Research Of Correlation Filters Network For Unsupervised Tracking By Solving Jigsaw Puzzles

As one of the most important topics in Computer Vision(CV),visual tracking is a task to estimate the position and area of the target in a given video,obtain the motion track of the target,and prepare for the subsequent advanced tasks.Visual tracking has important applications in many fields,including video surveillance,human-computer interaction,and visual navigation.In recent days,the lack of labeled data still largely limits the development of visual tracking,so this paper focus on unsupervised visual tracking algorithms.The main contents of this paper are as follows.(1)Aiming at the problem that unsupervised visual tracking algorithms cannot extract distinguishing features from unlabeled data,we train a siamese-ennead network by solving the Jigsaw puzzles in this paper.First,the image is divided into a grid,and a tile is randomly picked from each cell.Then,the image tiles are reordered via a randomly chosen permutation.After that,the network takes the image tiles as input,and predict the index of the chosen permutation.The proposed method can make the network obtain strong sematic learning capabilities and outperform the existing unsupervised visual tracking algorithms.(2)Based on jigsaw puzzles,this paper explores the impact of different feature fusion methods.The early fusion method will first fuse the extracted multi-layer features,and then use the correlation filters to predict the specific location and area of the target.The late fusion method predicts the corresponding response map for each feature using correlation filters,and then fuses the response maps to obtain the specific location and area of the target.Feature fusion can make the network get the target representations that both semantics and fine-grained details are simultaneous exploited.(3)Experiments are conducted on three large-scale databases,including OTB-50,OTB-100 and LaSOT.The qualitative and quantitative experimental results demonstrate that the proposed method can not only outperform the existing unsupervised visual tracking algorithms,but also show excellent tracking accuracy and real-time speed when compared with supervised tracking algorithms which trained by labeled data.