Safety-System-Design 🚗

Group E - Mattia Carlino, Mikael Motin, Lorenzo Paravano, Yusheng Yang

Introduction

This project was conducted in collaboration with Autoliv to automate the labeling process of side-view car windows using image segmentation

Before you get started

• It is recommended to use a Python virtual environment (venv) for libraries and packages

  • To create a venv with necessary libraries and packages, run the following file:
  • /src/installation_setup.py

• Create two empty folders in the root:

  • checkpoints
  • data

• Add into checkpoints/ folder the models' checkpoint provided OneDrive

• If you want to train the models (U-Net & YOLOv11) please download the images and labels from the GP22 dataset containing 1,480 pictures of cars with corresponding feature labels

Available notebooks

(If you want to train the models, you need to first run the preprocess notebook)

Preprocessing & Data Augmentation

The following notebook contains the necessary steps to prepare the dataset for training. Steps included: Removing background, Flipping cars, Scaling cars, Data augmentation

• preprocessing

Model training and prediction

• yolo_train_and_predict - train the YOLO model and/or predict with the trained model
• unet_train_and_predict - train the U-Net model and/or predict with the trained model

Convert to CAD Output

• output_in_cad- will convert predicted masks to CAD format (.DXF)

Model evaluation

The following notebooks are designed to evaluate the performance of both models using a set of 100 images and will find the best and worst prediction in the set.

• evaluate_yolo

• evaluate_unet

• Metrics used:

• Mean Hausdorff Distance Mean IoU Mean Dice Mean Precision Mean Recall Mean MAE Mean F1-Score