Object Tracking Based On Correlation Filter

Object tracking is one of the important research topics in computer vision,which is widely used in intelligent video surveillance,autopilot,intelligent transportation,aerospace,virtual reality,and other fields.In recent years,object tracking has developed rapidly,and the object tracking algorithm based on correlation filter has attracted attention.However,after several years of development,the object tracking algorithm based on correlation filter has entered a bottleneck period,and it seems that it has been impossible to achieve better results with correlation filter alone.In this thesis,based on the correlation filter object discrimination module,combined with the position correction module based on Kalman filtering and the bounding box optimization module based on Io U network,a single object visual tracking algorithm using multiple information is proposed.The main innovative work includes:(1)Based on the existing correlation filtering algorithm,the object discrimination module is optimized to improve the ability of object discrimination.The training set sample incorporation scheme is improved to obtain a more representative and simplified sample set;the updating frequency of filter template is adjusted to reduce the computation and over fitting,and at the same time,a more robust filter template is obtained.(2)In order to obtain the optimal object position output,a position correction method based on Kalman filter is proposed,which uses the motion information of the object to correct the object position obtained by the object discrimination module.(3)The bounding box optimization module based on the Io U network is designed.By using the general expression information of the object bounding box,the non-maximum suppression and gradient rise optimization of the candidate bounding box are carried out to obtain more accurate object bounding box,which effectively improves the accuracy of object tracking.The proposed algorithm is tested on many object tracking datasets.The results show that the performance of the algorithm in this thesis is better than many object tracking algorithms based on correlation filter.The AUC score of this algorithm on the OTB-2015 dataset is 69.8,which surpasses the classic correlation filter algorithm ECO and the recently proposed Siamese network algorithm Siam RPN++.The proposed algorithm is compared with the top 10 algorithms in the VOT2018 competition.It has the best Accuracy and Expected Average Overlap(EAO)score.The AUC of the algorithm on the UAV123 dataset reaches 64.8,which is 19% higher than UPDT and 23.4% higher than ECO.The indicators of this algorithm on the Tracking Net dataset greatly exceed UPDT,the AUC score exceeds 17%,and the normalized accuracy exceeds 10.3%.The AUC score of the algorithm in this thesis on the NFS dataset is 61.3,which greatly exceeds the C-COT by 25.6% and the UPDT by 14%.