Research On Object Tracking And Scene Perception Of Bionic Robotic Fish

As an underwater vehicle inspired by fish,the bionic robotic fish has attracted more and more attention due to its high efficiency,low noise and high maneuverability.In the research of the bionic robotic fish,better tracking of the bionic robotic fish and identifying the scene of the robotic fish can lay a solid control foundation for the follow-up research of the bionic robotic fish.This thesis takes the bionic robotic fish as the research object,and carries out target tracking and scene perception based on global vision.The main research contents are as follows:Firstly,the thesis described the research background and significance of the bionic robotic fish target tracking and underwater scene perception.It reviewed domestic and foreign research status of bionic robotic fish,target tracking and scene perception.And the research contents of each chapter are briefly described.Secondly,the target detection data set and scene perception data set of the bionic robotic fish are constructed.First,a robotic fish data acquisition experimental platform composed of bionic robotic fish,high-definition cameras,control computers,and pools was built,and 1000 robotic fish motion images were collected.Secondly,in order to improve the generalization ability of the depth model,four data enhancement operations,image rotation,Gaussian blur,brightness transformation,and mirror flip,are used to expand the data set.Finally,the data were annotated,and the target detection data set of bionic robotic fish was constructed successively.The scene perception data set including fish school,giant fish,coral reef,reef and marine vegetation was collected and sorted out,which laid the foundation for the follow-up research.Thirdly,a bionic robotic fish detection method based on YOLOv3-DeepSort is proposed.This method modifies the YOLO output layer,improves the K-means clustering method to determine the anchor frame,establishes a YOLO target detection algorithm suitable for bionic robotic fish,and optimizes the training of the detection model for different hyperparameters;The video jam problem has been optimized and improved by multi-threading in practical applications.On this basis,the DeepSort algorithm is used to realize the target tracking of the bionic robotic fish.Then,a comparison experiment of target tracking of bionic robotic fish was carried out.The experimental results show that the target tracking algorithm based on YOLOv3-DeepSort has better tracking performance.Finally,a scene perception method of bionic robot fish based on MobileNetV3 is proposed.This method uses transfer learning strategy to train the model,and optimizes the model for super parameters.In order to verify the effectiveness of the scene perception method of biomimetic robot fish based on MobileNetV3,the scene perception and tracking control software platform of biomimetic robot fish is developed,and the scene perception algorithm of robot fish based on SIFT-SVM is designed.The scene perception model based on MobileNetV3 is compared with that based on SIFT-SVM.The experimental results show that the scene classification method based on Mobilenet V3 is better.