Vision-Based Autonomous System for Counting and Inventory Monitoring of Metal Profiles

Maintaining up-to-date inventory data is essential for modern supply chains, yet transitioning to automated tracking frequently requires prohibitive structural modifications and substantial financial investments. To address this bottleneck and enable low-cost innovation within existing legacy warehouse infrastructures, we developed an industrially deployable, practical vision system that integrates deep learning, classical computer vision, and robotic elements, which we call the MOBOT system. Its vision pipeline unifies robotic image capture, depth-based ROI localization, perspective transformation, material identification, and lightweight geometric counting algorithms into an adaptive, field-ready framework. We evaluated this inventory-monitoring system in an operational warehouse storing semi-finished metal products. Experimental results confirm the system’s high robustness across various geometries: structured hollow shapes, such as tubes, hollow sections, and U-profiles, result in a low error rate of 2–4%, while for irregular, solid cross-sections, the fundamental error rate measured under real-world conditions is 7–20%. This robot-assisted platform delivers unambiguous semi-finished metal product inventory monitoring without requiring facility modifications.