Research On Object Detection And Grasping Method For Domestic Service Robot

With the continuously increasing computer power and development of artificial intelligence technology,a round of research boom has been triggered in the field of service robots.Accurately completing the detection,grasping and avoiding obstacles are the prerequistes for the robot to intelligently complete the task.However,it is the key factors that the poor detection effect of small objects and the difficulty of grasping pose estimation that restrict the development of domestic service robots.In order to solve these problems,many scholars use SSD,YOLO and RCNN series of deep learning algorithms,template matching,image segmentation and point cloud segmentation algorithms to complete the robot's object detection and grasp in the filed of object detection and grasp.In this paper,objection detection and grasping are studied for the domestic service robot system.In the stage of object detection by robots,aiming at the shortcomings of existing object detection algorithms,a kind of SSD object detection algorithm based on feature fusion is proposed.The pose estimation based on point cloud normal is used to estimate the pose,which reduces its complexity.In addition,aiming at the difficulty of obstacle avoidance without prior information,this paper proposed an improved Artificial Potential Field for Virtual Coordinate(VC-IAPF)obstacle avoidance algorithm based on the characteristics of domestic service robot to realize autonomous obstacle avoidance.Finally,the overall design of the object recognition and grasping application system of the domestic service robot is completed.The main research contents of this subject are as follows:(1)SSD object detection algorithm based on feature fusionIn order to solve the problem that the detection effect of small objects is poor in the fileds of object detection,a feature fusion based SSD object detection algorithm is proposed in this paper.This paper analyses the characteristics of each layer of SSD network structure,identifies the reasons for the inadequate performance of small object detection,and determines the fusion feature layer and fusion structure.Firstly,the convoluted features are normalized and activated by ReLU.Then,the activated features are fused and fed into the SSD classification detector to complete object classification and detection.By fusing the shallow and deep feature information,the algorithm makes full use of the shallow high resolution features and deep high semantic features,and effectively improves the detection performance of small objects.(2)Object grasping posture method based on point cloud normalAiming at the difficulty of estimating the position and pose of the object grasped by the manipulator,this paper uses the point cloud segmentation method to obtain the complete target object,calculates the tangent normal vector of the center point of the target object,and then estimates the grasping position and pose of the manipulator.Aiming at the problem that there are many bad data in the depth map of Kinect camera,the bilateral filtering algorithm is improved by adding dynamic Gauss similarity weight factor to enhance the filtering performance,so as to repair the depth image.A more accurate point cloud image is established with color image and repaired depth image,and then the point cloud filtering,clustering segmentation,normal estimation and other methods are used to further get more accurate grasp points and their positions and postures.The experimental results show that the method has remarkable restoration effect,accurate position and pose estimation,meets the grasping requirements of the manipulator,and can accurately complete the grasping task.(3)Design and implementation of autonomous obstacle avoidance algorithm based on VC-IAPFAiming at the requirement of obstacle avoidance in domestic service robot system,an obstacle avoidance algorithm based on VC-IAPF is presented.Based on the center of the robot,the virtual coordinate system is established,the orientation of obstacles is located by distance priority strategy,and the coordinate data of obstacles are determined by using the virtual coordinate mapping relationship of obstacles.According to the characteristics of the obstacle avoidance process of the robot,a mathematical model of the obstacle rejection velocity is designed.The model is used to calculate the real-time repulsive velocity of obstacles to the wheelchair bed,and the velocity gravitational field and repulsive field is established.With the repulsion force unit,the obstacle avoidance velocity vector of the robot is calculated by the obstacle avoidance speed module,so that the autonomous obstacle avoidance of the robot can be realized.The performance of the proposed object detection algorithm is validated with the open PASCAL VOC dataset.The experimental results show that the mAP of the proposed SSD detection algotithm based on feature fusion is 78.33%,which greatly improves the detection effect of small target objects,and further verifies effectiveness of the algorithm.At the same time,the proposed object detection algorithm is applied on the platform of the service robot to verify the performance of the method of grasping object pose based on point cloud normal estimation.The experimental results show that the above algorithm can effectively accomplish the task of object detection and grasping.Finally,using wheelchair bed as the experimental platform,the obstacle avoidance effect of the VC-IAPF algorithm is verified.The experimental results fully prove the effectiveness and accuracy of the obstacle avoidance effect of the algorithm.