With the rapid progress of deep learning and intelligent vision technologies, the retail industry is experiencing a transformative transition towards operations driven by Intelli-gent Media (IM). This paper introduces a novel intelligent shelf monitoring system to tackle major challenges in sustainable retail management, like high labor costs, inefficient inventory control, and limited real-time responsiveness. By leveraging deep learning and computer vision, the system combines enhanced object detection and robust multi-object tracking to enable automatic, accurate, and scalable shelf supervision. An improved YOLOv8 model with an integrated attention mechanism and Mosaic data augmentation is used to improve detection accuracy. Meanwhile, a refined Strong-SORT algorithm manages trajectory lifecycles to ensure stable tracking under occlusion.
The system is deployed through a lightweight asynchronous Web architecture based on Flask. It integrates real-time video stream processing, responsive visual interaction, and optimized FFmpeg encoding to reduce resource consumption. Experimental results indicate that the proposed system attains 86.7% mAP@0.5, a 14.4% improvement over the baseline. In occluded scenes, it maintains 78.4% MOTA with an 18.3% reduction in the ID switch rate. It also supports high concurrency with a 24.5 FPS processing speed and 200 requests per second.
This research contributes to the field of Intelligent Media by integrating deep learning into real-time retail surveillance. It aligns with Sustainable Technology Management by providing a low-resource, high-efficiency solution to enhance operational sustainability. The system shows high performance in precision, responsiveness, and scalability, offering a promising path for AI-driven intelligent retail environments.