An indoor inspection station running entirely on a Raspberry Pi 5 with a Hailo-8L NPU and 2D LiDAR. LiDAR-triggered detection avoids the latency, connectivity dependence, and privacy exposure of cloud-based vision systems, with all sensing, detection, and recognition running locally.
Developed as part of an MSc course in AIoT Edge Computing at FH Technikum Wien, this project addresses two problems with cloud-dependent vision systems: the latency and connectivity dependence of sending frames over a network, and the privacy exposure of raw video leaving the device. A related inefficiency is always-on inference, which burns compute on scenes that are empty most of the day.
The system pairs a 2D LiDAR, which scans continuously at low power, with a Raspberry Pi 5 and Hailo-8L NPU. The LiDAR triggers a capture only when a person enters a configurable proximity threshold, so the heavier detection and recognition pipeline runs only when there's something to see. It evolved across three milestones: real-time detection with LiDAR fusion, an event-driven proximity gate that eliminates always-on inference, and a full system with face recognition, a live dashboard, and Telegram alerts.
YOLOv6n at ~28 FPS on Hailo-8L. Distance from LiDAR overlaid on every bounding box. Real-time LiDAR polar map rendered alongside.
LiDAR monitors entry zones and triggers camera snapshots with stateful cooldown to prevent re-triggering events.
Face recognition (dlib) classifies known vs. unknown visitors. Flask dashboard for live monitoring. Telegram alerts on detections.
All detections logged to rotating CSV files with timestamps and confidence scores for post-analysis.
| COMPONENT | SPECIFICATION |
|---|---|
| Compute | Raspberry Pi 5 (4–8 GB RAM) |
| AI Accelerator | Hailo-8L M.2 neural processing unit |
| LiDAR | Hokuyo URG-04LX-UG01 (USB serial) |
| Camera | Raspberry Pi Camera Module v2/v3 (CSI) |
| Model | YOLOv6n — ~28 FPS on Hailo-8L |
| OS | Raspberry Pi OS 64-bit Bookworm |
| Proximity threshold | Configurable — default 1.5 m |