Automatic Pothole Detection System from Road Surface Images
This project was completed as the master’s thesis for the Data Science & Big Data program at U-TAD.
The work carried out for the project includes:
- A study of the state of the art in object detection
- Development of a Python implementation of YOLO V3 and YOLO V3 Lite
- Collection and preparation of a dataset of road images
- Training several YOLO models with different configurations and conducting a comparative evaluation of their results
- Conversion of the trained model for deployment on a mobile device
- Development of an Android application to run the model
The project is organized into the following repositories:
| Repository | Description |
|---|---|
| tfm | Contains a Python implementation of YOLO V3 and YOLO V3 Lite |
| tfm-android | Contains an Android mobile application that runs the trained model using TensorFlow Lite |
| tfm-doc | Contains the thesis document [ES] |
References:
- The images used to train the models were obtained from Kaggle
- The mobile application is based on a TensorFlow Lite example
- The YOLO V3 implementation is a fork of this repository, and the YOLO V3 Lite implementation is based on this one. Both implementations were unified into a single codebase, additional model configuration options were added, and support for more image label formats was included
