Research On Fish Detection And Tracking Method Based On Deep Learning

Fish detection and tracking technology is of great significance in the field of aquaculture as the premise and basis for solving practical problems such as fish count measurement and fish behavior analysis in marine pastures.In recent years,object detection and multi-object tracking algorithms which are based on deep learning technology have developed rapidly.However,the applications in the field of fish are not widespread.It is a key research to apply object detection and multi-object tracking algorithms to the field of fish tracking and to improve the performance of fish detection and tracking.Fish detection and tracking methods are the technical foundation of fish counting and behavioral analysis to create greater benefits in aquaculture.This paper achieves detection and multi-object tracking of fish based on deep learning technology,and finally deploys the model of fish detection and tracking into the system for application in fish farming.The main work of this paper is as follows:（1）A method,ROFD（Real-time occluded fish detection）,is proposed to detect fish in the case of occlusion.Aiming at the high rate of missed detection and false detection in the case of serious fish occlusion in fish detection,this chapter proposes a real-time occluded fish detection method.Occlusion can be divided into intra-class occlusion and inter-class occlusion.First of all,for the intra-class occlusion that occurs when a fish is occluded by other fishes of the same category,the repulsion loss is introduced in the loss function in order to penalize the predicted box for shifting to other ground-truth objects and then achieve high detecting accuracy.In addition,aiming at inter-class occlusion which occurs when fish is occluded by objects of other categories,a Cross-Stage-Partial-Connections module is added in the head of model to fuse more features of different layers and then further extract features for better prediction.In this way,the accuracy of fish detection is improved.Lastly,this chapter constructs the COFD（Complex overwater fish detection）dataset.The images of the COFD dataset are obtained by camera which is put above the fish pond in the laboratory.The images data are labeled using software before data preprocessing.By experimental results on the COFD dataset,the proposed method achieves 92.0%AP50,60.0%AP75 and 55.2%AP that effectively improves the detecting accuracy of fish under occlusion.（2）A multi-object tracking method of fish based on attention mechanism,Fish Track,is proposed.In this chapter,the attention mechanism-based algorithm,Trans Track,is transfered to the fish domain for achieving fish tracking.However,the loss function used in Trans Track exist the slow convergence speed and the low convergence accuracy,and the backbone network used keep a weak ability to extract features,which cannot effectively improve the accuracy of fish tracking.For the issue of low accuracy when the Trans Track algorithm is transfer to the fish tracking task,this paper improves the backbone network and loss function of the Trans Track algorithm to improve the accuracy of fish tracking.In this chapter,Res Ne Xt101 is utilized to improve the backbone of Trans Track,and thereby the number of layers of the backbone of the model is deepened,and the ability of the model to extract features is improved,thereby improving the accuracy of multi-object tracking of fish.In addition,this chapter uses the Smooth L₁ loss function and the DIo U loss function to improve the loss function used in the Trans Track algorithm,which improves the speed of convergence for network and the accuracy of fish tracking.Last,the experimental results show that the proposed methods achieve 79.3%MOTA and 80.3%IDF1 respectively,which effectively improve the accuracy of fish tracking.（3）A fish detection and tracking system is research and development.The fish detection model and the fish tracking model are deployed into the system.The code is programmed by the Python language,and the system is designed by the Pytorch framework and the Flask framework.In the system,the pictures or the videos uploaded by the user are saved in local,and the algorithm and the saved model are called to detect the fish in the pictures or videos.The system can be applied to improve the degree of automation in the aquaculture.