Grasping Position Estimation Of Unknown Objects Based On Deep Learning

Grasping is the main way of interaction between service robots and environment.The premise of its success is the accurate estimation of grasping position of target.Objects in unstructured environment have many categories,different shapes and are lack of threedimensional models,which put forward high requirements for the performance of grasping position detection algorithm.Aiming at the actual demand of service robot in unstructured environment,this thesis studies the detection algorithm of unknown object grasping position based on deep learning and builds the robot grasping system,which has strong theoretical and practical significance.In the research of target detection,the methods based on deep learning usually have higher accuracy.Therefore,this thesis solves the problem of grasping position detection for unknown objects based on deep learning.Based on the research of convolutional neural network structure,Faster R-CNN and R-FCN target detection algorithm,and the detailed analysis of the key problems in the detection of grabbing position,this thesis designs a grasping position detection algorithm for unknown objects based on proposal region.The experimental results show that the accuracy of the algorithm proposed in this thesis can reach 90.40%and the detection speed can reach 0.175 seconds per frame.In order to solve the problem that the detection algorithm based on proposal region is slow,this thesis designs an algorithm based on regression.Compared with the target detection methods based on proposal region,the regression methods such as YOLO and SSD have higher real-time performance,and the training of model is relatively simple,which are more suitable for deployment in the service robot platform.Therefore,on the basis of SSD algorithm,this thesis optimizes the detection of small targets and the feature extraction part of the network and proposes a grasping position detection algorithm for unknown object named SSGD(Single Shot Grasp Detector).The experimental results show that the detection accuracy of SSGD proposed in this thesis can reach 93.71%,and the detection speed can reach 0.051 seconds per frame.When SSGD algorithm detects grasping position,it quantifies grasping angle,resulting in loss of detection accuracy.Therefore,this thesis optimizes the detection of grasping position.Through the edge detection and fusion of RGB image and depth image,the accurate edge of grasping area of target is obtained.Then,the grasping contact boundary is extracted by line fitting.On this basis,the detection results of different grasping positions are optimized and adjusted based on the force-closure analysis method.At the same time,this thesis builds a robot grasping system and proposes an adaptive camera pose adjustment algorithm when there does not exist accurate grasping position of objects in some camera angles.The experimental results show that the algorithm proposed in this thesis has higher speed,applicability and accuracy,and the robot grasping system has a good performance in the real environment.Finally,the research work carried out in this thesis is summarized,and the future research content is prospected.