June 14, 2026
Drone drama at 46 FPS
Show HN: Dual YOLOv8n UAV Detection on RK3588S at 42 FPS Using NPU
This tiny board spots drones insanely fast, and the comments turned into a hardware cage match
TLDR: A cheap mini-computer was shown detecting drones in live video as fast as the camera can feed frames, then writing a short summary after the drone leaves. Commenters were impressed but immediately started arguing over whether the older AI model choice was genius pragmatism or just playing it safe.
A modest little computer board just posted a very big flex on Hacker News: it can spot drones in live video at the camera’s full speed, while barely sipping memory and leaving the main processor free. In plain English, the maker squeezed almost all the hard work onto dedicated chips built for camera and AI tasks, so even a cheap device can keep up. And because apparently that wasn’t enough, when a drone disappears, the system can even generate a short plain-language recap of what happened. Yes, the commenters absolutely noticed the “tiny gadget writes its own drone report” twist.
But the real popcorn moment was the YOLO version debate. One commenter basically asked, “Why is everyone still using YOLOv8 when newer versions exist?” and that instantly injected classic tech-forum energy: is this smart engineering, or are people clinging to an older favorite because it actually works on limited hardware? The creator jumped in to say the real magic wasn’t the model name at all, but the pipeline design and the choice to prioritize fresh live video over flashy benchmark tricks.
That sparked the second mini-drama: batching. Some people love stacking frames together for raw speed numbers, but the creator argued that for a live camera feed, that just makes the video feel stale. Translation: commenters got a familiar showdown between “best benchmark score” and “best real-world experience.” The vibe was equal parts impressed, nerdy, and slightly combative — with a side of disbelief that a budget board is now doing drone spotting and post-game commentary.
Key Points
- •The project runs real-time 1080p UAV detection on RK3588S hardware using YOLOv8n across all three NPU cores and reaches the OS08A10 camera’s 46 FPS limit.
- •The pipeline offloads capture, preprocessing, and inference to the ISP, RGA, and NPU respectively, keeping CPU use low and memory near 137–152 MB RSS per stream.
- •Two camera streams can run side by side, with reported memory use of about 276–304 MB RSS for dual-stream operation.
- •The software is organized as independent processes connected by Unix-domain sockets, with stages for detection, ByteTrack tracking, temporal features, presence FSM, and optional LLM summarization.
- •An on-device Qwen2.5-0.5B model can generate a natural-language summary after a tracked UAV exits the scene, with the pipeline temporarily handing full NPU access to the LLM.