Research On Maritime Target Detection And Tracking Technology Based On Deep Learning

With the development of marine resources,the problems of maritime transportation safety,marine environment safety and maritime homeland security are becoming increasingly serious,and the demand for high-precision maritime ship multi-target real-time detection and tracking technology is increasing in the civil and military fields.At present,there are few models dedicated to ship target detection and tracking,and they are easily affected by the change of ship target scale,inter-target occlusion,wave wake and complex weather,and the existing models have the problems of complex structure of detection algorithm,poor real-time performance and low tracking accuracy in the task of ship multi-target detection and tracking.To address the above problems,this paper proposes MA-YOLOv4 maritime ship target detection algorithm and combines it with the improved DeepSORT maritime ship target tracking algorithm to achieve accurate detection and tracking of ship targets in order to improve the intelligence of maritime supervision.The main research contents are as follows:(1)Aiming at the problems of complex structure and poor real-time performance of YOLOv4 algorithm,a new lightweight ship target detection network MA-YOLOv4 is proposed.Firstly,in order to improve the problem of random initialization of clustering centers by K-means clustering algorithm,the stability of K-means algorithm is enhanced by selecting 9 sample points with the largest spacing as the initial clustering centers to obtain the anchor frame suitable for the shape of the ship target;secondly,in order to simplify the model structure and reduce the number of parameters,lightweight S-MobileNetv3 is used as the backbone of the detection network,and depth separable convolution is used instead of normal convolution to improve the detection speed.Secondly,to simplify the model structure and reduce the number of parameters,the lightweight S-MobileNetv3 is used as the detection network backbone,and the depth separable convolution is used instead of the normal convolution to reduce the number of model parameters and improve the detection speed of the network.The experimental results show that compared with the YOLOv4 algorithm,the mAP of the algorithm in this paper is improved by 4.69%,the number of model parameters and the computational effort are reduced by 80.87%and 87.66%,and the detection speed is improved to 52 frames per second,which achieves accurate and efficient multi-target detection of ships.(2)Aiming at the problem of poor robustness of multi-target tracking model for offshore ships,an improved DeepSORT algorithm based on appearance feature enhancement combined with MA-YOLOv4 is proposed.Firstly,the MA-YOLOv4 proposed in this paper is used as a detector of the tracking algorithm to obtain accurate maritime ship target positions and features.Secondly,the accurate appearance features of the ship target are extracted adaptively using the deformable convolution aligned with the shape of the ship target,and fused with the target features extracted by the re-identification network to enhance the network’s ability to judge the same ship target and improve the ID switching problem.The experimental results show that compared with the DeepSORT algorithm combined with YOLOv4,the improved DeepSORT algorithm proposed in this paper improves MOTA and MOTP by 4.4%and 3.8%,reduces IDs by 19.66%,and achieves accurate multi-target tracking of ships at sea with good results for small-sized ships and obscured ship targets.