Computer Vision-based Ice Channel Identification And Ship Target Tracking In Ice Areas

With the intensification of global climate change,the strategic position of the Arctic is becoming more and more prominent due to its special geographical location and rich mineral resources,and at the same time,new requirements and research topics are raised for the development and utilization of the Arctic ice channel.Many technical problems need to be solved to ensure the safety and economy of navigation in the Arctic ice channel.Advanced icebreakers have excellent structural strength and advanced technology to clear a channel from the ice field,and it is a common scenario for merchant ships to navigate in the ice channel or follow the icebreakers.Identifying the ice channel from the first viewpoint of the ship and tracking the target ship ahead is the basis for realizing the automatic driving of ships in the ice channel.The research of identification technology and ship target tracking technology of polar ice area channel is helpful to ensure the safety and economy of navigation in ice area channel.The research work and achievements of this thesis are as follows:(1)In this thesis,we first analyze and summarize the construction of synthetic datasets,and propose to build an ice channel scene based on Unreal Engine to automatically generate ice channel pictures and use image annotation tools to annotate the channels to obtain synthetic ice channel datasets for the lack of real ice channel datasets.(2)Conduct research on the analysis and production of ship target tracking datasets and ice channel datasets.Existing marine datasets such as Mar DCT,Seagull,SMD,and LMD-Tship are analyzed in terms of image quantity,image quality,annotation quality,and work condition diversity,and the marine dataset to be used is determined.(3)Conduct research on two-staged ice channel identification method based on image segmentation and corner point regression network.The principle of the image segmentation network used is described,and the structure and loss function design of the corner point regression network are analyzed.Based on the image segmentation network,the navigable area is extracted from the channel picture,and based on the corner point regression network,the channel lines on both sides are extracted from the navigable area,and the two stages are combined to achieve the identification of the ice area channel.Based on the synthetic ice area channel dataset,the channel identification accuracy and identification speed are evaluated and compared with the non-intelligent angle-based filtering clustering method.(4)Conduct research on ship target tracking analysis based on Transformer model predictor.The integration of a priori features of ship tracking targets is achieved by offline training on the dataset and online training during tracking,so as to separate the ship targets from the ocean background and accurately predict the ship target bounding box.Based on the professional marine tracking dataset LMD-TShip,we compare and analyze the differences in tracking accuracy and success rate between the ship target tracking method in this thesis and the advanced tracking methods in recent years,and analyze the robustness of the tracking method in various marine conditions based on the dataset.