Edge AI

Research area: running ML inference on resource-constrained devices (Raspberry Pi, embedded systems).

---

Current Work

Project: KAIT 2026 — Two-phase context-aware detection on RPi 5

Paper: "Adaptive Hierarchical Detection: A Two-Phase Framework for Context-Aware in Edge Device"

---

Two-Phase Pipeline

Input (camera frame)
  └── Phase 1: Scene Classification
        └── Places365 GoogLeNet → scene label
              └── Phase 2: Context-Aware Detection
                    ├── Indoor/simple → YOLO26n (fast, lightweight)
                    └── Outdoor/complex → YOLOWorld (open-vocab)

Key insight: use scene context to switch between detection models — avoid running heavy models when a lightweight one suffices.

---

Runtime: NCNN

NCNN is a high-performance neural network inference framework optimized for mobile/edge platforms. Key advantages over ONNX Runtime on RPi:

• No dependency on large frameworks (PyTorch, TF)
• Vulkan GPU acceleration support
• ARM NEON SIMD optimization
• Smaller binary size

---

Hardware

• Yahboom Raspbot on Raspberry Pi 5
• SSH access via Tailscale: 100.88.131.75 (takano-lab)
• CPU-only inference (no GPU on RPi 5 for ML)

---

Model Performance

Model	Task	Notes
Places365 GoogLeNet	Scene classification	Phase 1 gating
YOLO26n	Object detection	Fast path
YOLOWorld	Open-vocab detection	Accurate path
YOLOv8s (finetune)	Objects365 finetune	1 epoch: mAP50=10.9% CPU

---

Tools & Stack

• PyTorch → training
• ONNX → export format
• NCNN → edge inference
• OpenCV → video pipeline
• Python 3.11