Privacy-Preserving Edge AI Traffic Monitoring
Research · Edge AIScroll ↓Overview
A crowd-sourced traffic monitoring platform where edge devices detect vehicles on-device and a web dashboard shows live analytics. Published at IEEE COMSNETS 2026.
Details
- Role
- Design & development
- Category
- Research · Edge AI
- Links
- GitHub ↗IEEE paper ↗
- 7road-user classes detected on-device
- 11REST API endpoints
- 0raw images leave the device
Problem
Cloud video analytics for traffic is bandwidth-heavy, slow, and exposes raw footage of the public. CrowdLense detects vehicles on low-cost edge devices so only anonymous counts — never images — leave the device.
01 — System Design & Edge Intelligence
- Raspberry Pi camera nodes run YOLO11-s, trained on the Indian Driving Dataset (IDD), to detect and classify 7 road-user classes (cars, motorcycles, auto-rickshaws, trucks, buses, bicycles, pedestrians) entirely on-device.
- Privacy by design: devices transmit only aggregated, timestamped vehicle-count telemetry to a serverless edge-to-cloud pipeline, removing the need to stream or store raw video.
02 — My Contributions — Full-Stack Platform & Vision Integration
- Built the React 18 + TypeScript web platform with role-based dual dashboards: Providers register devices, provision IoT credentials and track online/offline status; Viewers explore all live sensors on an interactive Leaflet map.
- Developed the Flask REST API with JWT authentication — 11 endpoints across auth, provider, viewer and device-ingestion routes — plus CSV export of historical traffic data.
- Integrated the edge vision model's output with the backend and built Recharts analytics (vehicle-type distribution, device-level trends) with a client-side device cache, so field detections appear on the dashboard in near real time.
Built with
YOLO11-s / Raspberry Pi / Python / OpenCV / Flask 3 / JWT / React 18 / TypeScript / Vite / Tailwind CSS v4 / shadcn/ui / Leaflet / Recharts