TrafficFab: An Autonomic Edge-Cloud Testbed Fabric forAI-Driven Traffic Management

TrafficFab is an autonomic edge–cloud testbed fabric designed for AI-driven traffic management in emerging megacities, addressing the need for real-time analytics over thousands of CCTV streams under strict latency, bandwidth, compute, and energy constraints.

The system validates a representative slice of a megacity deployment (approximating 10% of Bangalore’s camera network) by sustaining real-time analytics for ≈400 live camera streams. It combines RTSP stream emulation, heterogeneous edge inference (using DNNs on Jetson accelerators and Raspberry Pis), and cloud-based nowcasting/forecasting via Spatio-Temporal Graph Neural Networks (ST-GNN).

Key features include: Autonomic Control: Enables fine-grained scale-out/in of edge inference through energy- and migration-aware scheduling, and elastic scaling of GNN forecasting on public clouds. Continual Adaptation: Uses foundation-model (FM)-assisted Federated Learning to periodically adapt detection models without centralized video collection, preserving privacy and handling traffic drift. * Validation Scale: Demonstrated feasibility for closed-loop traffic analytics, short-term operational decision support, and longer-horizon planning at a scale practical for city-wide rollout.

TrafficFab addresses a core systems challenge in city-scale AI: how to run real-time video analytics over thousands of CCTV streams without exceeding practical limits on latency, bandwidth, compute, and energy. The paper presents an autonomic edge–cloud testbed fabric that emulates a representative slice of a megacity traffic-management deployment, rather than a single isolated model or application. Its central contribution is a platform-level architecture that can ingest camera video, for example via RTSP, distribute analytics workloads across edge and cloud resources, and adapt execution in response to system conditions. In effect, it treats traffic analytics not merely as an ML problem, but as a constrained, distributed systems problem.

The material is significant because it moves the discussion from “can an AI model detect vehicles or congestion?” to “can a city-scale video analytics stack operate reliably under real infrastructure constraints?” By providing a testbed that is both representative and autonomic, TrafficFab offers a useful substrate for studying workload placement, resource orchestration, energy-aware scheduling, and end-to-end latency tradeoffs in traffic management. For researchers and practitioners building AI-driven smart-city systems, the work matters because it bridges the gap between prototype-level analytics and deployable, measurable, city-relevant infrastructure.