Research On Ship Type Detection And Tracking Technology Based On Deep Learning

Ensuring the safe and orderly navigation of ships in China’s well-developed waterways with a large number of vessels holds significant practical significance for promoting the healthy development of the shipping industry.With the rapid development of computer hardware and software technology,perception algorithms based on computer vision and deep learning can serve as auxiliary means for intelligent ship navigation and vessel monitoring.The "Action Plan for the Development of Intelligent Ships(2019-2021)" released by the Ministry of Transport highlights the importance of intelligent ships as a determining factor for the future development of the maritime industry.Perception algorithms based on computer vision and deep learning are the primary means for intelligent ships to perceive the navigational environment.Ship detection and tracking are fundamental for assisting maritime supervision and enabling intelligent ship perception of the navigational environment.However,challenges such as ship missed detection and misclassification still exist in ship detection,and the model parameters and computational complexity need to be reduced to enable the application of algorithms on mobile devices.In ship tracking,it is necessary to improve the accuracy of the algorithm while ensuring real-time performance.Therefore,this thesis aims to enhance the accuracy of target detection algorithms,lightweight the network,and improve the accuracy of multi-object tracking algorithms.Using deep learning techniques,ship type detection algorithms and ship tracking algorithms are studied to achieve real-time detection and tracking of ship targets.The main research contents are as follows:(1)Designing the Ship Type Detection algorithm STD-Yolov5.This thesis proposes the STD-Yolov5 model based on the Yolov5 object detection algorithm as the baseline network.It considers the accuracy,inference speed,and computational cost of the model and realizes the detection of ship types.Firstly,the backbone network of Yolov5 is optimized by embedding the ECA attention mechanism module into the C3 module to enhance the network’s feature extraction capability.Secondly,to reduce model redundancy,the model is lightweighted by integrating the Ghost Conv module into the Feature Pyramid Network to reduce the model’s parameters and floating-point operations.Next,to address issues such as misclassification and missed detection of ship types,a novel receptive field amplification module is constructed based on the Spatial Pyramid Pooling network by adding the Ghost Conv module and depth-wise separable convolution.Finally,to improve the accuracy of bounding box regression,the CIo U bounding box regression loss function is replaced by a simpler generalized ɑ-CIo U to ensure that the model is biased towards positive regression.Compared to the Yolov5 algorithm,STDYolov5 achieves a 1.2% increase in m AP@.5:.95,a 32.75% decrease in model parameters,a14.46% decrease in GFLOPs,and meets the real-time detection requirements.Compared to other mainstream object detection algorithms,STD-Yolov5 still exhibits superior performance.(2)Designing the Ship Tracking algorithm Ship Sort.This thesis proposes the Ship Sort model based on the Deep Sort multi-object tracking algorithm,aiming to improve the tracking accuracy of ship targets.Firstly,the OD-Res Net is designed as the feature extraction network for the tracking algorithm,which combines the Full-dimensional Dynamic Convolution with Res Net to extract deeper semantic information of ship targets.Secondly,inspired by the Interacting Multiple Modelsconcept,an IMM-based Kalman filter algorithm is designed to construct multiple motion models to describe different motion states of the same target.The weighted combination of these models enhances the accuracy of ship trajectory prediction.Ship Sort outperforms Deep Sort in terms of accuracy and also performs well in six evaluation metrics: MOTA,MOTP,IDP,IDR,IDF1,and FPS.This thesis proposes the STD-Yolov5 ship type detection algorithm and the Ship Sort ship tracking algorithm,which have been experimentally validated using the Sea Ships dataset and ship navigation video data.The experimental results show that the STD-Yolov5 algorithm outperforms other mainstream algorithms in terms of accuracy and model size.The Ship Sort algorithm ensures real-time performance while surpassing Deep Sort in terms of accuracy.