Search papers, labs, and topics across Lattice.
CMU MLCenter of Data for Public Good (CDPG)Department of Computational and DataRobert Bosch Center for Cyber PhysicalMar 5, 2026arXiv:2603.05217This paper introduces a scalable AI-driven Intelligent Transportation System (AIITS) that processes hundreds to thousands of CCTV streams in real-time using an edge-cloud fabric. The system employs DNN inferencing on edge devices (Jetson Orin) for detection and tracking, followed by Spatio-Temporal GNNs in the cloud for traffic nowcasting and forecasting. A capacity-aware scheduler balances load across heterogeneous devices, and SAM3-assisted labeling with Continuous Federated Learning updates edge-based DNN detectors, achieving stable ingestion of up to 2000 FPS and accurate ST-GNN forecasts for up to 1000 streams.
Achieve real-time traffic analytics across city-scale camera networks by offloading DNN inference to edge devices and using cloud-based GNNs for forecasting, all while dynamically adapting to changing conditions with federated learning.
Real-time city-scale traffic analytics requires processing 100s-1000s of CCTV streams under strict latency, bandwidth, and compute limits. We present a scalable AI-driven Intelligent Transportation System (AIITS) designed to address multi-dimensional scaling on an edge-cloud fabric. Our platform transforms live multi-camera video feeds into a dynamic traffic graph through a DNN inferencing pipeline, complemented by real-time nowcasting and short-horizon forecasting using Spatio-Temporal GNNs. Using a testbed to validate in a Bengaluru neighborhood, we ingest 100+ RTSP feeds from Raspberry Pis, while Jetson Orin edge accelerators perform high-throughput detection and tracking, producing lightweight flow summaries for cloud-based GNN inference. A capacity-aware scheduler orchestrates load-balancing across heterogeneous devices to sustain real-time performance as stream counts increase. To ensure continuous adaptation, we integrate SAM3 foundation-model assisted labeling and Continuous Federated Learning to update DNN detectors on the edge. Experiments show stable ingestion up to 2000 FPS on Jetson Orins, low-latency aggregation, and accurate and scalable ST-GNN forecasts for up to 1000 streams. A planned live demonstration will scale the full pipeline to 1000 streams, showcasing practical, cross-fabric scalability.
