Low-Cost Vision-Based Inventory Monitoring: Implementation and Evaluation

Abstract

This paper presents a low-cost computer vision system for monitoring inventory drawer states using consumer-grade hardware and open-source software. As part of the broader goal of enabling digitalization in resource-constrained industrial environments, especially small and medium-sized enterprises (SMEs), this work supports practical steps toward smart intralogistics and human-centric automation. The prototype uses a lightweight convolutional neural network (CNN) to identify which drawer is open and to classify the drawer’s fill level (empty, partially filled, full). Optical character recognition (OCR) was implemented and tested, but a CNN-based approach proved more robust and was adopted in the final implementation. To reduce the annotation burden, a semi-automated HSV-based method was developed to generate bounding boxes. The system was evaluated on a dataset of over 11,000 images captured under controlled laboratory conditions. The drawer identification CNN achieved over 95% classification accuracy, while the fill-level classifier reached 93% accuracy. Total inference time averaged under 50 milliseconds per image on a standard CPU, demonstrating suitability for near real-time use without specialized hardware. The results demonstrate the feasibility of deploying accessible, modular, and scalable vision-based inventory systems as part of cyber-physical production environments. The approach aligns with smart manufacturing goals by lowering entry barriers to digital automation and addressing common challenges in SME adoption of intralogistics technologies.