Order picking is one of the most costly and labor-extensive operations in a typical warehouse. It entails 50% of operational expenses, thus being a high priority area for automation and cost reduction. Several attempts have been made towards an automation of order picking by companies and research groups around the Globe. However, there is still no solution suitable for order picking in a warehouse not specifically equipped for this purpose. The main bottleneck of applications where robots have to work in an environment built for humans is a perception ability. An appearance of low-cost depth sensors started with Microsoft Kinect, has opened a new research field in robotics that tries to provide better perception abilities to robots using a combination of depth and RGB data. Many methods have been proposed over last 5 years. These algorithms mostly use RANSAC-based techniques or one of the 3D point descriptors in order to fit a model into image. Many methods for deriving primitive shapes and mapping surrounding environments using RGB-D images originate in Photogrammetry and require a human input on processing stage for choosing potential candidates for detection. There are also methods built upon collected databases of CAD-models of objects to be detected. The goal of this work is to develop a computer vision algorithm for reliable registration and pose estimation of boxes on a pallet. We propose a novel approach to object detection and pose estimation of cuboid-like objects. It is based on a combination of several approaches (edge-based, descriptors based and region growing clustering-based). Such combination allows for more robust object detection in a highly cluttered environment where objects are occluded by each other. Proposed approach requires neither manual input nor a preloaded database of models to be detected.
Сколковский институт науки и технологий