Research On Moving Object Tracking Algorithm

Moving target tracking technology involves many advanced application tasks such as video understanding,analysis and behavior prediction.And it has been one of the hot issues in computer vision for many years.However,the complicated factors bring challenges to the development of target tracking technology,such as: scale change,deformation,fast motion,camera shaking,illumination and background clutter,etc.The unpredictability of these changes poses challenges to the development of target tracking technology.Based on the analysis and research of related principles,in order to improve the accuracy and robustness of the target tracking algorithm,the main work is as follows:1)An improved tracking-learning-detection(TLD)target tracking algorithm is proposed for the problem of low real-time and robustness of TLD algorithm,which improves the performance of the tracker and detector through the kernel correlation filter(KCF)and frame difference method.First,the KCF tracker is used to correct the tracking failures of optical flow tracker in illumination.Secondly,the foreground detection is introduced into the detector.The coarse-to-fine detection strategy eases the huge computation resulted in the large number of sliding windows.Finally,the detection results are used to update the filter model of the tracking module to ensure the efficient operation of the tracking system.The results show that the success rate and precision of the proposed algorithm are 20.4% and 17.3% higher than the TLD algorithm,and the tracking speed is 53.9 frames per second.It has better robustness and real-time performance in dealing with scenes such as illumination change,scale change and occlusion.2)Aiming at the problem that correlation filtering tracking methods are difficult to recover after after the target leaves the field of vision,a target tracking algorithm based on cascade neural network and correlation filter is proposed.Based on the KCF algorithm,a multi-task cascade convolutional neural network detector is introduced to correct the errors of the tracker and update the filter model.In addition,in order to ensure the tracking speed of the algorithm,the detector is started only when the tracker fails to track or the tracking accuracy is below a threshold.Experiments show that the success rate and precision of the proposed algorithm are 81.6% and 82.7%.The algorithm solves the problem that the correlation filter tracking algorithm fails to track when the target is severely blocked,and ensures the real-time and accuracy of the tracking target.3)To track multiple targets simultaneously,a real-time tracking algorithm by assigning target ID is proposed.Using the framework of tracking-learning-detection,a fixed ID is assigned to each target,then the single target is extended to multiple targets.The positive samples of each target are independent,while the negative samples are uniform.In each ID,the tracker and detector run in parallel.After obtaining the tracking results and detection results of all objects,the learning model synthesizes results and updates the target model.Results show that the success rate of the proposed algorithm is 80.1%.And the proposed algorithm can achieve good tracking performance.