From 50% to 95% detection accuracy, running unattended in the field
How a traffic monitoring system went from a working pilot to 300+ autonomous sensor units in production, with every architecture decision made for unattended field operation.
Commissioned by Head of Product, German traffic-technology company Traffic monitoring, smart city
- detection accuracy, up from ~50%
- 95%
- units in the field
- 300+
- on Jetson Orin NX
- 30+ FPS
- camera to result
- < 200 ms
The challenge
The client needed a traffic monitoring system that runs autonomously and reliably at intersections, multi-lane roads and roundabouts. The existing solution detected vehicles with about 50% accuracy. That was far from enough for a production rollout with hundreds of sensor units.
What was required was not just a better model, but a complete system that runs 24/7 under real conditions: at night, in rain, under changing light, unattended in the field for weeks and months.
The approach
The work was production-oriented from day one. Rather than optimizing detection accuracy in isolation, the team built the complete system, from hardware integration to security architecture.
The decisive difference: every architecture decision assumed that hundreds of these units would run unattended in the field for weeks and months.
Edge-AI.DE built the inference pipeline on NVIDIA Jetson Orin NX and used TensorRT INT8 optimization to reach 30+ frames per second at under 200 milliseconds end-to-end latency. In parallel, a Docker-based microservice architecture made updates possible during live operation.
Technical implementation
Inference pipeline
The core is a GStreamer / DeepStream pipeline. Seven stages run on the device; only compressed vehicle tracks travel over MQTT to AWS IoT Core and the cloud application.
On the device · Jetson Orin NX
-
CSI-2 camera
Sony IMX662
-
Undistortion
Lens correction
-
Detection
TensorRT INT8
-
Tracking
-
Classification
11 BASt classes
-
Track compression
Ramer–Douglas–Peucker
-
MQTT transmitter
TLS
In the cloud
-
AWS IoT Core
Mutual TLS
-
Cloud application
Tracking sits between detection and classification, so classification works on tracked vehicles rather than single frames. Track compression with the Ramer–Douglas–Peucker algorithm reduces each trajectory to its significant points before it is transmitted.
Hardware platform
Every hardware decision answered one question: what does a device need to stand unattended at an intersection for months?
- NVIDIA Jetson Orin NX
- GPU-accelerated inference at low power, suited to battery-powered outdoor deployments.
- CTI Hadron carrier board
- Industrial-grade board for continuous field operation.
- Sony IMX662 sensor
- Strong low-light performance for 24/7 operation under changing light.
- Teltonika RUT241 LTE gateway
- A separate gateway avoids LTE certification overhead and enables out-of-band remote management.
Security architecture
Layered security from the boot process to the network:
- Secure boot
- A chain of trust anchored in the Jetson’s write-once fuses: the boot ROM verifies the bootloader, the bootloader verifies the kernel. Only signed software starts.
- Encryption
- TLS for all communication in transit, encrypted OTA updates.
- Cloud communication
- Mutual TLS with AWS IoT Core, using pre-provisioned device certificates.
- Network
- Only ports 80 and 443 reachable from outside. No inbound SSH. Remote access via Teltonika RMS.
- Least privilege
- The admin app runs without root. Privileged operations are isolated in a separate system service with a minimal API.
Deployment and operations
- Mobile variant
- Battery-powered, 72 h battery runtime, mast-mounted
- Permanent variant
- Fixed installation on site power
- Operations
- 30+ days unattended, automatic recovery in < 5 min, OTA updates in < 15 min
- Fleet management
- Device inventory, health monitoring, remote diagnostics and remote access over a separate management channel
GDPR compliance
All AI processing happens on the device. Raw video never leaves it; only encrypted, compressed vehicle tracks are transmitted. Privacy by design makes GDPR compliance a property of the architecture rather than an afterthought.
Results
| Metric | Before | After |
|---|---|---|
| Detection accuracy | ~50% | 95% |
| Vehicle classification (11 BASt classes) | 56% | >90% |
| Inference performance | – | 30+ FPS, < 200 ms |
| Unattended operation | 1–3 days | 30+ days |
| Auto-recovery | manual remote reboot | < 5 min |
| OTA update | – | < 15 min |
| Units in the field | – | 300+ |
The system is in production today in two variants: a mobile unit for temporary traffic counts and a permanent installation for continuous operation. The client supplies it to customers across Germany.
Same stack, other sensors
Camera, radar, PTZ: the detection, tracking and classification stack that runs our traffic fleet also drives these prototypes on NVIDIA Jetson.
-
Quality control of machine output
Prototype Industrial inspection
Checks every part a machine puts out, automatically, and flags the defective ones.
On the device:
- Camera at the machine output
- Detection
- Defect classification
- Pass / fail result
-
People re-identification across rooms
Prototype Buildings and indoor spaces
Recognises the same person again in another room or on another camera and continues their track.
On the device:
- Face detection
- Embedding per face
- Re-identification
- Track across cameras
Embeddings are computed and matched on the device; face images are neither stored nor transmitted. Built for controlled settings with a legal basis.
-
Flying objects: radar plus PTZ camera
Prototype Airspace monitoring
Detects flying objects by radar, points a PTZ camera at each one and classifies it from close-up photos.
On the device:
- Radar detection and tracking
- PTZ camera slews and zooms
- Photo capture
- Classification
Have an edge AI project that needs to reach production?
Start with a 30-minute technical call about your hardware, your task and your timeline.
Free and without obligation. We reply within 24 hours.