Traffic Monitoring Customer · Traffic Monitoring / Smart City · NVIDIA Orin NX

A Leading Traffic Monitoring System

How a leading traffic monitoring system was optimized from 50% to 95% detection accuracy and hardened for production deployment with hundreds of autonomous sensor units in the field.

95%
Detection Accuracy
from 50%
100+
Units in the Field
autonomous operation
30 FPS
Inference Rate
real-time
<200ms
End-to-End Latency
camera to result

The Challenge

The customer needed a traffic monitoring system that operates autonomously and reliably at intersections, multi-lane roads, and roundabouts. The existing solution detected vehicles with approximately 50% accuracy -- far from sufficient for production deployment with hundreds of sensor units.

What was required was not merely a better model, but a complete system that runs 24/7 under real-world conditions -- at night, in rain, under changing lighting. Unattended in the field for weeks and months.

The Approach

The approach was production-oriented from day one. Rather than optimizing detection accuracy in isolation, the team developed the complete system from hardware integration to security architecture.

The decisive difference: every architecture decision was made under the premise that hundreds of these units would operate unattended in the field for weeks and months.

Gradion EdgeAI built the inference pipeline on NVIDIA Orin NX and pushed model performance with TensorRT int8 optimization to achieve 15 frames per second at under 200 milliseconds end-to-end latency. In parallel, a Docker-based microservice architecture was developed to enable updates during live operation.

Technical Implementation

Hardware Platform

Every hardware decision followed a clear question: what does a device need that will stand unattended at an intersection for months?

  • NVIDIA Orin NX -- GPU-accelerated inference at low power consumption, ideal for battery-powered outdoor deployments
  • CTI Hadron Carrier Board -- industrial-grade board for continuous field operation
  • Sony IMX 662 Sensor -- superior low-light performance for 24/7 operation under changing lighting conditions
  • Teltonika RUT 241 LTE Gateway -- separate gateway eliminates LTE certification overhead and enables out-of-band remote management

Inference Pipeline

The core pipeline is implemented as a GStreamer/Deepstream pipeline with five processing stages:

Video Source
GStreamer/CSI-2
Primary Inference
YOLOv4/TensorRT int8
Object Tracker
Luenberger Observer
Message Sink
JSON Serialization
Redis Queue
Async Decoupling

Deepstream leverages the hardware acceleration of the Jetson NX to its full potential. The Redis queue decouples the high-performance C++/GStreamer pipeline from the Python data processing -- when the data collector is slowed by network issues, the inference pipeline continues uninterrupted.

Deployment and Operations

Deployment Variants

  • Mobile: 12V battery, 72-hour runtime, mast-mounted
  • Permanent: PoE+ over Ethernet, permanent installation

Operational Reliability

  • 30+ days unattended operation
  • Auto-recovery in <5 minutes
  • OTA updates in <15 minutes

Security Architecture

Multi-layered security approach at the network, application, and data levels:

Encryption

LUKS full-disk encryption (at rest), TLS for all communication (in transit), encrypted OTA updates

Cloud Communication

Device-to-cloud communication secured via public/private key infrastructure with pre-shared certificates for mutual TLS authentication with AWS IoT Core.

Network

Only ports 80/443 externally reachable. No inbound SSH. Remote access via Teltonika RMS.

Least Privilege

Admin app runs without root privileges. Privileged operations isolated in a separate system server with minimal API surface.

GDPR Compliance

All AI processing happens on-device. Raw video never leaves the device. Only encrypted counting data is exported. This privacy-by-design approach makes GDPR compliance an architectural property rather than an afterthought.

Results

Metric Before After
Detection Accuracy ~50% 95%
Vehicle Classification Insufficient >90% (11 BAST classes)
Inference Performance — 30 FPS, <200ms latency
Unattended Operation Days 30+ days
Auto-Recovery Manual <5 minutes
OTA Update — <15 minutes
Deployed Units — Hundreds in the field

The system is in production today in multiple variants: as a mobile solution for temporary traffic counts and as a permanent installation for continuous operation. The customer deploys the system to clients across Germany.

Have an Edge AI project that needs to reach production?

Let us discuss a no-obligation assessment.

Discuss your project