Research On Occlusion-handling Tracker Based On Discriminative Correlation Filters

As an important branch of computer vision,object tracking has been widely used in human-computer interaction,intelligent video surveillance,robot visual navigation,unmanned driving,military defense and other fields.However,for a long time,due to problems such as changes in illumination,occlusion,background similar interference,and scale changes,object tracking cannot be widely applied to the industrial field.In the face of a wide range of application scenarios,it is a reasonable choice to improve the accuracy of object tracking under occlusion and realize the large-scale application of object tracking.This paper firstly researches and summarizes the technology development route of the object tracking algorithm in recent years.In addition,it analyzes and studies the basic principles of convolutional neural networks,and the effectiveness and advancedness of CNN(Convolutional Neural Networks)in extracting image features.After summarizing the advantages and disadvantages of the occlusion handling algorithm in recent years,I realized that the current tracking algorithm mainly has two problems:(1)the object tracking algorithm based on correlation filters will cause model degradation due to occlusion;(2)under long-time occlusion,the speed of object tracking algorithm based on correlation filters drops seriously.Therefore,this paper focuses on solving the problem of the effectiveness and real-time performance of the algorithm in the case of occlusion.In view of the above problems,this paper combined with the STRCF(Spatialtemporal Regularized Correlation Filters)model,proposed an anti-occlusion object tracking algorithm based on correlation filters,this algorithm mainly includes two technical points:(1)introduced a quality evaluation function and the occlusion determination function based on PSR(Peak Side Lobe Ratio)can accurately determine whether the target is occluded.(2)when the object is blocked,the block matching method is used to determine the blocking status.When the target is partially occluded,STRCF is used for target tracking,and when it is completely occluded,the Kalman filter is used for position estimation and the feature matching method is used for precise positioning.Extensive experiments on OTB-100,Temple-Color-128 and VOT datasets show that the proposed tracker performs favourably against the state-of-the-art methods in terms of occlusion and achieves a faster tracking speed than STRCF.In terms of speed,our tracker outperforms its baseline STRCF with gains of 8.8%.Moreover,on VOT-2018,our tracker outperforms its baseline STRCF with gains of 4.8% in EAO and 4.6% in accuracy respectively.Occlusion-Handling tracker based on discriminative correlation filters proposed in this paper can effectively solve the problem of tracking failure caused by model degradation,improve tracking efficiency and accuracy,and can meet the needs of moving target tracking in complex industrial environments,which is beneficial to promote the target tracking algorithm is widely used in actual production and life.