基于多尺度注意力机制的实时激光雷达点云语义的分割  

Semantic segmentation of real-time LiDAR point clouds based on multi-scale self-attention

作　　者：张晨 刘畅 赵津[1] 王广玮[1,2] 许庆 ZHANG Chen;LIU Chang;ZHAO Jin;WANG Guangwei;XU Qing(School of Mechanical Engineering,Guizhou University,Guiyang 550025,China;School of Vehicle and Mobility,Tsinghua University,Beijing 100084,China)

机构地区：[1]贵州大学机械工程学院,贵阳550025 [2]清华大学车辆与运载学院,北京100084

出　　处：《汽车安全与节能学报》2024年第4期591-601,共11页Journal of Automotive Safety and Energy

基　　金：国家自然科学基金地区项目(52265070);贵州省科技支撑计划项目(黔科合支撑[2022]一般045);贵州省创新人才团队项目(CXTD2022-009)。

摘　　要：为既能提高分割精度,又能克服车载计算资源局限,提出一种面向移动机器人平台的车载实时点云语义分割方法,并进行了综合实验。该方法采用基于投影的激光雷达语义分割方法,将三维点云投影到球面图像,并结合二维卷积进行分割。引入多头注意力机制(MHSA),实现轻量级语义分割模型,以一种全新的方式,将一种深度学习模型架构Transformer映射到卷积。将Transformer的MHSA迁移至卷积,以形成多尺度自注意力机制(MSSA)。结果表明:与当前主流方法CENet、FIDNet、PolarNet相比,本方法在NVIDIA JETSON AGX Xavier计算平台上保持了较高的分割精度(平均交并比为63.9%)及较高的检测速率(41帧/s),从而证明了其对移动机器人平台的适用性。A real-time point cloud semantic segmentation method was proposed for mobile robot platforms through digital experiments,to enhance segmentation accuracy within the constraints of in-vehicle computing resources.The approach used a projection-based LiDAR technique,projecting the 3-D point cloud onto a spherical image and applying 2-D convolution.The approach integrated the multi-head self-attention(MHSA)mechanism,adapting the Transformer,a software semantic segmentation,architecture into convolution operations to build a multi-scale self-attention(MSSA)framework.The results show that on the NVIDIA JETSON AGX Xavier computing platform,the proposed method achieves a high segmentation accuracy with the mean ratio of Intersection to Union(mIoU)being 63.9%,and a fast detection speed of 41 frame/s,compared to stateof-the-art methods like the CENet,the FIDNet,and the PolarNet,therefore,demonstrating the effectiveness of the mobile robot platforms.

关 键 词：移动机器人平台 激光雷达(LiDAR) 点云 多尺度注意力机制(MSSA) 语义分割方法TRANSFORMER 卷积神经网络 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]