preloader

30–60 FPS Object Detection on NVIDIA Jetson Orin – Without Cloud

Train YOLOv models, TensorRT-optimise (INT8/FP16) and deploy on NVIDIA Jetson Orin – we handle the complete pipeline from labelling to commissioning.

Performance Figures

Metric Value Platform
Inference latency < 10 ms Jetson Orin NX 16 GB
Object detection 30–60 FPS Full HD, YOLOv8, Jetson Orin NX
AI compute 20–100 TOPS ORINHSN4 to ORINHSX16
Quantisation INT8 / FP16 TensorRT-optimised
Operation 24/7, offline No cloud dependency

Our Services

Model Training & TensorRT Optimisation

Training on own GPU hardware, optimisation for edge deployment via TensorRT (INT8/FP16 quantisation). Typical results: 3–5× faster inference compared to non-optimised models.

Camera System Development

Selection and integration of CMOS cameras for industrial and scientific applications. Focus on Sony IMX sensors, resolutions from 4 megapixels, USB3 and GigE Vision.

Custom Image Processing Algorithms

Tailored solutions: debayering, noise reduction, multiscale fusion for HDR, projection algorithms. No licence overhead, no vendor lock-in.

Data Annotation & Model Quality

Structured labelling with LabelStudio, training data quality assurance, class balancing. Typical data requirement: 500–5,000 annotated images depending on complexity.

Embedded AI & System Integration

Integration on NVIDIA Jetson Orin as DIN rail module (OrinHS). Industrial I/O, OTA updates, SSH remote access – ready to operate from delivery.

Technologies & Sensors

AI Stack:

Layer Technologies
Annotation LabelStudio
Training YOLOv8, PyTorch, TensorFlow
Optimisation TensorRT (INT8/FP16), CUDA C/C++
Deployment NVIDIA Jetson Orin (JetPack, DeepStream)
Inference 30–60 FPS, < 10 ms latency

Sensor Types:

  • Monochrome (higher sensitivity, NIR option)
  • One-shot colour (OSC) with Sony IMX sensors
  • Near infrared (NIR) for special spectral ranges
/images/services/gallery1/01.jpg
/images/services/gallery1/02.jpg
/images/services/gallery1/03.jpg
/images/services/gallery1/04.jpg
/images/services/gallery1/05.jpg

Why Custom Algorithms Instead of Standard Software?

OpenCV and Halcon cover many cases – but edge deployment on embedded hardware follows different rules:

  • Latency: Custom algorithms optimised for CUDA architecture and TensorRT
  • No licences: Full control over the code, no licence overhead on each device
  • Quality: Fine-tuning to sensor properties and use case
  • Long-term: No dependency on software vendors over 10+ year product lifetimes

Project Examples

Autonomous Allsky Camera for Sky Monitoring NVIDIA Jetson Orin Nano as control unit – high-resolution scene detection, < 10 ms inference, 24/7 operation without cloud. Real-time detection of clouds, aircraft and astronomical events. Services: optics design, camera system, embedded AI, image processing algorithms.

Schmidt Cameras for Astronomical Imaging Development of highly light-sensitive 5.5" and 8" Schmidt cameras with modern CMOS sensors and custom image processing pipeline for scientific astrophotography. Quantity: 4 systems.

FAQ

On an NVIDIA Jetson Orin NX we achieve 30–60 FPS at Full HD resolution with optimised YOLOv8 models. Inference latency is below 10 ms. Exact performance depends on model size, image resolution and quantisation level (INT8/FP16). For your application we can create benchmarks with real data.

For simple classification tasks, 100–500 images per class often suffice. For complex object detection we recommend 1,000–5,000 annotated images. The exact number depends on object variability and environmental conditions. We also support building suitable training datasets.

We focus on edge AI – evaluation runs directly on the device, without cloud dependency. This guarantees latencies below 10 ms, works offline and avoids data protection issues. For applications requiring centralised evaluation we can also implement cloud connectivity.

For most industrial applications we recommend USB3 cameras with Sony IMX sensors – good balance of resolution, sensitivity and price. For GigE Vision, high frame rates or NIR we select accordingly. We advise you on sensor selection.

Yes, we analyse existing systems and identify optimisation potential: better algorithms, TensorRT optimisation of existing models or faster hardware. Often targeted adjustments achieve significant improvements in detection rate or speed.

We integrate and qualify optical components but do not develop our own lenses. For selection of lenses, filters and illumination we support you based on our experience from projects such as the Schmidt cameras. For special applications we work with optical partners.

Do you need edge AI or image processing?

From sensor selection to TensorRT-optimised inference on Jetson Orin – we develop the complete pipeline.

Request project

Copyright (c) 2026 LyconSys GmbH & Co.KG