SANet:空间注意力机制下的LiDAR点云实时语义分割方法  

SANet: real time semantic segmentation method of LiDAR point cloud based on spatial attention mechanism

作　　者：王玮琦 游雄[1] 苏明占 张蓝天 周雪莹 赵耀 WANG Weiqi;YOU Xiong;SU Mingzhan;ZHANG Lantian;ZHOU Xueying;ZHAO Yao(Information Engineering University,Zhengzhou 450052,China;Beijing Institute of Remote Sensing Information,Beijing 100192,China;Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]信息工程大学,河南郑州450052 [2]北京市遥感信息研究所,北京100192 [3]北京理工大学,北京100081

出　　处：《测绘通报》2022年第11期32-38,共7页Bulletin of Surveying and Mapping

基　　金：国家自然科学基金(42130112,42171456);国家自然科学基金青年基金(41801317);中原学者科学家工作室资助项目;国家重点研发计划(2017YFB0503500)。

摘　　要：语义分割是智能机器人由感知智能迈向认知智能的重要基础,当前针对点云数据的语义分割方法存在实时性差、精度低等现象。本文系统分析了点云经球面投影所得的距离图像与自然图像的差异,为基于距离图像的实时语义分割网络设计提供了思路。通过分析发现,距离图像具有强空间相关性的特点,将强空间相关性与注意力机制相结合,提出基于空间注意力机制下的LiDAR点云实时语义分割方法SANet。该方法能够高效地聚合空间分布特征与上下文特征,且模型参数量较少,满足实时性的要求。在SemanticKITTI数据集上的试验表明,与其他优秀算法相比,SANet兼顾了实时性与准确性,显著提高了LiDAR点云语义分割的精度,可为自动驾驶及其他机器人应用领域提供辅助支撑。Semantic segmentation is an important basis for intelligent robots to move from perceptual intelligence to cognitive intelligence. The current semantic segmentation methods for point cloud data have poor real-time performance and low accuracy. In this article, we systematically analyze the difference between the range images generated by spherical projection of point cloud and common images, and provide ideas for the design of real-time semantic segmentation neural network. Through the analysis, we find that the range images have the characteristics of strong spatial correlation. This article combines the strong spatial correlation with attention mechanism, then proposes a real-time semantic segmentation method SANet based on spatial attention mechanism. SANet can efficiently aggregate spatial distribution features and context features. And the model parameters are less, which can meet the real-time requirements. Experiments on the SemanticKITTI dataset show that SANet has both good real-time performance and high accuracy compared with other excellent algorithms. The spatial attention mechanism proposed in this article significantly improves the accuracy of semantic segmentation of LiDAR point cloud by efficiently aggregating spatial distribution features and context features, which can provide auxiliary support for autonomous driving and other robot applications.

关 键 词：空间注意力 点云语义分割 SemanticKITTI 距离图像 

分 类 号：P237[天文地球—摄影测量与遥感]