Containerized Deployment Strategies for Scalable AI Surveillance in Urban Environments

Scalable deployment of AI-driven surveillance remains a central challenge in modern urban infrastructures, where heterogeneous workloads, real-time requirements, and energy constraints intersect. This study presents a containerized system architecture integrating microservices, GPU-aware orchestration, and hybrid cloud–edge pipelines to enable reliable and efficient large-scale video analytics. A city-scale simulation comprising 520 concurrent video streams was employed to evaluate the proposed framework against monolithic and static-partition baselines. Results demonstrated 99.2% operational uptime and sustained sub-second average latency (~0.61–0.78 s) across loads up to 500 streams, with latency degradation remaining below 0.9 s at the 95th percentile. Quantization-aware training maintained model accuracy within 0.5 percentage points of full precision while reducing inference time by 19–24% and energy consumption per frame by 12–15% compared to baseline. Energy profiling revealed that GPU accelerators consumed 240–270 W under peak loads, whereas NPUs maintained stable power at ~70 W, highlighting the complementary potential of heterogeneous allocation. These findings collectively confirm that containerization, when coupled with adaptive scheduling and model-level optimizations, provides a robust pathway for transitioning AI surveillance from prototype systems to resilient city-scale deployments.