KN-LIO: Geometric Kinematics and Neural Field Coupled LiDAR-Inertial Odometry

TL;DR

KN-LIO addresses the limitation of traditional LiDAR-Inertial Odometry by tightly coupling geometric kinematics with neural field-based dense mapping. It represents the map as a neural point cloud decoded by an online SDF decoder and fuses LiDAR and IMU data via an error-state Kalman filter in both semi-coupled and tightly coupled modes, while supporting asynchronous multi-LiDAR inputs. Experimental results on VIRAL, HILTI2022, and Newer College show competitive pose accuracy with state-of-the-art LIO methods and significantly improved dense mapping over pure LiDAR approaches, with additional robustness when using multiple LiDARs. This approach enables robust online dense mapping in high-dynamics scenarios and broadens applicability to multi-sensor platforms for real-time navigation and mapping.

Abstract

Recent advancements in LiDAR-Inertial Odometry (LIO) have boosted a large amount of applications. However, traditional LIO systems tend to focus more on localization rather than mapping, with maps consisting mostly of sparse geometric elements, which is not ideal for downstream tasks. Recent emerging neural field technology has great potential in dense mapping, but pure LiDAR mapping is difficult to work on high-dynamic vehicles. To mitigate this challenge, we present a new solution that tightly couples geometric kinematics with neural fields to enhance simultaneous state estimation and dense mapping capabilities. We propose both semi-coupled and tightly coupled Kinematic-Neural LIO (KN-LIO) systems that leverage online SDF decoding and iterated error-state Kalman filtering to fuse laser and inertial data. Our KN-LIO minimizes information loss and improves accuracy in state estimation, while also accommodating asynchronous multi-LiDAR inputs. Evaluations on diverse high-dynamic datasets demonstrate that our KN-LIO achieves performance on par with or superior to existing state-of-the-art solutions in pose estimation and offers improved dense mapping accuracy over pure LiDAR-based methods. The relevant code and datasets will be made available at https://**.

Figures (4)

• Figure 1: Overview of Kinematic-Neural LiDAR-Inertial Odometry. The high-frequency IMU readings are used for recursive system state propagation on the manifold. Asynchronous heterogeneous point clouds are first integrated based on timestamps and extrinsics, and then undistorted according to the recursive state. In a semi-coupled mode, the undistorted point cloud is registered with a neural field by minimizing point-SDF residuals, and the error-state Kalman filter is updated with the registration result. In a tightly coupled mode, the system calculates the point-SDF residuals and updates the error-state Kalman filter based on them, iterating until convergence. The system represents the map as a set of neural points, where the feature of a given spatial point is obtained by weighted distance to its neighboring neural points, and its corresponding SDF value is decoded from its features through a simple MLP. Under the tracked state, the system performs bundle optimization on the local neural map together with the current point cloud. Finally, the dense mesh of the scene is reconstructed through SDF queries and marching cubes.
• Figure 2: Asynchronous LiDAR merging.
• Figure 3: Meshes reconstructed on VIRAL experimental sites by PIN-SLAM, ImMesh, and our KN-LIO.
• Figure 4: Mapping with the horizontal LiDAR (left) and with both the horizontal and vertical LiDARs (right).