Timorese Academic Journal of Science and Technology
ISSN : 2617 - 4944 (Print) ISSN : 2617 - 4952 (Online)
Implementation of U-Net Deep Learning Framework for Road and Road Line Segmentation
Author(s):
Vosco Pereira,
Frederico S. Cabral,
Lourenço A. L. Pereira,
Mariano R. M. da Cruz,
Hidekazu Fukai,
Full Paper Access: Donwload
Abstract: Monitoring the road condition in Timor-Leste continues to be considered a challenge, particularly to detect road lines. The road lines are essential for drivers in traveling as they can regulate the roadmap and minimize accidents. The impact of the inexistence of road lines for vehicles may result in an accident. In this research, we used U-Net deep learning framework for segmenting the road surface and road lines based on road images. Different types of road videos were collected using a camera followed by applying several image processing techniques before conducting the segmentation. The experiment result shows U-Net achieve 99.5% of accuracy and 93.7 for the mean intersection over union (mIoU), where the IoU of each class was 99.5 for background, 96.7 for the road surface, and 85.0 for the road lines, respectively.
Keywords: road, road line, deep learning, semantic segmentation, u-net