Research On 6D Pose Estimation Network Based On Region-level Feature Fusion

In order to lead the high-quality development of intelligent manufacturing,modern industry needs more intelligent robots.Machine vision is the core module of intelligent robot.6D pose estimation is very important for robot visual perception technology.6D pose estimation is to obtain 3-DOF translation information and 3-DOF rotation information of the object in the camera coordinate system,which provides the information of the target object for subsequent application.The 6D pose estimation is the key to the application technologies such as industrial robot clamping object and augmented reality matching object.For example,in robot manipulation of objects,the identification of object's 6D pose provides useful information for grasping and motion planning.6D pose estimation is of great research value and is of great significance to realize robot intelligence.A key point of pose estimation based on RGB-D data source lies in how to deal with depth map data,which is greatly different from RGB image data.Depth map data stores the distance value of each pixel from the acquisition device.Generally,the depth is converted into point cloud data for processing.PointNet is the originator of point cloud network,but it cannot extract local information.To solve this problem,this paper proposes an improved point cloud feature extraction network based on PointNet,which can enhance the local information extraction.In severe sheltering and complex background,due to the block can only see some objects,most of the data source for RGB-D pose estimation method usually respectively to extract information from the color and depth data,when the extracted features cannot be ruled out effectively the interference information,this ser Io Usly limits the algorithm under severe sheltering and complex background prediction performance.To solve this problem,this paper proposes a pose estimation network based on region-level feature fusion,and uses our improved point cloud feature extraction network based on PointNet for geometric feature extraction of this network.The main contents of this paper are as follows:1.Aiming at the particularity of point cloud data,in order to serve the 6D pose estimation network,an improved point cloud feature extraction network based on PointNet is proposed to enhance the network's extraction of local information.Verification on var Io Us data sets shows that the improved point cloud feature extraction network based on PointNet achieves good performance.2.This paper proposes a pose estimation network based on region-level feature fusion in the case of heavy occlusion and complex background.The network combines the geometric features of the network extracted from the improved point cloud features of PointNet into the corresponding color images,and then uses the convolutional neural network to process the fusion features to obtain multiple region-level fusion features for subsequent pose estimation.The region fusion features contain both color and geometry information,which can enhance the robustness of the algorithm under severe occlusion.The results show that the proposed pose estimation network based on regionlevel feature fusion achieves good performance.It provides a new way to deal with two complementary data sets of depth map and color map.