Research On Underwater Target Detection And Tracking Algorithm Based On Deep Learning

With the continuous progress of underwater robot technology,underwater target detection and tracking has become a widely concerned research field.Traditional underwater target detection and tracking methods mainly rely on hand-designed features and classifiers,but this method has many problems,such as inaccurate feature extraction and poor generalization ability of classifiers.In recent years,the performance of underwater target detection and tracking has been improved,which is attributed to the continuous development of deep learning technology.This thesis also studies the detection and tracking of underwater fish targets based on deep learning technology.The main research contents are as following:(1)Aiming at the problems of underwater image blur,color distortion,complex underwater scene and limited target feature extraction ability,an improved YOLOv5-FISH method for underwater fish target detection is proposed.Firstly,the fish detection dataset Underwater Fishes is constructed based on Labeled Fishes in the Wild,which lays a foundation for model optimization and algorithm research.Secondly,by using the UDCP(Underwater Dark Channel Prior)algorithm to preprocess the underwater image,the image is enhanced to make the image clear,which is helpful to correctly identify the target in different environments,thus improving the detection accuracy.Then,the ECA(Efficient Channel Attention)attention mechanism is introduced into the Backbone layer of YOLOv5,which increases the breadth and depth of YOLOv5 network,enhances the feature acquisition ability of small targets,and further improves the detection ability of the algorithm.Finally,the loss function is introduced to enhance the robustness of target detection and improve the accuracy and recall rate of target detection.Compared with the previous algorithm,the algorithm proposed in this thesis has significantly improved the recognition accuracy of underwater fish targets,thus effectively improving the accuracy of underwater fish target detection.Experiments on the Underwater Fishes dataset show that the YOLOv5-FISH model improves the accuracy of the original YOLOv5 from 96.23 % to 99.18 %,an increase of 2.95 %.The m AP＿0.5:0.95 increased from 70.02 % to 77.69 %,an increase of 7.67 %.According to the experimental results,the proposed method can effectively improve the target detection accuracy of underwater fish.(2)Aiming at the problems of poor underwater target tracking effect and low tracking accuracy,an underwater multi-target tracking method Underwater-Fish Track based on improved Deep SORT is proposed.Firstly,the YOLOv5-FISH underwater fish target detector proposed in this paper is used to replace the original target detector Faster-RCNN,so as to improve the accuracy of underwater fish tracking.Then,the full-scale deep feature learning model OSNet(Omni-Scale Network)network is introduced to extract the appearance features,so as to improve the matching degree of appearance features between the detection box and the prediction tracking box.Finally,the public data set Brackish MOT is used for experiments,and compared with other existing tracking algorithms SORT and Deep SORT.The experimental results show that the proposed algorithm performs better in both MOTA and MOTP indicators,which are improved by 8.97 % and 7.1 %,respectively,compared with the original Deep SORT algorithm.Compared with the SORT algorithm,the improvement of the proposed algorithm is more significant,which is increased by 18.23 % and 9.44 %respectively.