Study On UAV Vision Target Tracking Based On Deep Network

Among the many research contents in the field of computer vision,target tracking is one of the popular research techniques.After decades of research by computer vision researchers at home and abroad,target tracking has been popularized in our daily lives and applied to various industries,such as intelligent transportation,human-computer interaction,and video surveillance.Although there are already plenty of target tracking algorithms with superior performance,numerous difficulties have taken place in performing target tracking tasks when used in drone scenarios.Problems such as rapid target movement,severe occlusion,and too small targets in the drone scenarios affect the effectiveness of target tracking.Meanwhile,lack of unified labeling of aerial photography data sets and incomplete data sets are also important factors affecting target tracking in drone scenarios.This paper conducts a more in-depth study of target tracking algorithms in terms of the challenges of target tracking in drone scenarios.Before deep learning was applied to the field of target tracking,the correlation filtering method had achieved certain advantages in improving target tracking performance.And a large number of excellent target tracking methods based on Siamese Network emerged after the emergence of SiamFc,but the tracking effect will be affected if these algorithms are directly applied to target tracking in drone scenarios.Considering the incompleteness of the aerial photography related datasets,lack of data sets,and inconsistent labeling,this paper proposes to train the UAV target tracking model by using the unsupervised method.And due to the limitations of the drone's own computing power,this paper uses lightweight network and forward tracking and multiframe backward verification to achieve the unsupervised target tracking model.In addition,according to the characteristics of drone aerial data,such as small targets,high data density,and many background interference factors,this paper proposes a method of using attention mechanism to extract target's major information in video images.At the same time,the network model combines the advantages of Siamese network and the correlation filtering.In order to verify the performances of the two target tracking model structures designed,this paper uses the VisDrone2019 data set for experiments.And the results show that the two UAV target tracking models designed in this paper have better performances and more excellent effectiveness than other algorithms of the same type in the application of UAV target tracking.