Development of a Low-Cost Monocular Vision System for Robotic Grasping of Dairy Bottles on Flexible Conveyor Lines

Abstract

This paper presents a low-cost monocular computer vision system for real-time detection and 3D localization of round objects (dairy bottle caps) on a conveyor belt, designed for robotic grasping by an industrial manipulator KUKA KR-3. The system addresses the challenge of achieving acceptable accuracy with minimal hardware — using only a single RGB camera and one static ArUco calibration marker. The proposed approach consists of two main modules: 1) object detection based on the Hough Gradient Method for precise (x, y) localization in the image plane; 2) monocular metric depth estimation using the Depth Anything V2 foundation model to generate a relative depth map, followed by simple scaling with a single ArUco marker placed on the conveyor. Experimental evaluation on a dataset of 45 image showed a mean absolute error in depth estimation of 1.81 cm. This accuracy level is sufficient for reliable grasping of dairy bottles, as the error is effectively compensated by gripping the object at its central height. The system demonstrates good robustness to typical production conditions, including specular reflections on plastic surfaces and varying illumination. The developed solution offers a cost-effective and easily deployable alternative to expensive stereo or LiDAR-based systems, making it highly suitable for flexible production lines in the dairy and organic food industry. The results confirm the practical applicability of the monocular approach for Industry 4.0 automation in agro-industrial enterprises.