IndustrialEdgeML - End-to-end edge-based computer vision systemfor Industry 5.0
Raphael Wagner, Mario Matuschek, Philipp Knaack, Michael Zwick, Manuela Geiß
Abstract
State-of-the-art object detection methods gain increasing attention in various industrial applications. However, current vision systems typically rely on cloud services for data storage and training of AI models. In this work, we present an end-to-end vision system for bin-picking applications at the edge of industrial premises via industrial computers, where bin-picking refers to the dedicated selection and displacement of objects into designated areas. Our system is designed for human-robot collaboration in a range of distinct challenging industrial environments with a focus on automatically learning new objects. To this end, we propose a new approach for automated oriented object detection consisting of three stages: (i) automatic training data labeling for the new object based on autoencoders, (ii) object detection using the YOLOR-S architecture, and (iii) inference of the object's orientation, more precisely, rotated bounding box detection based on attention maps using the TS-CAM algorithm. Our method achieves state-of-the-art performance, without the need for annotations of rotated bounding boxes, when compared to the single-stage architecture S2A-Net for rotated bounding box detection. In particular, the system maintains high-quality performance not only for learning all objects at once but also for incrementally learning one object after another.