[论文解读] A Survey on Deep Learning for Localization and Mapping: Towards the Age of Spatial Machine Intelligence

Changhao Chen, Bing Wang+3 more|arXiv (Cornell University)|Jun 22, 2020

Robotics and Sensor-Based Localization参考文献 251被引用 106

一句话总结

This paper provides a comprehensive taxonomy and survey of deep learning approaches for localization and mapping, covering odometry, mapping, global localization, and SLAM, and discusses limitations and future directions toward Spatial Machine Intelligence Systems (SMIS).

ABSTRACT

Deep learning based localization and mapping has recently attracted significant attention. Instead of creating hand-designed algorithms through exploitation of physical models or geometric theories, deep learning based solutions provide an alternative to solve the problem in a data-driven way. Benefiting from ever-increasing volumes of data and computational power, these methods are fast evolving into a new area that offers accurate and robust systems to track motion and estimate scenes and their structure for real-world applications. In this work, we provide a comprehensive survey, and propose a new taxonomy for localization and mapping using deep learning. We also discuss the limitations of current models, and indicate possible future directions. A wide range of topics are covered, from learning odometry estimation, mapping, to global localization and simultaneous localization and mapping (SLAM). We revisit the problem of perceiving self-motion and scene understanding with on-board sensors, and show how to solve it by integrating these modules into a prospective spatial machine intelligence system (SMIS). It is our hope that this work can connect emerging works from robotics, computer vision and machine learning communities, and serve as a guide for future researchers to apply deep learning to tackle localization and mapping problems.

研究动机与目标

• 将定位与建图作为自治代理与空间 AI 基础的动机。
• 提出一个新的分类法，将深度学习方法在里程计、建图、全局定位和 SLAM 中进行组织。
• 调查端到端、混合和无监督学习方法及其在现实传感数据中的应用。
• 讨论局限性、挑战和未来方向，以引导机器人学、计算机视觉和机器学习等跨学科研究。

提出的方法

• 提出定位与建图的深度学习方法分类法（里程计、建图、全局定位、SLAM）。
• 回顾用于里程计估计和姿态/轨迹恢复的监督、无监督/自监督和混合学习框架。
• 讨论视觉、惯性和 LiDAR 传感模态以及 DL 方法如何与传统几何模型整合。
• 分析基于 DL 的 SLAM 后端的鲁棒性、泛化、尺度恢复和不确定性估计。
• 将 Spatial Machine Intelligence System (SMIS) 作为感知、定位与建图的整合蓝图进行介绍。

实验结果

研究问题

• RQ1 What DL-based methods exist for odometry estimation, mapping, global localization, and SLAM?
• RQ2 How do end-to-end, unsupervised/self-supervised, and hybrid DL approaches compare in accuracy and robustness to traditional model-based methods?
• RQ3 What are the limitations (generalization, interpretability, computation) of current DL-based localization and mapping systems?
• RQ4 How can DL-based perception modules be integrated into a unified Spatial Machine Intelligence System (SMIS) for long-term autonomy?

主要发现

• DL methods can learn pose transformations directly from data and can provide scale in monocular setups through learning.
• Unsupervised/self-supervised VO/VIO approaches use view synthesis and geometric consistency losses to recover depth and motion, addressing some scale and dynamics issues.
• Hybrid models that combine deep depth/pose predictions with classic geometric back-ends often outperform purely end-to-end or traditional VO/VIO systems.
• The DL-based SLAM ecosystem shows promise in robustness to sensor noise and dynamic environments, but generalization, interpretability, and high computational costs remain challenges.
• The survey frames DL localization/mapping within a Spatial Machine Intelligence System (SMIS), connecting robotics, computer vision, and ML communities.