Operational Stability Analysis of Edge-Based Video Analytics under Fluctuating Workloads

Abstract

Edge computing platforms are commonly used for real-time video analytics in transportation, security, and facility management environments. While accelerator hardware reduces inference latency, overall system stability remains difficult to maintain under fluctuating workloads. An edge-based video processing pipeline was deployed across multiple monitoring sites and evaluated over a ten-month period. Performance logs show that decoding, buffering, and network transmission frequently dominate end-to-end latency once inference time is reduced below a certain threshold. Storage queue congestion and synchronization delays contributed to periodic service interruptions. Several infrastructure-level optimizations were implemented, including adaptive buffering and task co-location. Although throughput improved, sustained reliability depended largely on operational tuning rather than algorithmic upgrades.