A novel data augmentation approach for ego-lane detection enhancement

Abstract

Utilizing vast annotated datasets for supervised training of deep learning models is an absolute necessity. The focus of this paper is to demonstrate a supervisory training technique using perspective transformation-based data augmentation to train various cutting-edge architectures for the ego-lane detection task. Creating a reliable dataset for training such models has been challenging due to the lack of efficient augmentation methods that can produce new annotated images without missing important features about the lane or the road. Based on extensive experiments for training the three architectures: SegNet, U-Net, and ResUNet++, we show that the perspective transformation data augmentation strategy noticeably improves the performance of the models. The model achieved validation dice of 0.991 when ResUNET++ was trained on data of size equal to 6000 using the PTA method and achieved a dice coefficient of 96.04% when had been tested on the KITTI Lane benchmark, which contains 95 images for different urban scenes, which exceeds the results of the other papers. An ensemble learning approach is also introduced while testing the models to achieve the most robust performance under various challenging conditions.