Detection Of Fish Feeding Behavior Based On Unmanned Ship System Study

In order to achieve autonomy in fish farming,reduce labor costs,and improve the intelligence of aquaculture,a fish feeding behavior detection system based on autonomous cruise unmanned ships was experimentally designed,and the YOLOv4 algorithm was improved and compressed into small embedded devices.In the Raspberry Pi,the shape,texture,and density features of the fish swarm are extracted and fused.Compared with obtaining the features of a single fish body,the overall feature recognition accuracy of the fish swarm was higher,and the autonomous cruise of the unmanned ship can realize fixed-point real-time feeding.Detection,improvement of the autonomous feeding device for unmanned boats,had high research value for liberating labor,reducing labor costs and smart fishery feeding.The main research work of this paper was as follows:First of all,this paper selected the unmanned boat as the carrier,designed an autonomous cruise system,carried the Raspberry Pi to the designated feeding point to detect the feeding status of the fish,and designed the host computer system to monitor the status of the unmanned boat in real time,and established a system for the detection of the feeding behavior of the fish.test platform.The autonomous cruise system mainly included functions such as driving,positioning,speed,heading,obstacle avoidance,and real-time communication.Among them,the speed control adopted PID algorithm and PWM control to realize that the unmanned ship can reach the designated location at the specified speed.Heading control used GPS and IMU to collect positioning information to achieve accurate real-time positioning of unmanned ships.The lidar and steering gear form an obstacle avoidance system,and the differential steering movement of the unmanned ship was controlled by the lower computer.When the unmanned ship cruised autonomously,the Madgwick gradient descent algorithm was used to filter the data collected by the sensor,and Google Earth was used to send the position to the upper computer to realize automatic cruise.Secondly,this paper studied the YOLOv4 algorithm based on one-satge target detection.In order to realize the requirement of carrying the detection model to small embedded devices,the model was optimized,and the F-YOLO model was designed to realize the detection of fish feeding behavior.The feature extraction network of the FYOLO model used Mobile Net V3 instead of the backbone feature extraction network CSPDarknet53 of YOLOv4,which improved the real-time detection performance of the network when the detection accuracy was less sacrificed.In order to solve the problem of occlusion in the multi-target detection of fish schools,the algorithm model improved the accuracy of anchor recognition,and replaced the K-means algorithm that comes with YOLOv4 into binary K-means,and added global non-maximum suppression to the DIo U loss function.Determine the anchor frame to improve the detection performance of small fish targets.In order to further improve the detection speed,the channel pruning compression model was carried out on the convolutional layer of the network structure,which reduced the number of floating-point operations and the amount of computation of the model,and increased the knowledge distillation process,which made up for the decrease in recognition accuracy caused by pruning.Finally,in the detection of fish feeding behavior,this paper proposed a real-time network detection model of light-weight fish feeding behavior based on the fusion of fish texture,shape and density features for traditional detection algorithms.The shape feature of the picture was composed of the outer contour lines of the fish school.The shape image was obtained by the canny operator in the experiment,and the shape parameter of the fish school was used as the basis for judging the feeding state.The texture feature of the fish was the background image of the fish.In this paper,the background difference method was used to extract the background,and the frequency domain feature of the background image was extracted by Fourier transform.The fish density index was the number of fish within a certain range.In this paper,the counting range was set according to the placement of the bait.According to the pictures obtained in the experiment,each fish will generate a Bounding Box.The F-YOLO model calculated the Bounding Box within the set range.The number of fish determined the number of fish,thereby obtaining the fish density index.After a large number of experiments,the threshold value of 15 was determined.If it exceeded 15,it was judged as feeding,and if it was less than 15,feeding was completed.The experimental results show that the improved F-YOLO detection model was compressed from 245 MB to 49.7MB,the average recognition time per image reached50 ms,the accuracy reached 99.13%,the FLOPs was only 16.4G,and the FPS in embedded devices can reach 52 frames.The fish feeding behavior detection studied in this paper can be applied to the automatic feeding system of unmanned boats.Using machine vision technology to control feeding can make feeding decisions accurately and quickly,which was more effective than manual experience feeding.Reducing the cost of aquaculture,improving the precision of aquaculture and the intelligence of aquaculture can provide theoretical guidance for actual fish farming.