DNN-based Service Robot Capture Pose Estimation

Service robots as a new form of robot automation outside of industry,which often used to help or replace humans in performing trivial,time-consuming,repetitive,or dangerous tasks.The autonomous grasping in the service robot technology determines the dexterity of the robot to perform many tasks,which is of great significance.Robotic grasping technology has been widely used in industries production with a single target and relatively stable environment.However,the autonomous grasping ability of service robot is poor.At present,the objects in the family environment are diverse,and the detection and recognition of the object are easily affected by factors such as ambient light and object occlusion,which make the pose calculation and the robotic arm grabbing planning of the object more difficult.Therefore,this thesis is aimed at the problem of self-acquisition of the service robots,and deeply studies the methods of robot capture object and scene description,position estimation by object grab and the motion control by robot arm grab etc.This implements a comprehensive crawling method for a general environment and arbitrary objects.Meanwhile,the approach in this paper can be verified by robotic grasping physical experiments.This study would include the following parts: 1.A service robot framework is described based on stereo vision and a light-weight bionic robotic arm.The framework consists of a stereo camera based on binocular and structured light,a humanoid light-weight robotic arm and a computer.Meanwhile,the framework uses the hand-eye system of “eye” on the “hand” to make the grasping more flexible.In addition,the robot can observe the deviation between the object and the hand during the movement to adjust the execution plan of the grasping action by the robot arm.2.Established a grasping model of the service robot,defined the model parameters,and designed a deep neural network model for the quality assessment.Then,the network model is analyzed to get the basic structure of the network.3.A training dataset is generated from Cornell's dataset based on model parameters,which covers common grasping objects in the home environment.Meanwhile,a specific grasping quality evaluation depth neural network is designed,which takes the depth image as input and predicts the grasping quality and the grasping gesture in units of pixels.In addition,the accuracy of the captured pose detection on the existing data set can reach 91.5%.4.This thesis builds an experimental environment based on the open source ROS robot development system,and combines tensorflow and Keras to design a grasping quality assessment based on depth neural network.The neural network model training is completed on the filtered dataset.At the same time,the objects repeated grasping experiment are conducted in a real home environment,the detection success rate was 91.2%,and the final objects grasping success rate are 86.7%.This thesis designs a comprehensive grasping method for real-time and arbitrary objects that can be used for closed-loop control,which realizes the three-dimensional description and capture of the grasp object in the unstructured space.