A Cylindrical Convolution Network for Dense Top-View Semantic Segmentation with LiDAR Point Clouds

Accurate semantic scene understanding of the surrounding environment is a challenge for autonomous driving systems. Recent LiDAR-based semantic segmentation methods mainly focus on predicting point-wise semantic classes, which cannot be directly used before the further densification process. In this paper, we propose a cylindrical convolution network for dense semantic understanding in the top-view LiDAR data representation. 3D LiDAR point clouds are divided into cylindrical partitions before feeding to the network, where semantic segmentation is conducted in the cylindrical representation. Then a cylinder-to-BEV transformation module is introduced to obtain sparse semantic feature maps in the top view. In the end, we propose a modified encoder-decoder network to get the dense semantic estimations. Experimental results on the SemanticKITTI and nuScenes-LidarSeg datasets show that ourmethod outperforms the state-of-the-artmethods with a large margin.

Automated summary

In this paper, a cylindrical convolution network is proposed for dense semantic understanding in the top-view LiDAR data representation in autonomous driving systems. The method involves dividing 3D LiDAR point clouds into cylindrical partitions and conducting semantic segmentation in the cylindrical representation. A cylinder-to-BEV transformation module is introduced to obtain sparse semantic feature maps in the top view. Finally, a modified encoder-decoder network is used to achieve dense semantic estimations.