Research On Binocular Vision Based Grasping Detection Method Of Deep-sea Manipulator

Target recognition and positioning are the key technologies for the autonomous and accurate operation of deep-sea manipulator.It is of great significance and value to carry out research on this.Based on the task of "Intelligent accurate Operation module Development" of the strategic pilot science and technology project of the Chinese Academy of Sciences,this paper takes the deep-sea 4500 seven-function manipulator as the research object,studies the binocular vision recognition and positioning method,realizes the grasp detection of the manipulator based on convolutional neural network,writes the intelligent accurate operation module control software,and carries out the test and testing work.The autonomous operation function of the manipulator is realized.The main work of this paper is as follows:1.A binocular stereo camera based vision recognition and positioning system for deep-sea robotic arm was constructed,the binocular hand-eye imaging model was established,the internal and external parameters of the vision system were calibrated,and the parallax depth map was obtained by combining with the semi-global stereo matching algorithm.The depth information of the captured target was calculated and the positioning function of the robot arm was realized.The experimental results show that the positioning error of the constructed hand-eye system is less than 5 mm,which can meet the real-time grasping operation requirements of the manipulator.2.This paper proposes a robotic arm grasp detection method based on lightweight feature fusion convolutional neural network.In this method,three-channel RGB images and single-channel depth map are taken as the input of learning network to construct one-to-one mapping relationship between objects and pixels,effectively avoiding time-consuming operations such as discrete sampling and candidate sorting.Multi-constraint prediction evaluation method is adopted.It improves the accuracy of target positioning in the grasping process.The experimental results show that this method can accurately predict the grasping pose in unstructured environment,the grasping success rate is up to 86.7%,and the processing speed of a single image is about 0.12 s,achieving the balance between accuracy and speed.3.Developed the deep-sea robotic arm system control software,and realized multiple functions such as target positioning,grasp detection,motion planning and state monitoring on the multi-threading architecture,which not only ensured the independence and stability of each algorithm,but also ensured the readability and robustness of the human-computer interaction interface,and also provided a simple human-computer interaction interface for the manipulator operators.The control function of the manipulator is realized by the upper computer.Figure[77] Table[9] Reference[61]...