Research On Unknown Object Grasp Technology Based On Depth Learning

Chinese seriously aging population and the large number of disabled people make the Wheelchair Mounted Robotic Arms(WMRA)become the focus of current research.With the goal of improving WMRA's intelligence,ease of use and adaptability,this topic proposes and develops a laser point intuition interactive WMRA,and studies the key issue of autonomously grasping any unknown objec t.It mainly involves key points such as laser point recognition,point object separation,unknown object grasp frame detection,and grasping pose mapping from 2D grasp frame to 3D gripper in unstructured home environment.The details are as follows:Firstly,in order to solve the problem of tiny laser recognition in the real environment,we have carried out research on light spot recognition and point object extraction based on convolutional neural network.With the purpose of distinguishing between laser points and reflective points,the image is processed at the pixel level.Then,in order to extract the pointed object 3D information,research is carried out on image filtering,segmentation algorithm and object centroid calculation.By comparing the coordinates of the center of mass of the object with the converted coordinates of the laser point,the laser point is captured to take the object,and the fast separation of the selected object is realized.Secondly,in order to quickly and accurately calculate the grasping pose of the pointed object,an improved multi-modal algorithm based on Faster R-CNN is proposed.First,the topic uses multi-modal RGD as input,designs the Grasp Faster R-CNN model to identify pose of the object,and outputs the five-dimensional grabbing frame(x,y,w,h,?),solving the problem of low recognition accuracy caused by more grasping features.Then,in order to reduce the impact of the complex environment on the generation of the grasping representation,multi-modal information input of the object make the grasping more accurate.Finally,the five-dimensional grabbing frame is used as the learning feature to solve the problem of the lack of gesture information caused by the traditional grabbing point representation the problem of low success rate.Thirdly,In order to obtain the position and pose of the gripper,the topic studies the mapping relationship between the two-dimensional grasping rectangular and the three-finger gripper position and posture information.At First,the topic p erforms the parameter calibration of the Xtion PRO Live camera to obtain the internal and external parameters of the camera,then determines the grasp depth according to the minimum depth value of the point cloud neighborhood,and uses the calibration parameters to convert to obtain the grab coordinates.Finally,the grasping pose is obtained by calculating the average normal vector to accurately map the grasping five-dimensional representation into the grasping pose that the three-finger gripper can perform.The mapping method is versatile and solves the problem that the actual grasping in China is avoided and the actual grasping success rate is low.Finally,System integration and experimental research based on the robot operating system ROS to verify the feasibility of the algorithm.Integrate the laser spot detection algorithm with the grasp rectangular generation algorithm,so that it directly receives the posture information of the pointed object,and transmits the result to the robot arm to realize data communication between the grasp information and the robot arm.Finally,unknown object grasping experiments are conducted using typical objects in a home environment to verify the effectiveness of the algorithm in this topic.