Research On Grasping Posture Detection Technology Based On Deep Learnin

In recent years,the rapid advancement of science and technology has driven significant progress in robot technology.This has resulted in the widespread application of robots in various aspects of daily life,with robot grasping emerging as a fundamental task within the field of intelligent robotics.However,the increasing diversification in the types,shapes,and sizes of objects in retail environments has rendered traditional robot structured grasping methods inadequate.In light of these challenges,this paper focuses on the crawling problem in retail environments and proposes crawling strategies based on deep learning technology for both regular and irregular objects.Specifically,the two methods proposed in this study have been rigorously validated via both simulation platforms and practical applications.Firstly,the present study developed a comprehensive framework for the robot grasping system.This framework was divided into three essential components: environmental perception,grasping detection,and grasping execution.Initially,the study analyzed the calibration and image registration methods of Kinect V2 depth camera in order to facilitate environmental perception.Additionally,the study investigated the hand eye calibration and coordinate conversion relationship of robots,aiming to establish a theoretical basis for robot performance in grasping tasks.Secondly,a grasping strategy based on regular objects is proposed,and the grasping and detection tasks of regular objects are divided into two parts: target positioning and twoclassification angle grasping,so as to improve the grasping efficiency of regular objects.This paper focuses on E-YOLOv4-Lite,a lightweight object detection method for retail goods.This method replaces the backbone network with Mobile Net V3,designs a lightweight E-CBAM attention mechanism,optimizes the loss function,and significantly improves the detection ability of occluded objects and small target objects.Thirdly,the study proposed a novel grasping strategy for effective handling of irregular objects.An end-to-end detection method based on grasping quality regression was developed,incorporating Focus module for lossless down-sampling,a multi-scale fusion R-Resblock,a lightweight RFB-SE model,and optimized loss function to enhance accuracy.Comparing the proposed method with existing grab detection techniques on the Cornell and Jacquard datasets,yielded high levels of accuracy,with a score of 97.8 and 91.5 respectively.These findings demonstrate the suitability of the method for facilitating stable and efficient grasping of irregular objects.Finally,a retail sorting simulation platform,utilizing V-REP,was constructed to enable the verification of two grasping methods in a simulated environment.The regular object grasping detection method was applied to the BR280 mobile robot for automatic grasping experiments,demonstrating the effectiveness of the grasping strategy.Additionally,the irregular object grasping detection method was applied to the Baxter robot for grasping experiments,demonstrating the effectiveness of the grasping pose estimation model.The experiments validated the robustness and suitability of both proposed methods for crawling detection and their appropriate application in retail environments.This advancement in the field of robotics has great potential for streamlining retail operations and increasing efficiency.