Ship Vision System For Recognition And Tracking Of Sea-and-Air Multi Targets

China is a major maritime country with vast sea areas,long coastline and abundant marine materials.Therefore,the development of ship vision system is vital importance in our country.In military areas,it can be used to find enemy troops,strengthen oversight of maritime areas and safeguard maritime rights and interests.In the civilian areas,it can be used to manage customs transportation,safeguard border and coastal security,dispatch port and prevent guilty of misconduct at sea.At present,most of the recognition and tracking of sea targets systems merely focus on ships.But in practical applications,objects in air(such as aircrafts,birds)and other objects at the sea surface(such as islands,reefs,larger fishes on the sea)may also become the monitoring targets of ship monitoring.This thesis focuses on the development of ship vision system that is able to accurately recognize and track multi-target of sea and air in real time.The main researches are as follows.Firstly,the monitoring images based on the digital video at sea are sensitive to lighting conditions,such as night,cloudy,sea foggy.All of these can degrade the quality of the monitoring video images.In order to make better use of the collected ship monitoring video in the subsequent process,image enhancement and image filtering are utilized to improve the quality of monitoring video images.And the suitable datasets of this thesis are set up based on open source datasets and the collected video images.And then,all of images are processed into the unified format and packaged.In this way,the corresponding training set and test set can be used in the training of identification network.Secondly,the multi-target of sea and air recognition algorithm is designed based on convolutional neural network.By analyzing and comparing the speed and accuracy of common recognition algorithms,this thesis uses YOLO v3 to identify multi-target of sea and air.The Inception module is added to the YOLO v3 in order to improve the feature extraction network of YOLO v3,so that the fine feature information of sea and air targets can be extracted.According to the test results,the improved recognition network shows more effective recognition and can accurately identify small or occluded target.Then,the multi-target of sea and air tracking algorithm is designed based on video sequence.This part of the thesis mainly studies the KCF tracking algorithm.The confidence judgment mechanism is utilized to improve the KCF tracking algorithm.Utilizing the detector designed in multi-target of sea and air recognition algorithm,one can obtain the class and position information that can help initializing the tracker.Then,the data association algorithm is used to realize multi-target of sea and air tracking.Subsequently,the detector could continue detecting and correcting the tracker,to prevent false detection,missed detection and target loss,and ensure the robustness and accuracy of the multi-target tracking algorithm.Finally,the interface is designed by Qt software and the overall performance of ship vision system is debugged.The ship vision system is constructed by recognition algorithm and tracking algorithm.And the overall test of system shows that the system can accurately identify and track multi-target of sea and air in real time.