Modern cities demand smarter traffic systems. Legacy setups rely heavily on cloud-only processing, leading to bandwidth strain, high latency, and privacy risks. TrafNet, running on the NEMO Meta-OS, changes that by bringing AI-driven traffic analytics to the edge, with seamless orchestration across edge and cloud.
At the core of TrafNet is a containerized microservice architecture. Each processing task—whether video decoding, object detection, or vehicle counting—runs as an independent component, connected via lightweight brokers such as NATS and MQTT. This ensures:
- Fast deployment through DevOps automation
- Scalability across diverse hardware (from Jetson edge devices to cloud GPUs)
- Continuous integration and updates without disrupting live services
NEMO’s Meta-Orchestrator dynamically allocates workloads. Busy junctions get more compute power when needed, while low-traffic zones conserve resources. The Secure Execution Environment (SEE) ensures all deployments meet privacy and regulatory requirements like GDPR.
Video Decoder: Streams RTSP feeds into NumPy frames with ~20–30ms latency.
Resize & Crop: Keeps distant vehicles detectable while cutting unnecessary pixels.
Keyframe Extraction: Reduces redundant frames by up to 80% in low-traffic hours.
Object Detection: YOLO/DETR models identify vehicles and pedestrians in real
time.
Object Tracking: Algorithms like DeepSORT maintain persistent IDs across frames.
Segmentation: SAM and YOLOv8 provide pixel-level masks for precise analysis.
Lane Detection: Lightweight process at 60 FPS, mapping lane boundaries instantly.
Multi-Camera Re-ID: Vision Transformers match vehicles across junctions for travel-time
analysis.
Object Counting: Region-based counting helps optimize signals.
Speed Estimation: Calculates real-time vehicle speeds for enforcement and safety.
TrafNet isn’t just about processing—it’s about enabling smarter mobility decisions. Application services expose REST APIs for real-time integration into traffic control systems:
1. Vehicle Queuing – Counts waiting vehicles at red lights.
2. Vehicle Moving – Detects when to end green phases based on approaching cars.
3. Pedestrian Counts – Replaces outdated push-button systems with live detection.
4. Green Requests – Ensures signals respond to actual demand, reducing red-light running.
5. Travel Time Across Junctions – Tracks vehicle journeys across multiple cameras to measure congestion.
By blending edge AI, event-driven microservices, and city-scale orchestration, TrafNet delivers:
- Lower bandwidth costs by cutting raw video uploads
- Sub-second response for safety-critical use cases
- Privacy-first operations aligned with European regulations
- A scalable blueprint for smart cities
