Detection And Tracking Algorithm Of Vessels Based-on Multi-source Information Fusion In Maritime Supervision

With the development of shipping,multi-source information fusion has become an important direction to increase the safety of ship navigation.The current VTS system uses AIS,radar and CCTV cameras to identify,track and supervise ships on the way.But these three still have some defects.AIS information acquisition depends on shipborne equipment.Radar often produces false alarm in water surface environment under the influence of water wave and weather factors,and the information update rate of these two is slow;At present,the images taken by shore-based cameras still need to be processed manually.With the development of computer vision technology,it is possible to introduce cameras that can collect rich information and update quickly into ship target detection,tracking and information fusion.Under the condition of saving labor cost,it is possible to carry out dynamic monitoring of ship navigation and improve water perception.Aiming at the ship detection and tracking algorithm based on multi-source information fusion of maritime supervision,the main work of this paper is as follows:Firstly,this paper studies the image ship target detection algorithm based on deep learning,the detection results of detectors Faster R-CNN,YOLO9000 and YOLOv3,as well as the backbone networks of Squeeze Net,Res Net50 and Dark Net53 are compared and analyzed,the corresponding simulation results are given,and the structure of YOLOv3 detector is modified.In view of the small scale of the existing open-source ship dataset,which is not suitable for ship target detection in complex weather on water,this paper studies the data expansion method based on Atmospheric scattering model,which can better simulate the ship image collected in fog and improve the detection effect of the model in complex weather conditions;Secondly,this paper studies the multi-target online tracking of video ship,and compares and analyzes the tracking effects of Kalman filter,extended Kalman filter and traceless Kalman filter with motion model of constant and constant acceleration.Aiming at the situation of trajectory switching and omission when matching only using the state prediction of the filter,combined with a variety of features including label features,depth features and position features,the cost matrix is improved to reduce the occurrence of trajectory errors caused by matching failure and improve the effect of ship multi-target tracking;Then consider the data preprocessing before multi-source information fusion.The real-world position of the ship target in the image can be calculated through the camera matrix.Aiming at the problem that it is difficult for the shore-based camera to calculate the matrix through the calibration object with the same size as the ship,a method is proposed to calibrate the shore-based camera using the information provided by the shipborne class-A AIS and eliminate the influence of the real-time water level,and a method for calculating the real-world coordinates of the ship target according to the pixel information in the target detection frame.In addition,the target information of AIS and radar,and the methods of coordinate unification and time alignment of different information sources are introduced;Finally,this paper studies the track fusion between information sources,and introduces the two structures of detection fusion and track fusion suitable for AIS,radar and camera.For the radar and shore-based camera that produce false alarm,a track quality selection method is designed,and the route scene is designed and simulated.The results show that the shore-based camera can be used as a supplement to AIS and radar to improve the effect of information fusion.In addition,track fusion using local track quality selection can achieve better results.