Real-time Multi-Object Tracking and Fashion Attribute Analysis: An Asynchronous Edge-Cloud Architecture for Interactive Digital Signage

Personalized customer experiences in offline fashion retail increasingly depend on real-time visual analysis. However, performing multi-object tracking and deep learning-based attribute analysis simultaneously on a single computing resource introduces significant latency and frame drops. This paper presents an asynchronous edge-cloud hybrid architecture that decouples lightweight visual tracking from cloud-based fashion attribute analysis. On the edge device, YOLOv8 object detection and the BoT-SORT algorithm maintain real-time multi-object tracking with persistent identity assignment. To prevent blocking of the main processing loop during cloud communication, a multi-threaded asynchronous pipeline is employed, incorporating an exponential backoff retry mechanism for network resilience. A Finite State Machine (FSM) with spatial-temporal thresholds filters incoming detections, triggering cloud analysis only for objects exhibiting sustained proximity to the signage display. The system was deployed in a pop-up store operated by a major Korean fashion e-commerce platform, where it maintained a stable frame rate above 30 FPS under concurrent analysis workloads and demonstrated dynamic resolution adaptation via HDMI hot-plug detection. As this work was conducted within a commercial project, the evaluation is limited to a single deployment environment without controlled baseline comparisons, which represents an area for further systematic investigation.