Research On Intelligent Removal Of Dead Fish Based On Deep Learning

The domestic industrialized circulating aquaculture model has been vigorously promoted and developed,in which water treatment technology and equipment are quite mature,but the supporting production auxiliary equipment such as intelligent feeding,grading and catching are undeveloped,especially in removal technology and equipment of floating dead fish in the factory circulating water fish pond.The closed aquaculture systems,due to various factors such as water quality,temperature,dissolved oxygen and disease,the dead fish will cause more infections.Currently,the disposal of dead fish is mainly achieved by manual labor.However,because of the increase of labor costs,high labor intensity,and shortage of human resources,fewer and fewer people are engaged in such low-level monotonous work.Therefore,this research object of this paper,the intelligent removal of floating dead fish in industrial circulating water fish ponds,can solve the problem of manual and cumbersome removal of dead fish.By focusing on the requirements of intelligent fishing of robotic arms,this paper investigates the detection and tracking of dead fish objects,pixel position conversion of dead fish and fishing methods based on robot operating system for robotic arms in a factory circulating water environment.The specific content of the paper containing the following perspectives:(1)Conducting object detection studies on dead fish.Firstly,we collect 8970 pictures of dead fish under different working conditions in the factory circulating water environment,which are annotated with the help of Ro Label Img software,then we introduce the network structure of YOLO v5,a current advanced object detection algorithm.We also analyze the object feature of dead fish with an aspect ratio by using a suitable rotating object detection algorithm and evaluated the performance of rotating object detection and found that the average accuracy of the mean average precision reaches 90.44%.Finally,we make a comparison between a 3D object detection developed by depth camera and 2D image mapped by a common camera and determined the camera selection for the removal test.(2)Target tracking and coordinating system transformation for dead fish.Firstly,we expound the DeepSORT algorithm framework for target tracking by taking rotating target tracking as input and innovated the extraction method by using rotating IOU matching and taking the rotating object as the appearance feature.Then,we evaluate models of improved and unimproved multi-target rotation tracking.We find that the multi-target rotation tracking accuracy is 79.6%,which is 22.6% higher than the YOLO v5-DeepSORT algorithm and the multi-target rotation tracking precision is 68.4%,increased by 4.9%.we also analyze the path of the floating dead fish in the factory circulating water fish pond by using multi-target rotation tracking and find that the dead fish basically rotate around the center of the fish pond,which establishes the polar coordinate system in the dead fish removal method.Secondly,we establish the mathematical model based on the principle of small hole imaging and obtain the internal parameters of the camera and use the hand eye calibration method to calculated the position and attitude conversion from the camera coordinate system to the basic coordinate system of the manipulator.Finally,we achieve the transformation of the dead fish coordinates and obtained the salvage coordinates by the discrete polar coordinate method.(3)Design and removal test based on ROS manipulator.First of all,according to the need to remove the floating dead fish in the factory circulating water fish pond,we select the robotic arm,design the end effector and its subsidiary mechanisms.Besides,we configure the control of the manipulator based on ROS,realize the static obstacle avoidance,common position settings and its motion control in joint space and Cartesian space.Then,we integrate the multi-target rotation tracking and coordinate system transformation into ROS to realize the real-time communication between the manipulator and the algorithm.Secondly,we introduce the components of the removal test system.It is divided into two cases to conduct real-time intelligent removal test of dead fish in aquaculture ponds: one is the case with normal water intake and aeration,and the other is with aeration air but no water.The test shows that the one-time successful removal rate of dead fish without water inflow is 60%,while the one-time successful removal rate of dead fish under normal conditions is 77.14%.Finally,we analyze the influencing factors of fishing test from four aspects: the accuracy of target tracking method,the influence of camera position on coordinate conversion,the reasons affecting the first-time successful removal rate and the direction of improvement.The research results have reference value for the intelligent removal of floating dead fish in the industrialized circulating water fish pond,and provide a basic research direction for the intellectualization and unmanned of industrialized circulating water.