Radial Transformer for Large-Scale Outdoor LiDAR Point Cloud Semantic Segmentation

Semantic segmentation of large-scale outdoor point cloud captured by light detection and ranging (LiDAR) sensors can provide fine-grain and stereoscopic comprehension for the surrounding environment. However, limited by the receptive field of convolution kernel and ignoration of specific spatial properties inherent to the large-scale outdoor point cloud, the existing advanced LiDAR semantic segmentation methods inevitably abandon the unique radial long-range topological relationships. To this end, from the LiDAR perspective, we propose a novel Radial Transformer that can naturally and efficiently exploit the radial long-range dependencies exclusive to the outdoor point cloud for accurate LiDAR semantic segmentation. Specifically, we first develop a radial window partition to generate a series of candidate point sequences and then construct the long-range interactions among the densely continuous point sequences by the self-attention mechanism. Moreover, considering the varying-distance distribution of point cloud in 3-D space, a spatial-adaptive position encoding is particularly designed to elaborate the relative position. Furthermore, we fusion radial balanced attention for a better structure representation of real-world scenes and distant points. Extensive experiments demonstrate the effectiveness and superiority of our method, which achieves 67.5% and 77.7% mean intersection-over-union (mIoU) on two recognized large-scale outdoor LiDAR point cloud datasets SemanticKITTI and nuScenes, respectively.

I. Introduction

Light detection and ranging (LiDAR) sensor can provide point cloud data to accurately describe outdoor scene structures [1], and is the utmost important data source for environmental perception in various intelligent domains, such as intelligent robotics, autonomous driving and remote sensing [2], [3], [4]. Based on the raw LiDAR data, point cloud semantic segmentation can predict semantic category label for each point to achieve fine-grained and stereoscopic perception [5]. The outdoor LiDAR point cloud is typical large-scale, and its accurate semantic-level perception results can offer the fundamental guarantee for various downstream tasks such as path planning and behavior decision-making. Thus, large-scale outdoor LiDAR point cloud semantic segmentation serves as the essentially foundational tasks for modern intelligent domains [6], [7].