Research On Surface Ship Detection And Tracking Method Based On Deep Learning

With the rapid development of Marine transportation industry,more and more ships are coming and going from the port,and the safety problems of port and waterway are increasingly prominent.Therefore,it is particularly important to install unmanned monitoring system in the port and ship to realize automatic detection and tracking of surface vessels.Ship detection and tracking is an important method to obtain ocean target information and has wide application.For example,it is widely used in the field of maritime intelligent monitoring,the supervision of maritime rights safeguarding fishing and the discrimination of maritime combat situation.However,the traditional ship detection and tracking method is slow in calculation and low in accuracy.The Marine environment is complicated,so the automatic detection and tracking of ships are limited.This paper studies the surface ship detection and tracking method based on deep learning to improve the accuracy of ship detection and tracking.The specific contents are as follows:Firstly,according to the characteristics of the surface ship surface ship detection method study,on the basis of the existing ship data sets,through the Internet to collect the surface ship more pictures manual annotation,expanded its data set and set up the surface ship image data sets,and then establish a framework based on depth study of the surface ship detection network.The network framework based on YOLOv3,which introduced the Mixup data enhancement,residual connection,the attention mechanism module,bottom-up path enhancement,CIOU loss function and significance test method,and multiple methods are combined for optimization.The method is tested on the surface ship image data set to select the combination mode with the best effect and verify the feasibility of the proposed method.Secondly,the surface ship tracking method is studied,and a surface ship tracking method based on improved Deep Sort is proposed.The YOLOv3 network framework is learned and trained through the surface ship image data set migration,which is used as a detector.The improved Deep Sort method was used as the tracker to optimize the similarity calculation formula and correct the position of the tracking ship with successful correlation,so as to improve the tracking accuracy of the ship and carry out secondary data correlation for the tracking target with failed correlation.Experiments were carried out on surface vessel video and open data sets of MOT16 and MOT20,and compared with Deep Sort to verify the effectiveness of the proposed method.Finally,A cascade method of surface ship detection and tracking based on deep learning is studied.First of all,the improved YOLOv3 network framework was used as the detector,which was combined with the improved Deep Sort algorithm in a cascading way to form a cascade method for detection and tracking of surface ships,and realize the automatic detection and tracking of surface ships.Experiments were carried out on the surface ship video and the open data set of MOT16 and MOT20,and compared with Deep Sort method and improved Deep Sort method.The results showed that the surface ship detection and tracking cascade method reduced the times of missing and false detection in the surface ship video,and improved the tracking accuracy.