Research On Grabbing Object Detection For Warehouse Automatio

Robotic grasping is one of the most basic tasks of industrial robots.Because grasping involves many factors,robot grasping is faced with huge challenges.In the traditional warehousing environment,the robot is in a structured environment and the grasping method is also manually taught to complete the fixed posture grasping of fixed objects,which cannot meet people’s requirements for intelligent grasping.Recently,the shapes,sizes,positions,types and placement directions of objects are becoming more and more diverse in the retail warehousing environment.It is necessary to develop a robot autonomous grasping technology suitable for unstructured scenes.However,the first thing to implement this technique is to find a suitable grasping pose.This paper studies the grasping of unknown objects in the storage environment.Combined with the current mainstream deep learning technology,two robot grasping detection methods based on two-dimensional planes are designed,which are verified by the Coppelia Sim simulation platform.The main contents are as follows:Firstly,the basic composition of the robot grasping system framework is introduced,the structure of the Kinect depth camera and the transformation relationship between the four coordinate systems in the robot grasping process are studied,and the camera calibration principle is studied.The image registration is carried out,and the parameter representation of the depth camera model in this paper is established.Secondly,given the problem that the object to be grasped and the grasping pose are separated in the current robot grasping detection,a cascade grasping detection method is proposed that combines the object detection network to be grasped and the grasping detection network.The grasping frame evaluation method obtains effective grasping frames and improves the accuracy of robot grasping detection.The method first uses feature fusion and attention mechanism to optimize the RFBNet network to realize the positioning of the object to be grasped by the warehouse robot.Then,the loss function is designed based on the Faster RCNN network to make it suitable for grasping,meanwhile,the classification balance of sample data and the processing method of feature information are optimized to realize the positioning of the grasping pose of the object to be grasped by the warehouse robot.Finally,the position information is used to judge the pose information,the grab box evaluation method is proposed to screen out more accurate grab boxes.Experiments show that the proposed method can obtain accurate grasping pose and improve the accuracy of grasping detection.Thirdly,given the problems of complex network model,a large number of parameters,and pool real-time based on the cascade grasping detection method,an end-to-end grasping detection method is proposed,which realizes the rapid grasping of robots.This method first introduces multi-scale receptive field blocks into the convolutional neural network to improve the network’s ability to distinguish features.Second,connecting 5 lightweight convolution blocks improve the network model’s ability to extract feature information.Finally,an attention mechanism is introduced into the network,so that the network can pay more attention to the grasped area.Experiments are carried out on Cornell grasping dataset,the accuracy reached 96.6% and 95.5%.Finally,a simulation experiment platform is built.Coppelia Sim is used to build a simulation experiment environment,and the robot grasping experiment is realized through python and Coppelia Sim,the effectiveness of the proposed method is verified by experiments.