Attention-based deep learning framework for semantic segmentation of 3D point clouds for autonomous driving

Abstract

Perception is an important yet challenging task for self-driving cars. Particularly, the semantic segmentation of LiDAR point cloud can yield rich semantic information about the outdoor environment. However, current methods dealing with LiDAR point clouds are not efficient enough in terms of computational cost or accuracy. In this dissertation, we propose an end-to-end network for efficient semantic segmentation of outdoor point clouds. Our contributions include a novel data augmentation technique, a lightweight 2D network and a newly designed weighted loss function. The data augmentation technique interpolates more object instances randomly, thus providing a balanced dataset for training. Then, our ADLA-Net, comprising the Deep Layer Aggregation (DLA) and the Axial-Attention, is employed on 2D range-view inputs to predict class labels. The Weighted Asymmetric Loss (WASL) adds a class frequency-based penalty to alleviate the persisting class imbalance issue. We evaluate our framework on the NuScenes dataset against mainstream methods. Experiments demonstrate that our method can produce better results even at a low computational cost. Overall, it outperforms existing methods by a margin of \(3.2\%p\) in terms of mIoU.