Design And Verification Of Maritime Target Detection Algorithm For Small USV

With the advantages of small size,good maneuverability,wide application range,and low cost,the unmanned surface vessel(USV)has good application prospects at the civilian and military levels.Hence deep research on USV not only has great economic value,but also has important and profound significance for China’s national defense.The detection and recognition of maritime targets are one of the most important key technologies in the research of USV.It is used to identify and detect the obstacles around the USV in real time and accurately,and provide a guarantee for safe navigation.Compared with other methods,vision-based target detection is suitable for USV for the reason of low cost and wide application range.However,the existing algorithms are mostly verified in the laboratory and short of the understanding of reality application demand.And it could lead to pool realtime performance and low accuracy when applying them to USV directly,resulting in the weakened ability to sense the surrounding environment and the inability to effectively guarantee safe navigation.Therefore,this thesis builds a vision system based on a small USV and designs a maritime target detection algorithm based on deep learning according to actual demand,which is tested in the lake.The main research contents of this thesis are summarized as follows:For the problem of low accuracy and poor real-time performance,a maritime target detection model is proposed based on YOLOv3.After analyzing the application requirements and the structure of classic networks,a backbone network for feature extraction based on four basic modules is designed.Furthermore,we optimize the architecture of YOLOv3 and design an improved maritime object detection model Net-Y.By replacing the traditional convolution in the backbone with the deep separable convolution,we propose a smaller model,called Net-S.Compared with the Net-Y,the parameters and computation are reduced by 45% and 35%,respectively.For the problem of performance degradation when applied in USV,we analyze the application demand and characteristics.Moreover,we construct a special dataset for the maritime target.We re-label more than 75000 images selected from the public dataset and the self-collected,and expand the data with image enhancement.Then the Net-Y and NetS with intersection over union loss function are trained on the dataset.Compared with other algorithms,the performance of the proposed model is improved.With the input resolution of 720 × 1280,the accuracy of the Net-Y and the Net-S model reach 83% and 81%,and the detection rate is 12 FPS and 17 FPS respectively.For the problem of limited computing power on the controller of the USV,the Net-ST with faster inference speed is obtained by compressing Net-S network with Tensor RT.And the compressed model is integrated into the autonomous driving software of the USV and tests in the lake.The results show that with the input resolution of 720 × 1280,the detection accuracy of the Net-ST is similar to Net-S,but the detection rate is greatly improved to 35 FPS.Therefore,the model proposed in this thesis meets the real-time requirements while ensuring high accuracy,thereby improving the ability to perceive the surrounding environment and the obstacle avoidance performance of USV.