Study On Stereo Visual Real Time Localization And Target Detection Technology Of AUV

Ocean exploration has become a new battlefield of great powers.Underwater vehicle,as one of the important means to explore the ocean and maintain marine equipment,has become the research focus of relevant institutions.Nowadays,in order to complete the difficult underwater operations,underwater vehicles must have accurate underwater realtime positioning and underwater object recognition technology.Due to the high cost of sonar and other traditional sensors,and the limitation of distance in close underwater operation,vision-based underwater positioning method has been developed rapidly.In this thesis,real-time vision positioning and target detection of underwater vehicle are studied from three aspects: calibration of stereo camera under water,selection of point and line feature algorithm,realization of underwater binocular real-time positioning algorithm and application of underwater target recognition algorithm.Aiming at the traditional underwater positioning method,in this thesis a real-time underwater positioning method is proposed based on binocular vision,which is based on the characteristics of low texture of underwater scene.ORB-LSD is selected as the pointline feature extraction algorithm by designing a scoring mechanism.The re-projection error function based on the unification of point-line features,the solution of the motion posture between frames based on point-line features,the construction and optimization of local maps,Closed-loop Detection and closed-loop optimization are studied.Finally,experiments show that the binocular real-time location algorithm based on point-line features has better accuracy in low texture scenes,and has better real-time performance in underwater vehicles.For underwater target recognition and detection,we compare and analyze various object detection algorithms.For the sake of high accuracy and low computational complexity,we choose YOLOv2 algorithm as the object detection algorithm of underwater vehicle.The structure of feature extraction network Darknet-19 and some improvement measures are analyzed.The prediction principle of its boundary box and loss function are explored.A large number of underwater images were collected during the operation of underwater vehicles.By training our own data sets,we have realized the recognition of fish swarm,and achieved high accuracy and recall rate.Based on this,we achieve a simple underwater tracking function to detect targets.Aiming at the existing underwater vehicle in the laboratory,we have redesigned and developed the software control organization structure of the aircraft.The software control system written by the original Qt interface is encapsulated in the form of ROS to subscribe to the underlying data and update the system status.Each sensor data is transmitted to the Linux operating system through the underlying STM32,which is published in real time through ROS,so that the control system can update the system in real time by subscribing to the sensor messages.