Research On Multi-Object Tracking And Azimuth Estimation Algorithms For Ships On The Sea Surface

The realizations of multi-object tracking of ships on the sea surface and the estimation of ship object azimuth can be applied to many applications,such as maritime supervision,etc.However,due to the influence of many factors such as sea surface imaging conditions and ship motion,the performance of ship detection and multi-object tracking will be seriously degraded.Different from the estimation of ship azimuth from remote sensing images,the information loss during horizontal view imaging procedure is serious,and the appearance difference is small for different types of ships with the same azimuth.Therefore,how to improve the performance of multi-object tracking and azimuth estimation of ships on the sea surface in complex scenarios is still one of the important problems to be solved.The research in this thesis is conducted to solve the above problem and the main work is as follow:(1)YOLOv5 detection model is adopted to realize the multi-object tracking of sea surface ship under DeepSORT framework,and the influence of the main hyper parameters on multi-object tracking performance is discussed.Firstly,YOLOv5 model is trained using a detection dataset containing two types of ships,which are military ship and civilian ship,respectively.Secondly,to verify the influence of Re ID model on association performance of DeepSORT algorithm,the tracking performances using Re ID model with different loss functions are compared after constructing a ship re-identification dataset.Finally,a ship multi-object tracking test dataset with a total of 3310 sequence images are constructed to verify multi-object tracking performance of sea surface ships,.The experimental results show that,MOTA is improved by 5% when YOLOv5 is used as detection model instead of YOLOv3.Due to large appearance differences of the same ship caused by azimuth change,Re ID model has little effect on ship multi-object tracking performance.After experimental comparison and analysis of hyper parameters in DeepSORT algorithm,MOTA for the constructed test dataset reaches 73.3%,which is 0.6% higher than the default parameters.(2)To realize image-based ship azimuth estimation,a YOLOv5 ship azimuth detection model based on the idea of angle classification is proposed.On the basis of the aforementioned ship detection dataset,8 azimuth intervals are defined according to actual needs,and the self-developed software is used to add azimuth information as a property to the annotation file,and a dataset for ship direction detection is constructed.To realize ship azimuth detection based on YOLOv5,the direction classification dimensions are added to the output module,and multi-task joint optimization of object position prediction,class prediction and angle prediction is realized through decoupled design.In the multi-scale feature fusion stage,CBAM mixed attention is used to improve fusion performance and Focal loss function is used to reduce the impact of uneven angle distribution.The experimental results show that,m AP50 for the improved YOLOv5 model has increased from 49% to 67%,where m AP50 of the proposed model has increased by 13% using decoupled design,2.1%using CBAM attention module,and 4.4% Using Focal loss function,respectively.(3)To solve the problem of ship azimuth estimation in video images,the result of Siam Mask tracking algorithm is used to extract the boundary line between ship and sea and estimate the ship azimuth from the perspective of object tracking and instance segmentation.Firstly,the dataset used to train the Siam Mask model was constructed by manual labeling based on the ship detection dataset.After completing Siam Mask model training and obtaining ship segmentation results,the ship-sea boundary line is first estimated by projection of object mask image,and then least square method is used to fit a straight line,and the slope of the straight line is approximated as the ship azimuth.The experimental results show that,the method proposed in this paper can realize ship azimuth estimation when only part of the target area is selected because the segmentation results of Siam Mask is used to realize pixel-by-pixel classification.The performance of azimuth estimation using the proposed method are more stable and accurate compared with the method using min-area bounding box estimation.