Google Coral Dev Board Experiments

A collection of computer vision experiments running on the Google Coral Dev Board, exploring what the Edge TPU can do in practice.

Hardware

• Google Coral Dev Board — NXP i.MX8M SoC + Edge TPU (4 TOPS)
• USB Webcam — Logitech C110
• OS: Mendel Linux (Debian-based)

Projects

Folder	Description
detect_people/	CLI script — detects people in a live webcam feed and prints results to stdout
web_app/	Flask web app — live MJPEG stream with bounding boxes around detected people
detect_family_app/	Transfer learning via weight imprinting — trains a custom classifier to recognize specific people, served as a live web app

Getting Started

Board setup

Flash Mendel Linux to the board, then install dependencies:

sudo apt-get install libedgetpu1-std python3-pycoral python3-opencv python3-flask

Models

Download models to ~/coral/ on the board:

# Person detection
wget https://github.com/google-coral/test_data/raw/master/ssd_mobilenet_v2_coco_quant_postprocess_edgetpu.tflite
wget https://raw.githubusercontent.com/google-coral/test_data/master/coco_labels.txt

# Face detection
wget https://github.com/google-coral/test_data/raw/master/ssd_mobilenet_v2_face_quant_postprocess_edgetpu.tflite

# Weight imprinting base model
wget https://github.com/google-coral/test_data/raw/master/mobilenet_v1_1.0_224_l2norm_quant_edgetpu.tflite

Notes

• All inference runs on the Edge TPU — expect ~5ms per frame for detection
• Weight imprinting needs ~5–50 labeled images per class and trains in seconds, no GPU required
• Family photos used for training are excluded from this repo (see .gitignore)