Akademik Veri Yönetim Sistemi
INTERNATIONAL SYMPOSIUM ON APPLIED SCIENCES AND ENGINEERING, 26 - 28 Kasım 2018, ss.150-153
Lane detection plays a vital role in self-driving
cars and advanced driver assistance systems as it helps cars to
position themselves within the road and avoid collision. A
machine learning technique known as Transfer Learning was
used in this study. TL is the reuse of a previously developed
model for a task as the starting point for a new one with no need
for a big dataset. VGG-16, a 16 layers convolutional neural
network trained on more than million images for classifying
images into 1000 object categories, was used in this study. This
model was selected as the starting point to perform semantic
segmentation which is evaluating each pixel in the image and
categorizing it into a specific class. The Cambridge DrivingLabeled Video Database (CamVid) which provides ground truth
labels as each pixel associated to one of the 32 semantic classes
was used in the study for both training and test purposes
changing the classes from 32 to 2 as Lanes and Others. Based on
the experimental results the lane detection accuracy was 97.42%
under various illumination conditions and presence of shadows
on the road. The experimental results demonstrated the
effectiveness of the proposed method in reducing learning time
and improving performance, making it suitable for real-time
applications.