Research On Fine-grained Object Detection Technology For Unmanned Surface Vehicles

China is a large maritime country,with a wide range of maritime strategic significance.Improving the level of marine equipment,strengthening the capabilities of the sea target detection,surveillance and sea area control is of great significance for the construction of a strong maritime state.At present,it is difficult to distinguish the maritime targets by means of satellite or radar detection,which can no longer meet the needs of modern combat and supervision.Unmanned surface vehicles(USVs)have gradually changed traditional combat methods with their advantages of flexible deployment and no casualties,and are playing an irreplaceable role in military reconnaissance,maritime supervision and other military and civilian fields.The rapid development and application of deep learning technology and fine-grained detection technology has created favourable conditions for the improvement of the level of autonomy and intelligence of USVs.However,in the water scene,the target has problems such as high similarity,large scale span,and many interference factors,which increase the difficulty of detection and fine-grained classification.Therefore,based on the application background of USVs,this paper focuses on the research of fine-grained target detection technology,and is dedicated to improving the environmental sensing capability of USVs on the basis of ensuring real-time algorithms.The following is the main work of this paper:(1)The difficulties and challenges of fine-grained detection of USVs are summarized,the framework and training strategy of typical target detection algorithms are analyzed,and a surface target simulation dataset,a surface target fine-grained classification dataset and a surface scene fine-grained detection dataset are established to provide data support for subsequent research.(2)A coordinate attention and a two-layer cascade modules are designed based on the need for fine-grained detection of multi-scale targets.An improved fine-grained target detection algorithm(MC-YOLOv5)based on YOLOv5 is proposed using these two modules.The feature parameters of targets at different scales are aggregated while enhancing the spatial dependence and location information of the targets.Experimental results show that the MC-YOLOv5 algorithm in this paper achieves an mAP of 80.9 on the self-built surface scene dataset while ensuring real-time performance,outperforming most target detection algorithms.(3)Aiming at the practical problems of detection algorithms with low fine-grained classification accuracy and easy overfitting during training,a fine-grained detection algorithm based on MC-YOLOv5 and multi-network self-supervised learning is proposed.During training,the methods of the mosaic data enhancement and freezing parameters are used.The experimental results show that the fine-grained detection algorithm improves the recognition accuracy of different subclass ships,reaching 83.3%,which effectively alleviates the overfitting phenomenon in the training process.(4)Through the construction of the USV environmental perception system and the embedding and transplantation of the algorithm in the USV terminal,the algorithm in this paper participates in the integrated sea trial in a certain sea area and completes the visual perception task of the USV.After testing,the accuracy of the fine-grained detection on the NVIDIA platform reaches 83.1%,and it only takes 0.06 seconds to process a single 640×640 image.There is no lagging of the optical video streams with a frame rate of 25,meeting the real-time requirements of most visual perception tasks.