LSPU: A Fully Integrated Real-Time LiDAR-SLAM SoC with Point-Neural-Network Segmentation and Multi-Level kNN Acceleration

📋 论文概要

本文提出了一款名为LSPU的全集成实时LiDAR-SLAM系统级芯片，解决了传统RGB视觉SLAM在自动驾驶中视场受限、深度感知不准确且易受环境影响的问题。该芯片通过点神经网络分割和多级kNN加速，实现了高精度、长距离360°点云地图构建与定位。

💡 主要创新点

• 首次将点神经网络分割与多级kNN加速集成于单芯片，实现实时LiDAR-SLAM处理。
• 通过硬件加速器优化点云处理流程，显著提升能效比和吞吐量。

🏷 关键词

📄 原文摘要

mobile robots require Simultaneous Localization and Mapping (SLAM) for autonomous driving and seamless interaction with the surrounding objects. Previous RGB-based visual SLAM processors [1-2] cannot be deployed for autonomous driving due to their restricted FoV, inaccurate depth perception, and vulnerability to environmental changes. In contrast, LiDAR-SLAM algorithm [3] provides precise depth information with long-range 360°-view point clouds to build fine details of the surroundings, making it the most essential component in autonomous robots. Figure 20.6.1 shows the overall computation flow of the proposed LiDAR-SLAM (LP-SLAM) system that comprises 3D spatial perception, odometry and mapping, where each stage consists of multiple algorithms. For each frame of LiDAR, a point neural network (PNN)

👥 作者与机构

Jueun Jung1, Seungbin Kim1, Bokyoung Seo1, Wuyoung Jang1, Sangho Lee1,

Jeongmin Shin1, Donghyeon Han2, Kyuho Jason Lee1 Ulsan National Institute of Science and Technology, Ulsan, Korea Massachusetts Institute of Technology, Cambridge, MA 1 2 Emerging