YOLOv8 Image Recognition on RPi 5 with 26 TOPS M.2 Hailo AI Hat+ (doleron.substack.com)

A how‑to walkthrough shows how to deploy a custom YOLOv8 detector on a Raspberry Pi 5 equipped with Hailo’s AI Hat+ (13/26 TOPS). The author trains a YOLOv8 model (Ultralytics) on a barcode/QR dataset, exports the PyTorch weights (best.pt) to ONNX, compiles the ONNX to Hailo’s HEF format with the Hailo DFC toolchain, and runs the HEF on the Pi to achieve ~30 FPS object detection in Full HD. The writeup highlights that the Hailo AI Hat+ brings real‑time, edge‑native inference to Pi users but that the compile/deploy toolchain is nontrivial. Key technical details: workflow is train → yolo export model=... format=onnx → hailomz compile --ckpt <best.onnx> --calib-path <images/> --yaml <yolov8s.yaml> --classes 2 --hw-arch hailo8 to produce yolov8s.hef. DFC/ HailoRT / HailoRT‑Python must be installed (example packages and git checkout given), and calibration requires >1024 images and a GPU for optimization. The author recommends Ubuntu 22.04 and Python 3.10 (use venv or Docker), shows example label JSON and detection.py usage on the Pi, and notes common pain points (versioning, environment separation, and lengthy compilation). The result demonstrates practical, high‑performance custom model deployment on a low‑cost edge platform, though tooling and setup remain the main barrier.