Back to browse
GitHub Repository

Real-time YOLOv8n UAV detection at the sensor's 46 FPS ceiling, in ~140 MB of RAM

47 starsC++

Dual YOLOv8n UAV Detection on RK3588S at 42 FPS Using NPU

by alebal123bal·Jun 14, 2026·71 points·9 comments

AI Analysis

●●●BangerWizardryNiche GemBig Brain

Hits the 46 FPS sensor ceiling with 140 MB RAM using all 3 NPU cores in parallel.

Strengths
  • Multi-threaded NPU inference saturates the camera, not the pipeline — genuine constraint engineering.
  • NPU hand-off between vision pipeline and Qwen2.5-0.5B LLM is clever resource multiplexing.
  • Runs on €90 RK3588S boards with 2GB RAM, not just expensive dev kits.
Weaknesses
  • Very niche audience — embedded CV engineers and drone detection is a narrow use case.
  • Zero stars and forks suggests early stage; no production deployments shown yet.
Category
Target Audience

Embedded systems engineers, edge AI developers, drone detection specialists

Similar To

NVIDIA Jetson inference pipelines · Rockchip RKNN Toolkit · Edge Impulse

Similar Projects

AI/ML●●Solid

True and4500 FPS Edge Detection at 1K (CPU-only)

4500 FPS edge detection on CPU via bitwise hacking; clever but unclear practical advantage over GPU.

WizardryBig Brain
andrespi
103mo ago