| Multi target tracking of ships in videos is one of the important tasks in surface traffic management,which can provide effective quantitative information for water traffic management departments.In recent years,with the rapid development of deep learning theory,multi-objective tracking methods based on deep learning have attracted more and more researchers’ attention and have become the mainstream direction of multi-objective tracking research.The Deep SORT algorithm can utilize deep feature information to achieve object tracking under occlusion.However,this algorithm has the following problems in ship tracking scenarios: the model needs to be trained through a large number of ship datasets for feature extraction,resulting in poor adaptability of the model after scene replacement;Secondly,due to the complete occlusion problem in the scene and the similarity of ship appearance features,it leads to misidentification in the target tracking process.This thesis proposes a Deep SORT model based on SIFT features to address the issue of low tracking accuracy caused by complete occlusion of ship targets in video scenes using traditional methods.SIFT features do not require pre training and have fast computation speed,thus adapting to different scene requirements.Moreover,SIFT features are invariant to rotation,scaling,and brightness changes,and stable to some extent to view angle changes,affine transformation,and noise.The experimental results show that the proposed method can effectively improve the accuracy of ship tracking.In response to the slow tracking speed and inability to meet real-time requirements,this thesis proposes a reserve pool optimization method to improve the computational efficiency of the model.Provide the optimal reserve pool capacity for ship tracking problems through experiments to achieve balance between accuracy and speed during the tracking process.The experimental results show that the proposed method can meet both tracking accuracy and real-time requirements. |
PDF Full Text Request