Research On Target Attitude Estimation And Manipulator Grab Based On Deep Learning

In recent years,with the deepening of deep learning technology,deep learning has shined in the field of various target detection events around the world,especially in the field of target detection and recognition.But in the actual industrial environment,it is not enough to detect and identify the target.With the advent of the era of artificial intelligence,robots are gradually being used in the industrial and service industries.In the factory environment,the robotic arm of the automated production line repeats the "grab-place" operation countless times.To accurately and quickly complete this series of operations,a simple detection and recognition algorithm is not enough,and you need to know the target Position and posture information.Therefore,the specific work contents of the research paper on the target attitude estimation research and the robot arm grabbing experiment are as follows:1?Introduce the background significance of 3D object recognition and deep learning research,at the same time elaborate on the current status of 2D target recognition and 3D target recognition at home and abroad,and explain the basic knowledge and theory related to the research of this article,including the theory of related deep learning algorithms,3D targets Detection algorithm,camera calibration principle,target attitude estimation principle;2?On the basis of studying the camera imaging model and the principle of target pose estimation,the Kinect camera used for the data collection collected in this paper and the camera used for the robot arm grab operation were calibrated respectively;3?The production method of LINEMOD data set is studied and improved.The data set mainly includes the eight corner points and centroid points of the smallest enclosing rectangle of the target,the mask of the target and the accurate 3D model of the target.First build a data set production platform,and use the Aruco module in Open CV to calibrate the two-dimensional code to obtain the pixel coordinates of the nine control points of the smallest bounding box of the target;then for the target with a regular shape,connect the eight vertex coordinates in pairs,and then pass The fill Poly function in Open CV is used to obtain the mask of the target;for irregular targets,the positioning frame of the target is first obtained by the YOLOv3 algorithm,and then the target segmentation algorithm Grab Cut is used to accurately segment the target mask;finally,the target minimum bounding model(MBM)Instead of the accurate 3D model to avoid the problem that the accurate 3D model of the target object is difficult to obtain;and completed a set of codes to automatically generate a data set,only need to input the length,width and height of the target to output the complete LINEMOD format data set.4?The attitude estimation algorithm using RGB as network input is studied.Based on the self-made data set in this paper,a target pose estimation algorithm based on the target minimum bounding model(MBM)is proposed.This method converts the target pose problem into two-dimensional coordinates that predict the nine points of the target MBM,and then combines the data The 3D MBM model of the target made in advance is concentrated,and the initial posture matrix of the target is calculated by the Pn P algorithm.This method achieves end-to-end target posture estimation,which can meet the "grab" operation that requires high real-time performance.And train and compare the analysis results on the public data set and the improved home-made data set,and simultaneously train the single target detection model and the multi-target detection model;the experimental results show that the method has obtained a comparison between the homemade data set and the official data set.Good prediction results.5?Simulate the construction of a robot gripping experimental platform in a real industrial environment and perform the calibration of the hand-eye of the robotic arm,and deploy the trained model to the robot for the gripping experiment.The Tsai-lenz algorithm was used for hand-eye calibration of the Co602 a robotic arm and the trajectory planning algorithm for the robotic arm was analyzed.Finally,the calculation results of the target pose were input into the robotic arm system to complete the grasping of the target object.