Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Road Segmentation for ADAS/AD Applications

Mathanesh Vellingiri Ramasamy, Dimas Rizky Kurniasalim

TL;DR

The paper tackles robust road segmentation for ADAS/AD by comparing a modified VGG-16 trained on Comma10k with a modified U-Net trained on KITTI Road, and then performing cross-dataset validation to assess generalization. The results show that the VGG-16 variant generalizes better across unseen data, while the U-Net struggles despite significantly more training, highlighting the impact of pretraining and dataset characteristics. Analysis points to factors such as dataset size, environment (California highways vs Karlsruhe suburbs), tree shadows, pavement similarity, and annotation conventions as key drivers of cross-domain performance. The work underscores the need for diverse, well-annotated datasets and appropriate transfer learning and augmentation strategies to improve road segmentation performance in real-world ADAS/AD deployments.

Abstract

Accurate road segmentation is essential for autonomous driving and ADAS, enabling effective navigation in complex environments. This study examines how model architecture and dataset choice affect segmentation by training a modified VGG-16 on the Comma10k dataset and a modified U-Net on the KITTI Road dataset. Both models achieved high accuracy, with cross-dataset testing showing VGG-16 outperforming U-Net despite U-Net being trained for more epochs. We analyze model performance using metrics such as F1-score, mean intersection over union, and precision, discussing how architecture and dataset impact results.

Road Segmentation for ADAS/AD Applications

TL;DR

The paper tackles robust road segmentation for ADAS/AD by comparing a modified VGG-16 trained on Comma10k with a modified U-Net trained on KITTI Road, and then performing cross-dataset validation to assess generalization. The results show that the VGG-16 variant generalizes better across unseen data, while the U-Net struggles despite significantly more training, highlighting the impact of pretraining and dataset characteristics. Analysis points to factors such as dataset size, environment (California highways vs Karlsruhe suburbs), tree shadows, pavement similarity, and annotation conventions as key drivers of cross-domain performance. The work underscores the need for diverse, well-annotated datasets and appropriate transfer learning and augmentation strategies to improve road segmentation performance in real-world ADAS/AD deployments.

Abstract

Accurate road segmentation is essential for autonomous driving and ADAS, enabling effective navigation in complex environments. This study examines how model architecture and dataset choice affect segmentation by training a modified VGG-16 on the Comma10k dataset and a modified U-Net on the KITTI Road dataset. Both models achieved high accuracy, with cross-dataset testing showing VGG-16 outperforming U-Net despite U-Net being trained for more epochs. We analyze model performance using metrics such as F1-score, mean intersection over union, and precision, discussing how architecture and dataset impact results.
Paper Structure (22 sections, 11 figures)

This paper contains 22 sections, 11 figures.

Table of Contents

  1. Introduction
  2. Background Theory
  3. Semantic Segmentation
  4. Encoder-Decoder Networks
  5. VGG-16 and U-Net
  6. Transfer Learning
  7. Loss Function and Optimizer
  8. Evaluation Metrics
  9. Pixel Accuracy
  10. Intersection over Union
  11. Precision and Recall
  12. F1 Score
  13. Methodology
  14. Cross-Validation and Discussion
  15. Dataset Size
  16. ...and 7 more sections

Figures (11)

  • Figure 1: An example of image segmentation with multiple classes, each class represented by their respective colors.
  • Figure 2: The VGG-16 backbone model used without modification. Courtesy of Very Deep Convolutional Networks for Large-Scale Image Recognition by Karen Simonyan et al.
  • Figure 3: The U-Net backbone model used without modification. Courtesy of U-Net: Convolutional Networks for Biomedical Image Segmentation by Olaf Ronneberger et al.
  • Figure 4: An example of image segmentation from the Comma10k dataset. The black areas represent values of 0 while the white areas represent 1.
  • Figure 5: Training and evaluation data after training the modified VGG-16 model on the Comma10k dataset for 8 epochs.
  • ...and 6 more figures