Research On Target Tracking In Remote Sensing Videos Based On Deep Convolutional Network

In recent years,with the development of remote sensing earth observation technology,such as video satellite technology and drone technology,target tracking technology based on remote sensing videos has broad application prospects in the fields of intelligent traffic management,video surveillance,battlefield dynamic analysis and so on.At present,there are many related researches based on remote sensing video target tracking,but most of them only use single or several combined features to represent the target.The feature expression ability of deep learning is very powerful.The features obtained by deep convolutional network contain more abundant information.The tracking method based on deep convolutional network has made breakthroughs in close-range video.However,there are still many difficulties and challenges in directly applying such methods to the field of remote sensing video: 1)Remote sensing video has complex background,large size,and is taken in a bird's-eye view,which is different from the target size and viewing angle in close-range video.2)In remote sensing videos,the target size is small resulting in less features.So,the same kind of objects are similar.And the interference of similar objects greatly affects the tracking accuracy.3)In practical applications,target tracking in remote sensing videos requires higher accuracy of tracking results.In order to solve the above problems,this thesis develops a target tracking in remote tracking video method based on deep convolutional network.While maintaining real-time performance,the accuracy of tracking is enhanced to suppress interference targets.Main tasks as follows:1.Based on the continuity of the video and the essence of the target tracking,that is,identifying and tracking the target instead of any other object,the Siamese network is proposed to be applied to the remote sensing video target tracking.The basic process of target tracking using Siamese network is analyzed and summarized.The experiments on the UAV dataset and remote sensing satellite video dataset demonstrate that the Siamese network can track small targets in large-scale remote sensing videos and achieve real-time requirements.2.Aiming at the problem of low the accuracy of remote sensing video tracking,this paper proposes a target tracking method based on Siamese network and template updating.The Siamese network is pre-trained on the target recognition data set.The template image is only the calibration position of the first frame,but the tracking accuracy will be degraded for the target whose direction or appearance changes drastically.First,the target is to track the video sequence,and the time-consistent drone video training network can make the network model more targeted.Then,according to the target prediction position of the previous frame,the template image is updated with the weighted average idea.The location of the target is determined by calculating the similarity between the template image and the search area.Experiments on the UAV dataset and satellite video dataset show that the proposed method can effectively improve the tracking accuracy.3.For the case of similar interference objects around the target of remote sensing video tracking,the correlation filtering layer is constructed according to the correlation filtering ideas and embedded in the Siamese network.And the target discriminating ability of the network is improved.The depth convolution network is used to extract the feature of the input template image and the search image,and the correlation filter layer filters the template image feature correlation,which can effectively suppress the interference target and improve the tracking accuracy.Experiments on the UAV dataset and satellite video dataset prove that the proposed method can effectively enhance the anti-interference ability and improve the tracking accuracy.