A Study On Adaptive Multiple Correlation Filter Algorithm For Visual Tracking

Target tracking is one of the research hotspots in the field of computer vision.Its aim is to continuously estimate the scales and locations of a preselected target in real time.The target may be subject to various external disturbances in the tracking process,such as deformation,occlusion,motion blur,scale change,background clutter,etc.,the tracking algorithm is required to handle the various situations of the target in real time.In recent years,there have been many excellent methods proposed.However,no one could address all the above challenges.This article carries out the following research work on the deficiencies of existing tracking algorithms and tracking platforms.(1)For the problem that the discriminative ability of traditional correlation filter model is limited,this paper establishes an adaptive multiple correlation filter model.The model is based on the use of high-speed calculations using correlation filter method,adaptively generates multiple filter modules,and uses the cluster method to selectively update each filter,to make different types of training data updating only its corresponding filter module.During the tracking process,each filter individually calculates the input candidate boxes and integrates the results of all filters to predict the target position.Each filter in the model could obtain a high sensitivity and discriminative ability to a particular type of target,so that the robustness of the entire tracking algorithm has been improved.(2)For the scene of fast moving target,this paper proposes an adaptive multiple correlation filter tracking algorithm based on hand-crafted features.The algorithm uses the fast computation of hand-crafted features and combines it with the adaptive multiple correlation filter model to obtain a tracking algorithm for handling fast moving targets.The algorithm achieves a speed of 65 fps on the test data set,the accuracy of distance precision rate reaches 0.735.(3)For the scene of complex situation target,this paper proposes an adaptive multiple correlation filter tracking algorithm based on deep convolution features.The algorithm utilizes the deep learning network to get the deep convolution features of rich semantic information,and the feature is input into the adaptive multiple correlation filter model to obtain a tracking algorithm for adapting to various complex scene changes.The algorithm achieves a speed of 23.8 fps on the test data set,the accuracy of distance precision rate reaches 0.852.This paper not only contrasts the algorithm on the data set,but also designs and produces a fine target tracking platform in a large field of view for the actual tracking task,for testing the algorithm in the real scenes.The platform uses adopts a dual camera cooperative working mode.One camera tracks the target in real time over a wide field of view,its maximum field of view is 6.8 x 4.9 m~2.The other camera detects various fine changes of the target in a small field of view,and obtains more accurate position information of the tracking object.The maximum operating speed of the platform could reaches 60 fps.