Multi-target Detection And Tracking At Sea Based On Deep Learning

China has a vast sea area,numerous rivers and developed shipping.The theoretical research and practical application of marine target detection and tracking are of great significance in the areas of surrounding sea area monitoring,inland river port ship information management,and environmental perception decision-making.At present,marine targets are mainly detected through synthetic aperture radar imaging,remote sensing imaging,infrared imaging,and visible light imaging pairs.Compared with other imaging methods,visible light imaging contains rich color,texture and spectral information,which can realize efficient detection and tracking of multi-target ships in a small area.Its low price is suitable for large-scale promotion and application,and has become a current research hotspot.This article takes marine targets as the research object,and focuses on the detection and tracking methods of multi-ship targets on the sea based on visible light images.The specific contents are as follows:First,this paper collects real sea target image data for training.And the data set is divided into 10 categories such as warships,speed boats,cargo ships,passenger ships,sailing ships,other ships,large fish,birds and reefs.Label the data set with label Img software.Aiming at the problem that the accuracy of marine target recognition is easily affected by complex environment,this paper uses image mixing technology to mix the image of complex environment and the image of marine target in proportion to enhance the robustness of the marine target detector.Secondly,in view of the low accuracy of traditional ship detection algorithms,this paper introduces the deep learning YOLOv3 algorithm into the field of ship and sea target detection.At the same time,in order to solve the problem of low recognition of YOLOv3 small and weak targets at sea,an improved YOLOv3 based sea target detection method is proposed.Increasing the number of small target samples at sea during training makes small target detection better.the introduction of Octave convolution CFE module makes the model feature extraction stronger and faster.the use of GIOU to improve the regression loss function makes the network border regression more accurate.The test results show that the proposed improved YOLOv3 algorithm has a greater improvement in target detection accuracy and better detection effect on small targets.Thirdly,the ship’s tracking process mainly has the problems of ship deformation and reidentification caused by occlusion and intersection.This paper uses the improved YOLOv3 detector to match the kernel-related filter tracker every 10 frames to solve the problem of ship motion deformation.Improve the original data association by introducing visual features based on neural network,and solve the loss of tracking target caused by ship occlusion and cross motion.Based on the improved frame-by-frame detection of the YOLOv3 detector,this paper achieves ship-sea multi-target tracking by matching the tracking target with the detection target,and this method is used as an experimental control.Finally,the ship motion video is made into a data set to verify the effectiveness of the proposed algorithm.The test results show that the multi-target tracking algorithm based on KCF and improved data association achieves a good balance between real-time and tracking accuracy.The algorithm can track ships at sea stably,and the tracking target can be re-tracked after a period of loss due to occlusion.