each different feature separately to get the weight ratio of feature fusion,give larger discriminative features a larger weight,and obtain the target estimated position based on the fused response map.Then calculate the target scale through the scale correlation filter to complete the tracking.The experimental results show that the algorithm of this paper has different image characteristics based on the expression of features,and adaptively fuses the layered convolutional features and HOG features to better express the appearance model of the target and has strong tracking robustness.Secondly,in order to improve the training sample set,a method of diverse training sample set is proposed.The clustering algorithm is used to increase the types of samples.The similarity of the samples within each sample subset is high,the differences between different subsets are large,and the stored features are different.This sample clustering model makes up for the shortcomings of the traditional sample training set and increases the discrimination ability of the tracking filter.In the filter update,a variable learning rate is selected to control the contribution of the filter corresponding to different samples to improve the tracking performance of the related filters.The results show that compared with the mainstream algorithms and the multi-feature adaptive fusion algorithm proposed in Chapter 3,the algorithm in this paper is more tracking robust and achieves the expected improvement.Finally,in order to reflect the application value of the algorithm in this paper,a Graphic user interface experimental platform based on the Matlab programming environment was developed.The user-oriented operating system has the ability to read the tracking video,change the parameters,display and track in real time,save,and evaluate.Track these five modules of results.The user interacts with the underlying algorithm through the buttons of the Graphic user interface platform.Based on this experimental platform,this paper selects two sets of airport monitoring videos for experiments.The tracking results prove the tracking robustness of the algorithm in this paper,and the experimental platform has certain applications value.