BEV-LIO(LC): BEV Image Assisted LiDAR-Inertial Odometry with Loop Closure (2502.19242v1)

Abstract: This work introduces BEV-LIO(LC), a novel LiDAR-Inertial Odometry (LIO) framework that combines Bird's Eye View (BEV) image representations of LiDAR data with geometry-based point cloud registration and incorporates loop closure (LC) through BEV image features. By normalizing point density, we project LiDAR point clouds into BEV images, thereby enabling efficient feature extraction and matching. A lightweight convolutional neural network (CNN) based feature extractor is employed to extract distinctive local and global descriptors from the BEV images. Local descriptors are used to match BEV images with FAST keypoints for reprojection error construction, while global descriptors facilitate loop closure detection. Reprojection error minimization is then integrated with point-to-plane registration within an iterated Extended Kalman Filter (iEKF). In the back-end, global descriptors are used to create a KD-tree-indexed keyframe database for accurate loop closure detection. When a loop closure is detected, Random Sample Consensus (RANSAC) computes a coarse transform from BEV image matching, which serves as the initial estimate for Iterative Closest Point (ICP). The refined transform is subsequently incorporated into a factor graph along with odometry factors, improving the global consistency of localization. Extensive experiments conducted in various scenarios with different LiDAR types demonstrate that BEV-LIO(LC) outperforms state-of-the-art methods, achieving competitive localization accuracy. Our code, video and supplementary materials can be found at https://github.com/HxCa1/BEV-LIO-LC.

BEV-LIO(LC): BEV Image Assisted LiDAR-Inertial Odometry with Loop Closure

The paper introduces BEV-LIO(LC), a novel LiDAR-Inertial Odometry (LIO) framework aimed at enhancing the performance of simultaneous localization and mapping (SLAM) by integrating Bird's Eye View (BEV) image representations of LiDAR data. This framework innovatively combines geometry-based point cloud registration with loop closure detection through BEV image features. The authors propose an approach that utilizes a lightweight convolutional neural network (CNN) to extract both local and global descriptors from BEV images to advance localization accuracy and robustness.

Methodology

BEV-LIO(LC) employs a three-stage process to optimize SLAM performance:

1. BEV Image Projection and Feature Extraction: The key innovation lies in the projection of LiDAR point clouds into BEV images, enabling efficient feature extraction and matching. BEV images maintain scale consistency and spatial relationships, distinguishing them from spherical range images which suffer from distortions. A lightweight CNN is then applied to extract unique local and global descriptors from the BEV images.
2. Odometry Framework and Reprojection Error Minimization: The local descriptors are utilized for frame-to-frame matching by creating reprojection error models through BEV image features, which then integrate into a tightly coupled iterated Extended Kalman Filter (iEKF) framework. By deriving analytic Jacobians, a combination of geometric and reprojection residuals is employed to optimize the state estimation process, aiming to enhance front-end odometry accuracy.
3. Loop Closure Detection and Factor Graph Optimization: The global descriptors facilitate loop closure detection by constructing a KD-tree-indexed keyframe database. Upon identifying loop closure candidates, Random Sample Consensus (RANSAC) computes a coarse transformation from BEV image matching, providing an initial estimate for refinement by Iterative Closest Point (ICP) alignment. This refined transformation is incorporated into a factor graph along with odometry factors to ensure global localization consistency and reduce drift.

Experimental Evaluation and Results

The authors conducted extensive experiments across diverse scenarios using differing LiDAR setups, demonstrating that BEV-LIO(LC) outperforms state-of-the-art methods in localization accuracy. The experimental results reported in public datasets such as the Multi-Campus Dataset (MCD) and the Newer College Dataset (NCD) confirm the competitive performance of BEV-LIO(LC) under challenging conditions. The integration of BEV-based methods into the odometry framework and loop closure correction showed significant improvements in trajectory estimation and global localization accuracy compared to existing techniques such as FAST-LIO2 and COIN-LIO.

Theoretical and Practical Implications

The paper contributes to the SLAM research by bridging the gap between real-time odometry and effective loop closure detection. The integration of BEV images into SLAM can potentially generalize across various LiDAR configurations, tackling issues like data sparsity and scale distortion inherent in existing methods. Furthermore, the use of BEV-based loop closure methods in factor graphs highlights a path forward for enhancing global consistency in SLAM systems. Practically, BEV-LIO(LC) holds promise for real-world applications in autonomous navigation where robust localization is crucial.

Future Directions

Looking ahead, the research invites exploration into eliminating reliance on CNN-based feature extraction to further improve system efficiency and decrease computational costs. The adaptability of BEV-LIO(LC) across various environmental scenarios and LiDAR types ensures that this approach can continue to be refined and evaluated for further improvements in real-time SLAM applications. This work lays the groundwork for advancing the integration of LiDAR data and odometry functions, bolstering both the research domain and practical applications in autonomous systems.