A new open-source pipeline achieves real-time YOLOv8n UAV detection at the sensor's 46 FPS ceiling while consuming only ~140 MB of RAM on Rockchip RK3588S chips. The system captures 1080p MIPI frames and distributes inference across all three NPU cores in parallel, pushing throughput from approximately 31 FPS to the camera hardware limit.

All heavy lifting offloads to dedicated silicon blocks: the ISP handles capture, RGA manages color conversion and resizing, while the NPU runs inference. This hardware-accelerated approach keeps the CPU idle and memory usage flat, enabling deployment on budget-friendly boards with just 2 GB of RAM rather than expensive development kits.

When a tracked UAV exits the frame, an on-device Qwen2.5-0.5B LLM generates natural-language event summaries using the same NPU through a hand-off control plane. The architecture chains independent processes via Unix-domain sockets, connecting detection, ByteTrack multi-object tracking, temporal feature extraction, and presence state management.

Built and tested on Khadas Edge2 hardware, the project demonstrates that sophisticated computer vision with AI-powered analytics can run on commodity ARM boards costing around €90. The complete pipeline architecture and build instructions are documented in the repository.