A Minimal Subset Approach for Informed Keyframe Sampling in Large-Scale SLAM

TL;DR

Large-scale LiDAR SLAM faces loop-closure bottlenecks as candidate pairs explode with mission duration. The authors propose the Minimal Subset Approach (MSA), an online keyframe sampling method that optimizes two descriptor-space criteria—redundancy minimization and information preservation—within a sliding window to reduce keyframes while preserving loop-closure potential. By jointly considering place recognition and pose-graph optimization in a descriptor-space framework and avoiding manual tuning, MSA improves ATE/RPE and PR-AUC while lowering memory usage and computation. Empirical results across KITTI, MulRan, and Apollo-SB datasets, plus ablations, demonstrate robust performance gains across urban, campus, and rural scenarios, highlighting MSA’s scalability for large-scale SLAM.

Abstract

Typical LiDAR SLAM architectures feature a front-end for odometry estimation and a back-end for refining and optimizing the trajectory and map, commonly through loop closures. However, loop closure detection in large-scale missions presents significant computational challenges due to the need to identify, verify, and process numerous candidate pairs for pose graph optimization. Keyframe sampling bridges the front-end and back-end by selecting frames for storing and processing during global optimization. This article proposes an online keyframe sampling approach that constructs the pose graph using the most impactful keyframes for loop closure. We introduce the Minimal Subset Approach (MSA), which optimizes two key objectives: redundancy minimization and information preservation, implemented within a sliding window framework. By operating in the feature space rather than 3-D space, MSA efficiently reduces redundant keyframes while retaining essential information. Evaluations on diverse public datasets show that the proposed approach outperforms naive methods in reducing false positive rates in place recognition, while delivering superior ATE and RPE in metric localization, without the need for manual parameter tuning. Additionally, MSA demonstrates efficiency and scalability by reducing memory usage and computational overhead during loop closure detection and pose graph optimization.

Figures (9)

• Figure 1: An example on KITTI 06. Comparison of two trajectories from KISS-ICP, after Pose Graph Optimization (PGO). In the top figure, keyframes are sampled using an entropy-based approach, while in the bottom figure, keyframes are sampled with the proposed Minimal Subset Approach (MSA) which achieves a lower Relative Pose Error (RPE).
• Figure 2: Redundant keyframes effect. Example of removing redundant keyframes (light blue). The updated odometry edge set, $\mathcal{E}_o^*$, influences the pose graph optimization, while the sparsified loop closure edges, $\bar{\mathcal{L}} = \mathcal{L} \:\backslash \langle 5,7 \rangle$, can further reduce the computational complexity.
• Figure 3: Overall pipeline. Block diagram of the overall pipeline, highlighting the interplay between front-end (descriptor extraction and odometry modules) and back-end optimization (place recognition, loop closure detection and pose graph optimization), facilitated by the proposed keyframe sampling scheme.
• Figure 4: Box plot comparisons. Translational (t) and Rotational (R) Absolute Trajectory Error (ATE) difference of every method compared to the baseline of All Samples, as well as the Memory Allocation in Gigabytes (GB) and the total Execution Time in seconds (s).
• Figure 5: Trajectory comparisons. Comparison of the ground truth, raw KISS-ICP, and the sampled poses after the pose graph optimization for the proposed Minimal Subset Approach (MSA) and the entropy-based approach on the KAIST sequence of the MulRan dataset and the SanJoseDowntown sequence of the Apollo-SouthBay dataset.