引言

本博文对基于learning的lidar odometry（包括lidar，lidar+IMU等）进行调研，并对一些经典的工作进行阅读。

本博文仅供本人学习记录用~

• 引言
• Paper List
• 经典工作介绍
  • LONet: deep real-time LiDAR odometry
  • DMLO: Deep Matching LiDAR Odometry
  • Self-supervised Learning of LiDAR Odometry for Robotic Applications
  • Efficient 3D Deep LiDAR Odometry
  • Translo: A window-based masked point transformer framework for large-scale lidar odometry
  • DiffLO: Semantic-Aware LiDAR Odometry with Diffusion-Based Refinement
  • Pointconv: Deep convolutional networks on 3d point clouds
  • DELO: Deep Evidential LiDAR Odometry using Partial Optimal Transpor
  • PWCLO-Net: Deep LiDAR Odometry in 3D Point Clouds Using Hierarchical Embedding Mask Optimization
  • Pointnet++: Deep hierarchical feature learning on point sets in a metric space
  • A Consistency-Aware Spot-Guided Transformer for Versatile and Hierarchical Point Cloud Registration
  • Kpconv: Flexible and deformable convolution for point clouds

Paper List

Year	Venue	Paper Title	Repository	Note
2025	IEEE Sensor Journal	A Generative Hierarchical Optimization Framework for LiDAR Odometry Using Conditional Diffusion Models	—	—
2025	CVPR	DiffLO: Semantic-Aware LiDAR Odometry with Diffusion-Based Refinement		—
2025	arXiv	LIR-LIVO: A Lightweight, Robust LiDAR/Vision/Inertial Odometry with Illumination-Resilient Deep Features		Fast-LIVO+AirSLAM
2024	arXiv	LiDAR Inertial Odometry And Mapping Using Learned Registration-Relevant Features		—
2024	IEEE 100th Vehicular Technology Conference	LiDAR-OdomNet: LiDAR Odometry Network Using Feature Fusion Based on Attention	—	—
2024	IROS	LiDAR-Visual-Inertial Tightly-coupled Odometry with Adaptive Learnable Fusion Weights	—	F-LOAM+VINS-Mono
2024	TRO	PIN-SLAM: LiDAR SLAM Using a Point-Based Implicit Neural Representation for Achieving Global Map Consistency		—
2024	AAAI	DeepPointMap: Advancing LiDAR SLAM with Unified Neural Descriptors		—
2023	ICCV	NeRF-LOAM: Neural Implicit Representation for Large-Scale Incremental LiDAR Odometry and Mapping		—
2023	ICCV	DELO: Deep Evidential LiDAR Odometry using Partial Optimal Transport		website
2023	AAAI	Translo: A window-based masked point transformer framework for large-scale lidar odometry		—
2023	RAL	LONER: LiDAR Only Neural Representations for Real-Time SLAM		—
2023	TIV	HPPLO-Net: Unsupervised LiDAR Odometry Using a Hierarchical Point-to-Plane Solver		—
2022	TPAMI	Efficient 3D Deep LiDAR Odometry		—
2021	ICRA	Self-supervised learning of lidar odometry for robotic applications		—
2021	CVPR	PWCLO-Net: Deep LiDAR Odometry in 3D Point Clouds Using Hierarchical Embedding Mask Optimization		—
2021	ISPRS	Deeplio: deep lidar inertial sensor fusion for odometry estimation		—
2020	ACM international conference on multimedia	Lodonet: A deep neural network with 2d keypoint matching for 3d lidar odometry estimation	—	—
2020	ICRA	Unsupervised geometry-aware deep lidar odometry	—	website
2020	IROS	DMLO: Deep Matching LiDAR Odometry	—	—
2019	IROS	Deeppco: End-to-end point cloud odometry through deep parallel neural network	—	—
2019	CVPR	Lo-net: Deep real-time lidar odometry	—	—
2019	CVPR	L3-net: Towards learning based lidar localization for autonomous driving	—	—
2018	IEEE International Conference on Autonomous Robot Systems and Competitions	CNN for IMU assisted odometry estimation using velodyne LiDAR	—	—
2016	RSS workshop	Deep learning for laser based odometry estimation	—	—

• 其他有代表性的基于learning的lidar工作或者point cloud registration：

Year	Venue	Paper Title	Repository	Note
2024	arXiv	A Consistency-Aware Spot-Guided Transformer for Versatile and Hierarchical Point Cloud Registration		—
2020	CVPR	3dregnet: A deep neural network for 3d point registration		—
2020	CVPR	P2b: Point-to-box network for 3d object tracking in point clouds	—	—
2019	CVPR	Pointconv: Deep convolutional networks on 3d point clouds	—	—
2019	ICCV	Meteornet: Deep learning on dynamic 3d point cloud sequences	—	—
2019	ICCV	Deep closest point: Learning representations for point cloud registration	—	—
2019	CVPR	Pointnetlk: Robust & efficient point cloud registration using pointnet		—
2019	CVPR	3D local features for direct pairwise registration	—	—
2019	ICCV	Deepvcp: An end-to-end deep neural network for point cloud registration	—	—
2019	CVPR	Flownet3d: Learning scene flow in 3d point clouds	—	—
2019	IROS	Rangenet++: Fast and accurate lidar semantic segmentation	—	—
2019	ACM Transactions on Graphics	Dynamic graph cnn for learning on point clouds	—	—
2018	ECCV	3dfeat-net: Weakly supervised local 3d features for point cloud registration		—
2017	NIPS	Pointnet++: Deep hierarchical feature learning on point sets in a metric space	—	—
2017	CVPR	Pointnet: Deep learning on point sets for 3d classification and segmentation	—	—
2015	IROS	Voxnet: A 3d convolutional neural network for real-time object recognition	—	—

经典工作介绍

对于基于learning的lidar odometry主要有以下几个挑战：

1. 两帧离散的lidar scans如何建立准确的数据关联
2. 由于遮挡或者lidar的分辨率限制而导致的，属于同一个物体的点云，在两帧中不一致
3. 动态点云
4. 直接从原始3D点云学数据通常是低效的（由于点云的不规则及无序性），也就是如何获取更好的representation learning

LONet: deep real-time LiDAR odometry

LONet是首个基于learning的lidar odometry，依赖于CNN的拟合能力。输入两个lidar scans，直接输出两者的relative motion。

网络end-to-end训练，没有任何的几何约束，容易出现overfitting的情况.

DMLO: Deep Matching LiDAR Odometry

DMLO在框架中明确强制执行几何约束,将6DoF姿态估计分为两个部分：

1. learning-based matching network给两个lidar扫描提供精确的correspondence
2. 通过近似的奇异值分解（SVD）来估算刚体变换。

将所有的lidar信息编码成2D的图像（如下图所示），进而可以通过CNN来提取feature以及局部的相似性，进而计算出scans之间的数据关联。

而对于所获得的correspondence，也会额外估算其对应的权重，而对于获得的这些点云对（ matched pairs）再通过SVD来计算姿态变换。直观来看，有点像基于learning的point cloud registration+SVD进而实现lidar-based odometry.

Self-supervised Learning of LiDAR Odometry for Robotic Applications

关键点：

• no labeled or ground-truth data is required
• applicable to a wide range of environments and sensor modalities without requiring any network or loss function adjustments
• 6-DOF pose, being able to operate in real-time on a mobile-CPU.

对于lidar的点云，一般有三种处理的方式：

1. 将点云投影到2D image，然后用基于image的架构处理（比如Rangenet++）
2. 使用基于voxel的3D卷积（比如Voxnet，high memory-requirement）
3. 直接作用在disordered point cloud scans（比如Pointnet）

而本文采用的是第一种。

通过KD-Tree寻找到所有的点云的 source 和 target的对应，构建point-to-plane 和 plane-to-plane 的loss.

Efficient 3D Deep LiDAR Odometry

首先提出了一种projection-aware representation of the 3D point cloud，然后提出了一个Pyramid, Warping, and Cost volume (PWC) 架构，而关于点云之间的关联则是采用projection-aware attentive cost volume 针对点云表达、数据关联、（动态点云）、如何提取有效的信息这四个问题，分别提出对应的模块针对处理。

Translo: A window-based masked point transformer framework for large-scale lidar odometry

这个工作则是基于transformer的lidar odometry（the first transformer-based LiDAR odometry network） 而点云则是也是投影到2D 图像平面来处理的project point clouds onto a cylindrical surface to get pseudo images

DiffLO: Semantic-Aware LiDAR Odometry with Diffusion-Based Refinement

直观理解为将MLP或者ICP对coarse initial pose refine的过程用Diffusion来做 （the first diffusion-based LiDAR odometry network）

点云的特征提取则是采用PointConv。而对于图中的语义感知模块，在推理的时候都需要retraining

Pointconv: Deep convolutional networks on 3d point clouds

PointConv将点云的位置(xyz)作为输入，用MLP来学习权重函数，并对学习到的权重采用inverse density scale来补偿非均匀采样。 可以看成是2D图像的卷积扩展到3D点云，采用MLP来实现density re-weighted convolution同时通过memory efficient的实现让其可以易于拓展。

• treat convolution kernels as nonlinear functions of the local coordinates of 3D points comprised of weight and density functions.
• perform convolution on 3D point clouds with non-uniform sampling
• PointConv involves taking the positions of point clouds as input and learning an MLP to approximate a weight function, as well as applying a inverse density scale on the learned weights to compensate the non-uniform sampling.

对于一张2D图像，其可以展开为2D的离散网格阵列，对应的卷积可以看成如下表达：

对于每个CNN的filter都是一个固定的小区域（比如3x3或者5x5等）

而对于点云数据，其是一系列3D点，每个点包含了xyz的位置向量以及对应的特征（比如颜色、表面法线等）。 相比起2D图像而言，点云的形状更加的灵活，其不在是在固定的网格中的点，而是可以是任意的连续值。因此传统的离散卷积将不可以直接用于点云上。而本文所提出的PointConv则是回归到3D卷积的连续版本上：

而点云可以看成是连续3D空间的非均匀采样：

而本文的关键点其实就是用MLP来实现上式中的权重函数W与3D坐标的关联。而对上式中的逆密度函数S则是通过a kernelized density estimation+MLP来近似。 进一步原理如下图3所示：

而PointConv中MLP的权重则是在多有的点云间share weight的。

DELO: Deep Evidential LiDAR Odometry using Partial Optimal Transpor

则是采用将点云降采样为固定的数量的点，然后用graph cnn来编码获取特征。然后也用transformer进行数据关联，然后再通过一个网络来估算变换以及用GTSAM来优化姿态：

PWCLO-Net: Deep LiDAR Odometry in 3D Point Clouds Using Hierarchical Embedding Mask Optimization

这个工作是直接作用在点云上的(处理原始点云的思路参考Pointnet++).通过计算两帧点云的a weighted soft correspondence

通过一个 internal trainable embedding mask来滤除遮挡点云或者动态点云。

Pointnet++: Deep hierarchical feature learning on point sets in a metric space

在PointNet的基础上加了一个hierarchical structure

而PointNet的架构则是如下：

A Consistency-Aware Spot-Guided Transformer for Versatile and Hierarchical Point Cloud Registration

此处点云的预处理就是用的Kpconv，然后通过transformer以及coarse-to-fine matching network来计算两帧点云的相关性（如变换矩阵）

Kpconv: Flexible and deformable convolution for point clouds

收到image-based CNN的启发，对于一系列的kernel point定义一个kernel weight区域。如下图所示

相当于kernel weight是由点来携带的，类似于input features，而其影响的区域则是由一个相关的函数来定义。

下图是CNN与Kpconv的对比：