Research On Ship Target Detection And Tracking In Complex Weather Based On Deep Learning

The ocean maintains maritime interests such as defending national sovereignty and territorial integrity,defending against attacks and invasions from hostile countries at sea,and supporting the sustainable development of Chinese economy in the process of the rise of the Chinese nation.Therefore,this paper researches the detection and tracking method based on deep learning,and at the same time,due to the interference of complex weather such as sea fog,the data is pre-processed for fog removal and then the subsequent detection and tracking.The paper also investigates the pre-processing of the data before the subsequent detection and tracking.Firstly,the self-built maritime ship target dataset was obtained from publicly available datasets on the web,sea trial images and web search keywords.In order to increase the generalization ability and robustness,the categories of fighter jets and buoys non-ship targets were added to obtain ten categories with a total of 3254 images,and the self-built dataset was labeled with the help of labellmg tool,and the obtained dataset was in VOC format.Secondly,there are many situations in complex weather,mainly for in-depth research on foggy conditions.In order to remove the interference of sea fog on the detection accuracy,firstly,the algorithms of multi-size Retinex,histogram equalization,and dark channel defogging are explained.The subjective evaluation and objective evaluation indicators of environmental defogging results verify that the AOD-Net defogging algorithm is more effective.Thirdly,an improved YOLOv4 ship target detection algorithm is proposed to address the problems of low detection accuracy and high miss rate of small targets in YOLOv4 for maritime target detection tasks.The method adds a 104×104 feature scale layer to improve the small target detection accuracy,clusters the labels using the K-means++algorithm,and embeds the SE(Sequence-and-Excitation)module on the basis of the YOLOv4 network to enhance the beneficial weights while suppressing the invalid ones,and finally uses the Focal loss to optimize the loss function to overcome the positive and negative sample inhomogeneity problem.The experimental results show that the improved YOLOv4 has a 2.12%improvement in mAP50(mean Average Precision)and a 2.20%improvement in mAP75,which is better for small,multi-target and overlapping target detection.Finally,based on the principle of deepsort tracking algorithm based on tracking-by-detection strategy,the Kalman filter algorithm is used to predict the tracking trajectory based on the completion of detection,and the Hungarian matching algorithm is used to cascade match and IoU match the tracking trajectory and detection results,and finally the estimated value is updated by Kalman filter.The experimental results show that the deepsort tracking algorithm can achieve effective tracking of small targets under sea fog,multiple targets under sea fog and overlapping multiple targets in three scenarios,as well as achieve tracking of multiple targets under sea fog and overlapping targets under certain degree of occlusion.