mypocketkitchen/pk-fooddetector

Pocket Kitchen | Custom TensorFlow object detection for food items with Raspberry Pi, HD Camera and Coral TPU

Pocket Kitchen | Object Detector

Pocket Kitchen is a system for managing food more efficiently in the home towards reducing household food waste. These scripts send detected incoming/outgoing food items in a fridge to a database.

• On-device object detection - Pocket Kitchen uses the Coral TPU and EdgeTPU TensorFlow models to perform object detection on-device. This means that no data is processed on the cloud, and all processing is done on the device itself.
• Customizable - Pocket Kitchen is designed to be customizable. For example, the user can add their own model trained on their own objects.
• Open Source - Pocket Kitchen is open source, so anyone can contribute to the project.

System Architecture

While similar to an API, these scripts are running locally and populating a database. The database is then accessed by the Pocket Kitchen app to display the data.

Requirements

Hardware

• Raspberry Pi 4
• Coral TPU
• Pi Camera (We use HD Camera Module V2, but any Pi camera should work. Autofocus is recommended.)
• 3D-Printed Camera Mount

Model

• You can use the enclosed model or create your own custom fine-tune EfficientDet-Lite0 model from TensorFlow Edge Object Detection

Installation

$ bash setup.sh

Usage

$ python3 detect.py --model <model> 

Future Work

• Add support for true API

Resources

• TensorFlow EfficientDet-Lite0 Object Detection for the Edge TPU