Source-linked AI summary
LO-Net: Deep Real-time Lidar Odometry
Qing Li, Shaoyang Chen, Cheng Wang, Xin Li, Chenglu Wen, Ming Cheng, Jonathan Li
TL;DR
Lidar odometry needs accurate, robust, real-time motion estimation despite sparse point clouds and dynamic scenes, while prior pipelines often rely on separately designed stages. LO-Net learns scan-to-scan motion, normals, and dynamic-region masks with geometric consistency, then uses a mapping module; benchmark experiments report state-of-the-art performance comparable to geometry-based methods. The authors identify matrix encoding and dependence on ground-truth training data as remaining challenges.
Problem
Accurate real-time lidar odometry remains difficult because sparse point clouds and dynamic objects challenge conventional registration and prior deep-learning approaches.
Method
LO-Net jointly learns 6-DoF relative pose, point-wise normals, and moving-object masks from consecutive scans, using geometric consistency and an efficient mapping module.
Results
LO-Net achieves state-of-the-art benchmark results and is comparable to state-of-the-art geometry-based lidar odometry techniques.
Takeaways & Limitations
Learned normal and mask information can support more accurate real-time odometry when coupled with mapping.
Takeaways & Limitations
The network uses matrix-encoded point clouds and requires ground-truth training data, limiting practical processing and application scenarios.
Abstract
from arXiv · showhide
We present a novel deep convolutional network pipeline, LO-Net, for real-time lidar odometry estimation. Unlike most existing lidar odometry (LO) estimations that go through individually designed feature selection, feature matching, and pose estimation pipeline, LO-Net can be trained in an end-to-end manner. With a new mask-weighted geometric constraint loss, LO-Net can effectively learn feature representation for LO estimation, and can implicitly exploit the sequential dependencies and dynamics in the data. We also design a scan-to-map module, which uses the geometric and semantic information learned in LO-Net, to improve the estimation accuracy. Experiments on benchmark datasets demonstrate that LO-Net outperforms existing learning based approaches and has similar accuracy with the state-of-the-art geometry-based approach, LOAM.
1. Introduction
LO-Net targets real-time lidar odometry by replacing separately designed estimation stages with an end-to-end network that learns motion-relevant geometric and dynamic-scene information. Its learned outputs are combined with mapping to improve accuracy while retaining real-time operation.
- Lidar odometry provides 3D position and orientation information for navigation in robotics and autonomous driving.
- Classic lidar registration methods struggle with sparse, nonuniform point clouds, while feature-based methods are computationally expensive and sensitive to dynamic objects.
- LO-Net proposes a deep neural network for lidar odometry that incorporates pairwise scans, normal consistency, and an uncertainty mask for static-region selection.
- The network simultaneously estimates normals and dynamic-region masks, applies a spatiotemporal geometry consistency constraint, and couples an efficient mapping module to the pipeline.
- Experiments on KITTI and the Ford Campus dataset report state-of-the-art results and comparability with geometry-based lidar odometry techniques.
2. Related work
Prior lidar odometry work includes ICP variants, geometric feature registration, and emerging deep-learning methods. These approaches trade accuracy, robustness, computational cost, and generality across sparse or dynamic lidar data.
- ICP and its variants estimate transformations by matching adjacent scans, but can be computationally expensive and sensitive to initial poses.
- GICP and NICP extend ICP with local-surface covariance, normal, and curvature information to improve accuracy, robustness, or performance.
- Plane- and feature-based methods address nonuniform laser sampling by detecting planes or robust line and plane keypoints for registration.
- CLS improves registration over GICP by grouping points into polar bins and forming line segments, but its line-segment computation is not real-time.
- LOAM combines line and plane features with parallel scan-to-scan and mapping modules to achieve low-drift, real-time odometry, without modeling dynamic objects.
- Deep learning has succeeded in vision and point-cloud tasks, but applying CNNs to 3D lidar odometry remains challenging; prior work transforms scans into dense matrices.
3. Method
LO-Net estimates lidar motion end-to-end from consecutive encoded scans while jointly learning normals and dynamic-object masks. A scan-to-map module then refines odometry by registering scans to a global map using the learned geometric and semantic information.
- LO-Net takes two consecutive scans and jointly predicts their 6-DoF relative pose, point-wise normals, and moving-object masks.The normal, mask, and Siamese pose-regression modules are jointly trained but can be used independently during inference.
- 3.1. Lidar data encoding: Cylindrical projection converts sparse, irregular lidar points into H × W × C matrices containing intensity and range channels.When multiple points project to one position, the closer point is retained.
- 3.2. Geometric consistency constraint: Normals are estimated efficiently from weighted cross products over four neighboring points and then smoothed with a moving average filter.This replaces covariance-matrix eigenvector computation used by PCA, which the authors describe as inefficient for embedded network training and estimation.
- 3.2. Geometric consistency constraint: The geometric consistency constraint transforms corresponding point elements between consecutive scans and compares their normals to regularize pose estimation.Normal consistency is used because normals reflect smooth surfaces and edge structures, while lidar coordinate and intensity measurements contain noise.
- 3.3. Lidar odometry regression: A two-stream network concatenates features from the input scans and uses convolutional and fully connected layers to regress the relative 6-DoF pose.FireConv layers reduce model parameters and computation, while width-only downsampling reflects the shape of intermediate features.
- 3.3. Lidar odometry regression: Translation and rotation losses use learnable scale parameters instead of a manually tuned rotational weight, with quaternion rotation constrained to a unit hemisphere.The initial training values are sx = 0.0 and sq = −2.5 for all scenes.
- 3.4. Mask prediction: A mask-weighted geometric loss addresses dynamic objects, whose points can otherwise inhibit odometry learning.The mask has no ground-truth supervision, so a cross-entropy regularizer prevents the trivial all-zero prediction.
- 3.5. Mapping: scan-to-map refinement: The mapping module registers each current scan against a global map, selecting smooth planar points with learned normals while excluding masked moving-object points.The scan matrix supplied to mapping contains intensity, range, normal, and mask values, and LO-Net odometry provides the initial pose.
4. Experiments
Experiments evaluate LO-Net across KITTI and Ford datasets, comparing odometry, normal estimation, and mask behavior against established baselines. LO-Net+Mapping achieves strong accuracy, improves trajectories, and supports real-time operation.
- Trajectory visualization: LO-Net+Mapping produces accurate closed-loop trajectories on Ford and the most accurate trajectory among shown methods on KITTI Seq. 08.The Ford trajectory forms closed loops accurately without enforcing loop closure.
- Odometry results: The mapping module reduces the vertical drift generated by LO-Net in KITTI Seq. 10 trajectories.The module refines odometry by matching the current scan against a global map.
- Experimental setup: The network is trained on KITTI and directly tested on Ford without further training or fine-tuning.KITTI sequences 00–10 provide ground-truth poses, while Ford contains loop-closure sequences with more moving vehicles than KITTI.
- Odometry results: LO-Net+Mapping achieves slightly better results than LOAM and clearly outperforms the other evaluated methods on KITTI and Ford.The evaluation uses average translational RMSE (%) and rotational RMSE (°/100m) over path lengths of 100m–800m.
- Ablation study: Training with the full loss yields the best average translational and rotational RMSE in the ablation study.The study compares different combinations of the proposed loss components on KITTI.
- Normal results: The predicted normals outperform comparison methods under most metrics and provide a reliable property for geometric consistency.Normal evaluation uses mean and median angular error and percent-good-normals thresholds; PCA normals with r = 0.5m serve as ground truth.
- Mask visualization: The learned mask identifies dynamic objects and vegetation as less explainable regions for odometry regression.The mapping and odometry pipeline uses mask information to exclude problematic points, while the implementation supports real-time GPU processing.
5. Conclusions
LO-Net is a learning framework for lidar odometry with an efficient mapping module that improves performance. The authors identify direct point-cloud processing and dependence on ground-truth training data as remaining challenges.
- LO-Net provides a novel learning framework for lidar odometry estimation, coupled with an efficient mapping module.The mapping module is included to further improve performance.
- Experiments on public benchmarks demonstrate the effectiveness of the framework over existing approaches.
- Directly processing 3D point clouds could be more practical than encoding them into data matrices for network input.
- Training the current network with ground-truth data limits its application scenarios.
- Future work includes investigating learned geometry representations and incorporating recurrent units to build temporal-related features.The authors suggest this may support an end-to-end framework without costly ground-truth data collection.
A. Network parameters
The network uses lightweight SqueezeNet-based modules for mask prediction and odometry regression. It predicts a point-wise mask and regresses the relative 6-DoF pose as translation and rotation quaternion outputs.
- Network parameters: The mask prediction and odometry regression layers use fireConv and fireDeconv modules to reduce parameters and computation cost.This design targets execution on low-performance platforms such as mobile robots or backpack systems.
- Network parameters: During feature extraction, maxpooling downsamples only the width because intermediate features are much wider than they are tall.
- Network parameters: fireDeconv upsamples feature maps to produce an original-resolution point-wise mask prediction.
- Network parameters: Output features from the mask-prediction enlargement layer are concatenated before estimating the relative 6-DoF pose.The final fully connected layers output translation x and rotation quaternion q.
- Network parameters: Batch normalization, data normalization, and rescaling did not significantly improve performance and sometimes reduced odometry accuracy.
B. More odometry results
The additional evaluations report trajectories and quantitative results across KITTI and Ford benchmark datasets, including sequences with and without ground truth. LO-Net+Mapping achieves the best reported KITTI performance across path-length and driving-speed analyses, while LO-Net has accuracy similar to the state-of-the-art geometry-based approach.
- More odometry results: Evaluations cover KITTI sequences 00–21 and the Ford dataset, with ground truth available for some KITTI sequences and unavailable for others.KITTI sequences 00–10 receive quantitative evaluation, while trajectory plots are also shown for later sequences and Ford.
- More odometry results: The KITTI quantitative evaluation results are presented for sequences 00–10, while the trajectory plots include comparisons with GICP, CLS, LOAM, LO-Net, and LO-Net+Mapping.
- More odometry results: LO-Net outperforms existing ICP-based approaches and has similar accuracy to the state-of-the-art geometry-based approach.
- More odometry results: LO-Net+Mapping achieves the best performance among all evaluated methods for average translation and rotation errors across KITTI path-length intervals.
- More odometry results: LO-Net+Mapping achieves the best performance among all evaluated methods for average translation and rotation errors across KITTI driving speeds.