Real-Time Retail Shelf-Stock Detection with YOLOv7

This study developed a real-time shelf stock monitoring system for retail environments, leveraging the You Only Look Once version 7 (YOLOv7) deep learning-based object detection framework. The system effectively addresses the inefficiencies, delays, and errors inherent in manual stock auditing processes. The underlying model was trained on a comprehensive dataset comprising 15,397 annotated object labels across fifteen distinct retail product categories. The fully trained model was then integrated into a web-based platform designed to capture real-time shelf images via a webcam. These captured images undergo automated processing for product detection and counting. The detection results are dynamically displayed on an interactive dashboard and securely stored in a backend database. The system also incorporates voice alerts, which are triggered automatically when stock levels fall below predefined thresholds, thereby facilitating immediate restocking. Experimental validation indicates high performance, with both precision and recall exceeding 96%, and an average processing latency of less than one second per frame. The model achieved an [email protected] of 0.996 and an [email protected]:0.95 of 0.86. These findings underscore the system's effectiveness in providing a rapid, accurate, and efficient monitoring solution specifically tailored for small to medium-sized retail businesses. The primary contribution of this research lies in its comprehensive, end-to-end system integration, combining robust YOLOv7-based object detection with real-time web visualization and automated voice alerts, successfully addressing existing gaps in prior implementations.