A Pose Estimation Approach Based on Keypoints Detection for Robotic Bin-picking Application

Robotic bin-picking is a fundamental yet trouble-some task in robot autonomous manufacturing applications such as industrial parts feeding, assembling, and sorting. The parts might be randomly arranged or stacked on the clutter scene with heavy occlusion. To reliably pick each part, accurate object detection and pose estimation are the key technology. In this paper, we present a real-time machine vision-based robotic binpicking system, and we propose a novel Pixel-wise Keypoints Detection Network (PKDN) to address this problem, which first detects each 2D keypoint of each part with a pixel-wise detection style, and then solves its pose using PnP algorithm. Our approach is based on 2D vision and a low-cost imaging technique. Unlike prior works that required 6 degrees of freedom (6DoF) posture annotation or holistic regression, PKDN just needs to label the 2D keypoints, and can handle the object occlusion by the pixel-wise detection manner. Experiments on the various scenarios are implemented to illustrate the efficacy of our methods, and we also conduct an actual robotic bin-picking test on the presented system to evaluate its application performance.