Study On Texture-less Industrial Parts Pose Estimation In Unstructured Environment

With the advancement of industrial automation,the intelligent transformation and upgrading of traditional industries is the trend of the times.Robots are an essential tool in industrial production and it is necessary to endue robots with strong visual perception capabilities in a complex and unstructured environment.This paper focuses on the pose estimation of texture-less industrial parts in unstructured environments and the application of robot bin picking and proposes a deep learning method for the pose estimation of texture-less objects based on grayscale and depth images.The specific research content consists of the following four aspects:(1)A rapid dataset generator based on physical simulation is proposed to alleviate the problem of high cost and time-consuming of generating datasets in the field of deep learning.According to the pre-configured parameters,this method can quickly generate large-scale annotated datasets by importing the three-dimensional model of the objects.At the same time,this article introduces the strategy of domain randomization in the simulation,which improves the generalization ability of the network model trained by using simulation datasets.(2)This paper proposes an instance segmentation method for textureless objects based on the simulation datasets,which can predict the category,bounding box and instance mask of texture-less industrial parts in the presence of clutter and occlusion.This method introduces the binary classification constraint and attention module based on the one-stage instance segmentation network YOLACT,which makes the foreground part of the image occupy more attention,and improves the instance segmentation performance in the presence of clutter and occlusion and texture-less industrial parts.Experiments show that the proposed method can achieve effective recognition results on both simulated datasets and real datasets.(3)Based on the instance information of each industrial part obtained by instance segmentation network,a two-stage pose estimation method is proposed.This method adopts the "first coarse and then fine" pose estimation strategy.The initial pose estimation network takes the depth image and the grayscale as input,and predicts the initial pose of the industrial part through feature extraction and feature fusion.The pose iterative refinement network takes the model point cloud and the scene point cloud transformed by the initial pose as input,and predicts the residual pose by fusing global and local features.The residual pose and the initial pose are integrated to obtain better pose.Experiments show that the proposed pose estimation method can effectively identify the pose of the industrial parts on both the simulation datasets and the real datasets in the presence of clutter and occlusion.(4)In order to verify the utility of the pose estimation proposed,this paper builds an experimental platform for robot bin picking in a laboratory environment,and designs a bin picking software framework based on ROS.This paper adopts the idea of modular design,and separates the data acquisition module,the instance segmentation algorithm module,the pose estimation algorithm module and the robot control module into ROS nodes,which is convenient for debugging and maintenance.Experiments show that the pose estimation algorithm for texture-less industrial parts in unstructured environments proposed in this paper can grab and sort multiple types of industrial parts in the presence of clutter and occlusion,and also verifies the effectiveness and practicality of each part of the algorithm in real scenes.