Let's Talk
DE

Onboard Object Detection

Real-time AI inference at 10.5 Mpx/s on resource-constrained platforms. YOLOX optimized for space-grade hardware.

10.5 Mpx/s
Throughput
26W
Power
YOLOX
Architecture

Deploy state-of-the-art object detection on resource-constrained space platforms. Our solution brings YOLOX performance to radiation-tolerant FPGAs, enabling autonomous decision-making at the edge.

Developed under armasuisse S+T contract, this capability has been demonstrated on Xilinx UltraScale+ MPSoC — and validated on heavy-lifting drone platforms with live camera feeds.

See It In Action

Real-time object detection demonstrated on a heavy-lift drone platform with live camera feed.

Technical Demonstrations

First FPGA/Object Detection demonstration from armasuisse Study 2

Screen-filmed imagery demonstrating real-time detection under various conditions

Detection and tracking performance on ships and traffic

Performance

10.5
Mpx/s
Processing throughput
26
W
Power consumption
YOLOX
Nano
Network architecture

Key Features

Quantization-Aware Training

INT8 quantization with minimal accuracy loss. Optimized for FPGA inference engines.

Custom Dataset Support

Train on your mission-specific objects. End-to-end pipeline from data curation to deployment.

FPGA Acceleration

Vitis AI integration for Xilinx devices. Pathway to radiation-hardened Versal processors.

Real-Time Processing

Frame-by-frame inference for autonomous operations. No ground-in-the-loop required.

End-to-End Capability

We deliver the full ML pipeline, not just a model:

1

Curate

Dataset creation, labeling, train/val/test split

2

Train

Quantization-aware YOLOX with version control

3

Compile

Vitis AI compilation for DPU execution

4

Integrate

Embedded OS, camera interface, inference pipeline

5

Verify

CI/CD with QEMU-based hardware-in-the-loop testing

6

Deploy

Production-ready demonstrator

Applications

Maritime Surveillance

Ship detection in coastal and open-water imagery.

Infrastructure Monitoring

Change detection for pipelines, roads, facilities.

Defense & Security

Real-time alerting for autonomous platforms.

Earth Observation

Onboard filtering to reduce downlink volume.

Applicable to: satellites, drones, UAVs, autonomous vehicles.

Technology Stack

Model

  • YOLOX-Nano
  • INT8 quantized

Training

  • PyTorch
  • Vitis AI

Hardware

  • Xilinx UltraScale+
  • Xilinx Versal

Integration

  • Linux / PetaLinux
  • Bare-metal option

Development Heritage

Image Preprocessing Pipeline

Completed

TensorFlow-based onboard preprocessing: demosaicing, calibration, geometric correction, projection. CI/CD automated with QEMU emulation.

armasuisse S+T Prime 2021

Object Detection Model

Completed

YOLOX training on swisstopo imagery. Quantization-aware optimization for small object detection. First hardware integration.

armasuisse S+T Prime 2022 – 2023

Real-Time Demonstrator

Demonstrated

Full integration on Xilinx UltraScale+. Achieved 10.5 Mpx/s at 26W. Demonstrated on heavy-lifting drone platform.

armasuisse S+T Prime 2023 – 2024
armasuisse S+T Program

Interested in Onboard AI?

Contact us for technical discussions or demonstration opportunities.

Get in Touch View All Products