Research On Ship Multiple Object Tracking In Remote Sensing Image Based On Deep Learning

Ship target tracking is always the key point in the field of object tracking technology,which is of great significance to sea surface monitoring and national defense security.The most important thing in modern war is information contention and utilization.Making full use of the sea surface information of high-resolution remote sensing image is of great significance to improve the comprehensive observation ability of our army’s sea battlefield,to coordinate the command of the whole army and to strike the target accurately in a long distance.Ship multiple object tracking based on remote sensing image plays an important role in information collection,sea surface early warning and battlefield dynamic analysis of modern information war.This thesis mainly studies the multiple object tracking of ships in the visible light remote sensing image,and the specific research contents and results are as follows:(1)This thesis first introduces the application,development history and current research status of the object tracking technology of remote sensing image in various fields,and summarizes the difficulties existing in the current object tracking technology based on remote sensing image.(2)This thesis introduces the related ideas of deep learning and the theory of multiple object tracking.Firstly,it introduces the basic knowledge of deep learning and the common architecture of multiple object tracking framework.The mainstream multiple object tracking framework mostly adopts detection based tracking strategy,which is mainly composed of object detection,state prediction and data association.This thesis mainly introduces several one-stage and two-stage object detection methods.And several data association methods in multiple object tracking framework are introduced and compared.The performance evaluation index of multiple object tracking algorithm is also introduced,and the main evaluation indexes of this thesis are determined.(3)In this thesis,two deep learning based multiple target tracking algorithms,SORT and Deepsort,are studied.The network loss function is modified and the Deepsort ship tracking model is optimized.The ship targets in the remote sensing image are collected,and the ship target data set based on the remote sensing image is made.The ship depth feature model is trained.Experiments show that Deepsort algorithm has high tracking accuracy,but also see the shortcomings of the algorithm in the current real-time scene,which determines the research direction for the next work.(4)Finally,aiming at the shortcomings of the previous thesis,a new multiple object tracking algorithm based on improved YOLO v3 is proposed.The method includes three parts: object detection by improved YOLO v3,feature prediction and matching,data association.In the aspect of target detection,the improvement of YOLO v3 mainly includes modifying the network structure,prediction prior frame and loss function of YOLO v3.The ship depth feature model is retrained on the ship target data set.According to the target detection information,Kalman filter is used to predict the target position.In feature matching,a MGN multiple granularity network is introduced to extract the appearance information of the target,and the Mahalanobis distance is used to correlate the target motion information.According to the number of targets,the priority is set,and the data association process is completed by using Hungarian algorithm,and the detection results and prediction results of unmatched targets are further processed.Finally,the experiment proves that the algorithm proposed in this thesis has achieved good results in multiple object tracking of ships in remote sensing images.Compared with Deepsort algorithm,the tracking accuracy of this method is improved by a small margin,but the algorithm speed is more than twice as fast as possible,and the problem of lack of real-time performance is solved.