A lightweight camera-based solution for checking a blind spot programmatically using TensorFlow and Python on a Raspberry Pi.
This repository contains multiple elements of the project. These elements include:
- Jupyter Notebook going into how the model was made
- The trained model
- Python program that uses the model on two cameras
Using transfer learning on MobileNetV2, an accuracy of ~98% was reached for blind spot detection with an average prediction time of 0.09s on the Raspberry Pi 4 without any machine learning accelerators. Given an ML accelerator such as the Google Coral USB Accelerator, it would likely reach prediction times of 0.0026s (2.6ms, source).
Want to skip straight into the details? Check out this video demoing the machine learning algorithm.
This project is under the GNU General Public License, version 3. More info is available in LICENSE