Abstract: Most existing lidar odometry estimation strategies are formulated under a standard framework that includes feature selection, and pose estimation through feature matching. In this work, we present a novel pipeline called LO-Net for lidar odometry estimation from 3D lidar scanning data using deep convolutional networks. The network is trained in an end-to-end manner, it infers 6-DoF poses from the encoded sequential lidar data. Based on the new designed mask-weighted geometric constraint loss, the network automatically learns effective feature representation for the lidar odometry estimation problem, and implicitly exploits the sequential dependencies and dynamics. Experiments on benchmark datasets demonstrate that LO-Net has similar accuracy with the geometry-based approach.
Introduction
Estimating a vehicle's motion, or odometry, is a cornerstone of autonomous navigation. While traditional methods like Iterative Closest Point (ICP) or Lidar Odometry and Mapping (LOAM) have been the standard, they often struggle with dynamic environments—where moving objects like other cars can confuse the sensor—and the inherent sparsity of 3D point clouds. LO-Net addresses these challenges by leveraging deep learning to estimate 6-DoF (Degrees of Freedom) motion directly from sequential lidar scans.
The LO-Net Architecture
LO-Net is an end-to-end deep convolutional network designed to process lidar data efficiently. To handle the irregular nature of 3D point clouds, the system first projects the data into a structured 2D matrix using cylindrical coordinates. This allows the use of standard CNN architectures.
The architecture incorporates lightweight "fireConv" and "fireDeconv" modules, inspired by SqueezeNet, to maintain high performance while reducing the parameter count. This makes the model suitable for deployment on mobile robots with limited computational resources.
Geometric Constraints and Dynamic Masking
A key innovation of LO-Net is the "mask-weighted geometric constraint loss." The network performs two auxiliary tasks alongside motion estimation:
- Normal Estimation: It predicts surface normals to gain a better understanding of the environment's geometry.
- Dynamic Masking: It generates a mask that identifies reliable static regions versus unreliable dynamic objects (like moving pedestrians).
By combining these, the loss function ensures that the predicted motion is geometrically consistent with the static parts of the scene. This allows the network to "implicitly" learn to ignore moving objects without requiring explicit semantic labels during training.
Experimental Results
The model was rigorously tested on the KITTI Odometry Dataset and the Ford Campus Vision and Lidar Data Set. On KITTI, LO-Net achieved a mean translational error of 1.09% and a rotational error of 0.63°/100m.
As shown in the evaluation plots, LO-Net maintains competitive accuracy compared to highly optimized traditional algorithms like Generalized ICP (GICP). Notably, the system demonstrates strong generalization; a model trained on KITTI data performed effectively on the Ford dataset, indicating it learns universal geometric features rather than environment-specific patterns.
Conclusion and Future Directions
LO-Net represents a significant step toward robust, learning-based lidar odometry. By integrating geometric constraints directly into the learning process, it achieves high accuracy and robustness to dynamic objects. Future work aims to incorporate loop closure mechanisms to reduce long-term drift and explore unsupervised learning paradigms to eliminate the need for labeled ground-truth data.
Summary updated