The growing adoption of urban video surveillance and traffic monitoring increases demand for low‑latency, cost‑efficient, and privacy‑preserving analytics. Pure cloud offloading introduces network bottlenecks and unacceptable end‑to‑end delay for time‑critical use cases. We evaluate a hybrid edge–cloud architecture that performs preliminary inference at the edge and delegates aggregation, visualization, and archival tasks to the cloud. Using Raspberry Pi–class devices and a managed cloud back end, we benchmark latency, bandwidth consumption, CPU utilization, and accuracy on representative urban scenes (intersections, pedestrian zones, parking lots). Compared to cloud‑only processing, the proposed hybrid approach reduces median latency by 71% and bandwidth by 94%, while maintaining accuracy within 3–5% of full‑precision models. We discuss design trade‑offs, security and privacy considerations, and deployment guidance for city‑scale systems.