Resource-Aware Distributed Edge Framework for Heterogeneous Video Analytics

Abstract

The widespread deployment of cameras for surveillance and monitoring applications has created a growing demand for scalable real-time video analytics solutions. Although cloudcentric approaches provide high inference accuracy, they suffer from increased latency, bandwidth consumption, and privacy concerns. Edge computing addresses these limitations by enabling inference closer to the data source; however, heterogeneous edge infrastructures introduce significant challenges in workload allocation and resource management. This paper presents a resource-aware distributed edge framework for video analytics across heterogeneous compute environments. The proposed framework introduces a lightweight hardware abstraction and scheduling mechanism that dynamically maps video streams to edge nodes based on node capability scores and stream intensity characteristics. The scheduling framework combines hardware profiling, normalized compute scoring, stream motion analysis, and adaptive workload dispatch across resource-constrained and GPU-enabled devices. The framework is demonstrated using an intrusion detection application integrating YOLO and SSD object detection models, RTSP video streams, H. 265 compression, MQTT telemetry communication, and ThingsBoard CE visualization. Experimental evaluation across Raspberry Pi devices, Intel mini PCs, and GPU-enabled desktop systems demonstrates adaptive workload placement, efficient utilization of heterogeneous edge resources, and the ability of the framework to support increasing concurrent video workloads within the evaluated deployment scale while maintaining real-time inference performance.