.png){kind=icon}
.png){kind=icon}
Return
LiDAR Loop Closure Detection Method Based on 3D NDT Features
.png){kind=icon}
PRE
AI
DOI：10.3788/lop252050.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Objective In autonomous driving and mobile robotics, simultaneous localization and mapping (SLAM) with LiDAR is a key technology for achieving high-precision navigation and environmental perception, while loop closure detection is a critical step for suppressing trajectory drift and improving the global consistency of maps. Traditional LiDAR-based loop closure detection methods typically adopt two-dimensional polar partitioning and exploit only horizontal structural features, underutilizing vertical geometric information. This limitation leads to a significant drop in recognition accuracy in environments with complex three-dimensional structures and repetitive textures, such as urban blocks and indoor corridors. Furthermore, existing descriptor designs insufficiently explore the local statistical properties of point cloud distributions, making it difficult to leverage both geometric shape and distribution complexity information. As a result, regions with similar geometric appearances or markedly different complexities are often misidentified. To address these issues, this paper proposes a LiDAR loop closure detection algorithm based on fused three-dimensional normal distributions transform (3D NDT) features. By employing volumetric spatial partitioning and jointly incorporating a shape index and information entropy into a multi-level feature fusion framework, the proposed method significantly enhances the discriminative power, robustness, and efficiency of loop closure detection, thereby improving the global localization accuracy and map consistency of SLAM systems in complex environments. Methods First, the input point cloud is voxelized, and NDT statistical modeling is applied. Building upon conventional two-dimensional polar partitioning, a height dimension is introduced to construct a three-dimensional polar region of interest (ROI) partitioning framework, effectively capturing vertical structural information. The covariance matrix of each voxel is then decomposed to extract a shape index for distinguishing linear, planar, and volumetric geometric structures. Additionally, the determinant of the covariance matrix is computed to derive information entropy as a measure of point cloud distribution complexity. Geometric category voting combined with height-weighted summation is used to fuse geometric and entropy information, producing a multi-level NDT descriptor that balances discriminative ability with stability. During loop closure detection, a multi-stage matching strategy- consisting of spatial distance constraints, geometric keyword coarse filtering, and normalized cross-correlation fine matching-is adopted to reduce computational complexity and improve matching reliability. Once a loop is detected, the result is integrated into the back-end global optimization module to correct trajectory drift and improve map consistency. Results and Discussions The proposed algorithm is validated on the KITTI05 and KITTI07 sequences in KITTI dataset, and Walking dataset, it is compared with A-lidar odometry and mapping (A-LOAM), lidar inertial odometry via smoothing and mapping (LIO-SAM), and Scan Context-LIO-SAM (SC-LIO-SAM). Experimental results demonstrate superior performance of proposed algorithm in trajectory accuracy, loop closure reliability, and map consistency. Global trajectory results show that the proposed algorithm yields stable convergence at loop points, smooth and natural paths, and the closest alignment with ground truth, particularly excelling in long backtracking paths and intersecting loop scenarios.Quantitative absolute pose error (APE) evaluation further confirms that the proposed algorithm outperforms competing algorithms in both mean error and stability. This improvement is primarily attributed to the combined use of the shape index and information entropy, which enhances descriptor discriminability and allows differentiation between geometrically similar structures. Loop closure performance indicates that the method maintains high precision even at high recall, validating its robustness and discriminative advantage in complex environments. On the Walking dataset, height-based pseudo-color visualization reveals continuous ground elevation and coherent wall boundaries in top, bottom, and side views for proposed algorithm, whereas SC-LIO-SAM exhibits abrupt color changes and broken edges. Quantitative trajectory analysis further demonstrates that proposed algorithm produces smoother and more stable results in terms of height variation and trajectory smoothness, consistent with the visualization findings. Local detail comparisons further show that SC-LIO-SAM suffers from misaligned and blurred window frames, while the proposed algorithm preserves clear outlines and straight boundaries, achieving superior detail retention. Conclusions<br /> This paper presents a loop closure detection algorithm that integrates 3D NDT features. To address the shortcomings of traditional methods such as insufficient recognition accuracy in complex three-dimensional environments and susceptibility to repetitive structures or missing height information the approach consolidates and refines multiple technical elements to better adapt to challenging scenarios. First, vertical structural features are explicitly modeled through three-dimensional polar partitioning. enabling the descriptor to fully capture the spatial geometry of the environment. Second, a shape index is introduced to distinguish different geometric morphologies, while information entrop is employed to quantify point cloud distribution complexity, thereby achieving an effective fusion of geometric and statistical constraints, Finally, a multi-stage matching strategy is designed to maintain detection accuracy while significantly reducing computational complexity and improving robustness. Experimental results demonstrate that the proposed algorithm exhibits clear advantages in complex scenes such as urban blocks and long indoor corridors. This study not only substantially enhances the discriminative power, robustness, and efficiency of loop closure detection but also provides a new solution for achieving high-precision localization and globally consistent mapping in SLAM systems operating under challenging conditions. Future work will focus on evaluating the method's adaptability in dynamic environments, investigating its deep integration with other sensors (e.g.. inertial measurement unit and vision, and exploring the use of deep learning techniques for automatic feature extraction and optimized matching to further improve the intelligence and generalization capability of loop closure detection.
Keywords:
LiDAR
loop closure detection
geometric feature
voxelized point cloud
Journal
L
IF:
1
Papers:
596
Citations:
0