Intelligent Traffic Ecosystem — Edge Detection

01The system

The platform replaces manual police radio coordination with camera-based monitoring: a vehicle crossing an intersection becomes a live dashboard update in under 500 milliseconds. Getting there means no round trip to a central server for inference — the detection has to happen where the camera is.

02My component — detection-edge

I developed the edge detection service. It runs YOLOv8n for object detection with ByteTrack for multi-object tracking on a Raspberry Pi 5, so each vehicle keeps a stable identity across frames and a crossing is counted once rather than once per frame.

Detections are published as events to Kafka. Deployment across the device fleet is managed over the air with Balena, which matters when the hardware is mounted on a pole rather than sitting on a desk.

03The wider architecture

Downstream, Kafka feeds Flink stream processing with ST-GCN congestion forecasting, and FastAPI plus WebSocket backends push updates to the dashboards. The whole system deploys through Kubernetes-based GitOps with ArgoCD, monitored by Prometheus and Grafana.

Working inside that pipeline meant treating the edge component as a well-behaved producer: bounded latency, stable event schema, and graceful behaviour when the network is not.