For autonomous vehicles, knowing where they can drive and where they cannot is of utmost importance. The drivable area detection module serves precisely this purpose, ensuring that the vehicle operates within areas where it is permissible to drive while avoiding non-drivable regions. Recent efforts in deep neural networks (DNNs) have significantly improved drivable area detection performance for autonomous driving. Nevertheless, the majority of DNN-based approaches require a substantial volume of data to train their models. Acquiring extensive datasets with manually annotated ground truth can be an expensive, laborious, and time-consuming process, often necessitating the involvement of domain experts. As a result, the practical implementation of DNN-based methods in realworld applicatio...[ Read more ]

For autonomous vehicles, knowing where they can drive and where they cannot is of utmost importance. The drivable area detection module serves precisely this purpose, ensuring that the vehicle operates within areas where it is permissible to drive while avoiding non-drivable regions. Recent efforts in deep neural networks (DNNs) have significantly improved drivable area detection performance for autonomous driving. Nevertheless, the majority of DNN-based approaches require a substantial volume of data to train their models. Acquiring extensive datasets with manually annotated ground truth can be an expensive, laborious, and time-consuming process, often necessitating the involvement of domain experts. As a result, the practical implementation of DNN-based methods in realworld applications becomes challenging. In order to alleviate the challenges associated with annotating training data, this thesis introduces a module called the Automatic Data Labeler (ADL). The ADL module enables the automated generation of training data by combining RGB images and depth information from LiDAR ¹ sensors. Additionally, considering the inherent disparity between the automatically generated training data and the ground truth (manually annotated training data), we incorporate uncertainty to bridge this gap. Finally, we evaluate the proposed ADL module on the KITTI [3] and KITTI-CARLA [4] datasets, and experimental results demonstrate that our approach achieves the best performance.