Research On Robot Connection Parts 6DOF Pose Estimation Algorithm Based On Deep Learning

In recent years,facing the challenge of upgrading manufacturing industry,China has put forward a series of intelligent manufacturing strategies.machine vision is the key technology to realize intelligent manufacturing,and the robot grasping technology based on machine vision has always been the difficulty in this field.This paper studied the problem of 6-DOF pose estimation of parts in the process of robot grasping.Based on machine vision and deep learning,a 6-DOF pose estimation algorithm of robot connection parts based on deep learning is proposed.This method can estimate the real-time pose of parts through monocular vision system and meet the urgent needs in this field.In order to solve the practical problems of the difficulty of labeling the 3D datasets and the difficulty of real-time detection in actual engineering.In this paper,virtual reality technology and deep learning technology are fused.First of all,this paper expand a small amount of real data by generating virtual data and form a mixed data set.This way can reduce the workload of labeling.Then,an improved dope network structure is proposed to improve the real-time performance of monocular vision system for 6-DOF pose estimation of two types of robot connection parts without losing the accuracy.This algorithm greatly improve the accuracy and robustness of the system.The main work was as follows:(1)Aiming at the problems of low manual annotation efficiency of parts 3D pose estimation data sets,inaccurate semantic segmentation,and single background performance,in this paper,an improved strategy is proposed to generate a virtual data set based on the unreal engine for data expansion,and to mix the virtual data set and the real data set to form a mixed data set.We build 3D models of two types of robot connection parts,simulate the real lighting,randomly change the placement of objects,randomly change the background pattern of objects in the unreal engine 4,and generate virtual data sets.The experimental results show that the ADD pass rate of the mixed data sets using virtual data and real data is improved compared with the original data sets.(2)Aiming at the low efficiency of the original DOPE network in identifying parts with occlusion relationships,this paper proposes a random mask local processing method to generate a random size mask,which randomly occludes 0% to 80% of the object area,so as to improve the network's ability to occlude objects.Experimental results show that the ADD pass rate is increased.When the occlusion rate is 40%,60%,80%,the ADD pass rate is significantly increased.(3)Aiming at the problem of poor real-time performance of the original DOPE network,based on the DOPE network,a two-point improvement strategy is proposed in this paper.Firstly,use the depthwise separable convolution instead of the traditional convolution to lighten the original network structure and improve the network operation speed.The experimental results show that it is improved,Secondly,for the accuracy loss caused by parameter reduction,we introduce the attention mechanism module.The features extracted by the feature extraction module undergo channel attention and spatial attention to fuse the features of different sizes of receptive fields and improve the accuracy of the network.The experimental results show that the ADD pass rate is increased after the attention mechanism is introduced.(4)Aiming at the problems of misrecognition and missing recognition in the original DOPE network when recognizing parts with too large or too small scales,this paper proposes a multi-scale fusion pose estimation module.This module merges the feature maps of the three scales,replacing the original single-scale feature maps,and improves the network's ability to recognize parts of different scales.The experimental results show that the improved ADD pass rate is increased.In the end,we build a robot connection parts grasping system based on monocular vision,and conducts 200 repeated grasping experiments.The experimental results show that the recognition success rate is 94%,and the grasping success rate is 90.5%.It shows that the algorithm proposed in this paper can accurately and real-time estimate the 6-DOF pose of the two types of robot connectors.It is not sensitive to scale transformations and messy backgrounds,and has certain reference significance for practical engineering applications.