Show HN: Dual YOLOv8n UAV Detection on RK3588S at 42 FPS Using NPU (github.com)

A new project showcased on Hacker News demonstrates real-time UAV detection using the YOLOv8n model on the Rockchip RK3588S system-on-chip, achieving an impressive 46 frames per second (FPS) while utilizing only around 140 MB of RAM. The innovative setup fully leverages the SoC’s three neural processing unit (NPU) cores in a parallel processing configuration, moving beyond a previous throughput limit of ~31 FPS. This high-throughput, low-resource computer vision pipeline is designed for affordability, as it runs efficiently on budget-friendly 2 GB RK3588S boards, thereby expanding access to advanced AI applications in drone detection without requiring high-end hardware. Significantly, when a detected UAV exits the camera scene, an on-device language model (Qwen2.5-0.5B) generates a natural language summary of the event, showcasing the integration of AI models for contextual understanding. The architecture is modular, with distinct processes for object detection, tracking, feature extraction, and language summarization, all interconnected via Unix-domain sockets to optimize performance without taxing the CPU. This development not only enhances real-time surveillance capabilities but also paves the way for future applications of AI in resource-constrained environments and could inspire further innovations within the AI/ML community.