The Key Technologies Of Multiple Vessel Detection And Tracking Based On Computer Vision

Driven by the global "Industry 4.0" and "Made in China 2025",technologies such as information,computer,communication,network and artificial intelligence develop rapidly.As the core content of maritime power strategy,the innovation of marine science and technology becomes the key domain of competing for the maritime leader and right of voice.The intelligent vessel is a significant aspect in the field of the innovation of marine technology and is widely studied in the shipping field and scientific research institutions around the world.As the key technology of intelligent vessel,the information perception technology based on computer vision,which effectively overcomes the limitation of marine radar and automatic identification system in aspect of detecting and tracking for the object in sea,it assures the safety of vessel navigation.In the intelligent vessel field,the six-degree-of-freedom low-frequency shaking,high-frequency jitter of vesselborne cameras,and the lack of the public datasets of visible light at sea impedes the development of vessel detection and tracking.Therefore,this thesis explores the key technologies of multiple vessel detection and tracking based on computer vision,including sea-sky line detection algorithms,vessel motion feature modeling based on the image sequence,vessel detection algorithms,and multiple vessel tracking algorithms.In the sea-sky line detection algorithm,to solve the problem poor real-time and instability of the current algorithms,this study combines the data from vesselborne camera and inertial sensor and modifies the fast estimation model and the accurate detection model.The former analyzed the six-degree-of-freedom motion of the vesseborne camera on the sea.The sea-sky line estimation model was built based on the rolling angle and pitching angle data that is provided by inertial sensor.The latter applies the modified edge detection model to obtain the image edge in the candidate region,and uses the modified Hough transformed model to convert the edge information to Hough space.It realizes the accurate detection for the position of the sea-sky line.The experimental results show that the estimation model has low estimation error.The accurate detection model has good robustness for the false detection and missing detection of the sea-sky line with 95.71%for the precision and 96.88%for the recall,respectively.In the process of modeling of vessel motion features in image sequences,aiming at the problem of vesselborne camera's autokinesis,the vessel's motion characteristics in continuous frames are analyzed.The optical flow motion model is constructed through four steps,including feature points matching,sea-sky lines matching,feature points classification,and homography matrix calculation.It realizes the accurate motion estimation of dynamic and static targets,which provides an accurate potential location for multi-target vessel detection and tracking.In the marine vessel detection algorithms,the existing algorithms are improved by three types of algorithms.The algorithm1 applies the improved minimum barrier distance transformation algorithm in the candidate region to segment the target vessels,which realizes the fast extraction of target vessels and keeps the same level of accuracy.The algorithm2 uses the optical flow motion model to obtains the potential boxes of the vessels in the stable image and applies the image classification model to localize and identify the target vessels on the sea.The algorithm3 not only improves the YOLOv3 model from two aspects,including the block detection and adaptive anchor box clustering,but it also achieves the high accurate detection for vessels based on Singapore maritime dataset and self-built dataset.The test results show that,compared with the discrete cosine transform algorithm,the average precision and running speed of the algorithm1 is improved by 12.36%and 62.07%respectively,which makes greatly progress rather than the performance of the traditional detection algorithm.The average precision of the algorithm2 reaches at 90.03%,which is higher than SSD-resnet50-fpn,Faster-rcnn-resnet50 and YOLOv3 algorithms.The average precision of the algorithm 3 is 95.63%,which is higher than other comparative algorithms and achieves better vessel detection performance.In the algorithms of marine multi-target vessel tracking,the modified algorithm based on Deep SORT is studied.For the motion model,the optical flow motion model is used to replace the Kalman filter model,which eliminates the influence of the vesselborne camera's autokinesis.For the appearance model,the improved Resnet50 classification model is applied to replace the deep cosine model as the backbone network for re-identifying the vessel's identity.For the similarity measurement,the joint measurement is used to replace the cascade measurement for effectively matching the tracking boxes with the detection boxes,and the Hungarian algorithm is used to correlate the data between them for realizing the tracking of multi-target vessels frame by frame.The test results show that the overall performance of proposed algorithm is significantly better than other algorithms,and 4 and 5 performance indicators obtain the best tracking performance in the onshore and onboard datasets,respectively.The research works make up the deficiency of traditional information perception device,quickly establish a clear and intuitive consciousness of context perception,improve the ability of vessel collision avoidance,anti-piracy and rescue,ensure the safety of life and property at sea,provide reliable technical support for the development of intelligent vessels,provide reliable technical support for the development of intelligent vessels,promote the innovation of marine science and technology,and promote the development of marine power strategy.