Source-linked AI summary
Temporal-Aware Fusion for Robust Outdoor LiDAR Localization
Minghang Zhu, Zhijing Wang, Yuxin Guo, Chen Liu, Yongshu Huang, Wen Li, Sheng Ao, Cheng Wang
TL;DR
Single-frame regression-based LiDAR relocalization can struggle in dynamic or ambiguous environments, and existing sequence methods may impose computational overhead. TempLoc estimates per-scan coordinates and uncertainty, models inter-frame correspondences, and fuses them; it reports consistent benchmark gains, including nearly 30% higher NCLT localization accuracy than LightLoc.
Problem
Regression-based LiDAR relocalization often struggles with noise and geometric variation in complex or ambiguous environments, while existing sequence methods only roughly encode spatiotemporal global features.
Source: §1 · p. 2Method
TempLoc predicts per-point global coordinates and uncertainties, estimates inter-frame point transitions, and fuses the resulting correspondence predictions for RANSAC-based 6-DoF pose estimation.
Source: §1 · p. 2Results
Takeaways & Limitations
TempLoc integrates historical and current information, enhancing robustness in complex outdoor scenarios with dynamic objects.
Source: §5 · p. 8The authors note that sequence-based methods such as STCLoc and NIDALoc rely on long-term sequence information, causing computational overhead that hinders real-time performance.
Source: §2.1 · p. 2
Abstract
from arXiv · showhide
LiDAR relocalization aims to estimate the global 6-DoF pose of a sensor in the environment. However, existing regression-based approaches often encounter limitations in dynamic or ambiguous scenarios, as they typically prioritize single-frame inference, leaving the potential of spatio-temporal consistency across scans not fully explored. In this paper, we propose a Temporal-aware Localization framework (TempLoc) designed to enhance the robustness of outdoor localization by effectively modeling sequential consistency. Specifically, a Global Coordinate Estimation module is first introduced to predict point-wise global coordinates and associated uncertainties for each LiDAR scan. A Prior Coordinate Generation module is then presented to estimate inter-frame point correspondences by the attention mechanism. Lastly, an Uncertainty-Guided Coordinate Fusion module is deployed to integrate both predictions of point correspondence in an end-to-end fashion, yielding a more temporally consistent and accurate global 6-DoF pose. Experimental results on the NCLT and Oxford RobotCar benchmarks show that our TempLoc outperforms state-of-the-art methods by a large margin, demonstrating the effectiveness of temporal-aware correspondence modeling in LiDAR relocalization.
1 Introduction
Regression-based LiDAR relocalization offers map-free pose prediction but remains vulnerable to ambiguous scenes and often overlooks point-level consistency across scans. TempLoc models sequential correspondences and uncertainty to improve robust 6-DoF localization, reporting state-of-the-art benchmark performance.
1 Introduction
- Nearly 30% higher localization accuracy than LightLoc on NCLT is reported for TempLoc, which also achieves superior accuracy on NCLT and QE-Oxford.Figure 1 compares mean position error on both datasets.Source: Figure 1 · p. 2, §1 · p. 2
- Single-scan regression methods struggle with measurement noise and geometric variation from spatial changes or dynamic objects, while existing sequence methods only roughly encode spatiotemporal global features.The paper identifies omitted point-level sequential consistency as a source of suboptimal localization performance.Source: §1 · p. 2
- The framework produces a global 6-DoF pose through RANSAC-based optimization of refined 3D correspondences.The uncertainty-guided fusion mechanism is described as end-to-end differentiable.Source: §1 · p. 2
- The authors report that explicitly modeling temporal scene-coordinate evolution mitigates single-frame regression limitations and enhances localization robustness.They also report state-of-the-art performance on two public benchmarks.Source: §1 · p. 2
- The paper identifies three contributions: temporal point-correspondence modeling, uncertainty-aware per-point global-coordinate prediction, and experiments and ablations examining the method and its components.The coordinate prediction is made from a single LiDAR scan.Source: §1 · p. 2
- TempLoc extends scene-coordinate regression into the temporal domain by incorporating sequential constraints into point-correspondence estimation.Its components predict per-point global coordinates and uncertainties, estimate inter-frame point transitions, and fuse correspondence predictions.Source: §1 · p. 2
2 Related Work
The related work contrasts map-based localization with map-free regression, reviews absolute pose and scene-coordinate regression, and motivates uncertainty-aware temporal fusion to address their respective limitations.
2 Related Work
- Map-based methods use retrieval or registration with pre-built feature maps, while regression methods encode maps in network parameters and predict global poses without added map storage.Regression methods include Absolute Pose Regression (APR) and Scene Coordinate Regression (SCR).Source: §2 · p. 2
2.1 Absolute Pose Regression
- APR methods directly regress pose from a single frame, while sequence-based approaches use temporal context but can incur high computational overhead that hinders real-time performance.The reviewed sequence methods include shared-feature multitask learning, visual-odometry constraints, recurrent relative-pose estimation, Kalman filtering, spatial-temporal regularization, and memory modules.Source: §2.1 · p. 2
2.2 Scene Coordinate Regression
- SCR regresses world-coordinate points and relies on RANSAC for pose estimation, but coordinate outliers can still compromise localization accuracy.The paper motivates uncertainty estimation to mitigate outlier impact and describes fusing temporal information from neighboring frames for pose estimation and smoother trajectories.Source: §2.2 · p. 3
3 Methodology
TempLoc combines per-scan global coordinate estimates and uncertainties with adjacent-frame coordinate priors, then fuses them into refined correspondences for RANSAC pose estimation. Its modules address unreliable point predictions and use attention-based registration to generate temporal priors.
3 Methodology
- GCE, PCG, and UCF produce refined local-to-global 3D correspondences, from which RANSAC computes the global 6-DoF pose.GCE supplies per-scan coordinates and uncertainties; PCG estimates adjacent-frame correspondences and credibility assessments; UCF integrates their outputs.Source: §3 · p. 3
3.1 Global Coordinate Estimation
- Per-point uncertainty estimation augments SCR to assess coordinate quality in noisy, dynamic, or structurally sparse environments.The model predicts global coordinates and uncertainty scores together; the ground-truth uncertainty threshold decays during training.Source: §3.1 · pp. 3–4
- The coordinate objective minimizes mean L1 distance, while a dedicated mean-squared-error term supervises predicted uncertainty.The total localization loss sums the regression and uncertainty losses.Source: §3.1 · p. 4
3.2 Prior Coordinate Generation
- PCG generates current-frame priors by estimating soft correspondences and uncertainty for previous-frame points, then interpolating prior coordinates to current-frame points.It uses PCAM with self-attention to enhance within-cloud point features and generate discriminative descriptors.Source: §3.2 · p. 4
8 PO ls {IPO Ile
The temporal prior is generated by attention-based correspondences between adjacent frames, then propagated to actual current-frame points by neighborhood interpolation. Uncertainty-guided fusion combines prior and per-scan estimates into refined coordinates and uncertainties.
8 PO ls {IPO Ile
- A global attention matrix generates soft current-frame correspondences for points in the previous frame.It aggregates multi-level cross-attention matrices by element-wise multiplication.Source: §8 · p. 4
- Because soft correspondences may not match sampled current-frame points, neighborhood interpolation assigns prior coordinates to actual points.Each pseudo-neighbor inherits the world coordinate of its source point, and the prior coordinate is a distance-weighted average over k neighbors.Source: §8 · pp. 4–5
3.3 Uncertainty-Guided Coordinate Fusion
- The propagated prior coordinates and uncertainties receive joint supervision, and the overall loss weights its three losses at 0.3, 0.3, and 0.4.The prior-coordinate loss is analogous to the coordinate-estimation losses.Source: §8 · p. 5, §3.3 · p. 5
- Uncertainty-guided fusion combines per-scan measurement estimates with temporally propagated prior estimates.A Softmax over prior and measurement uncertainties produces positive weights summing to one, which are applied to coordinates and uncertainties.Source: §3.3 · p. 5
4 Experiments
TempLoc is evaluated on outdoor LiDAR benchmarks against APR, SCR, temporal, and retrieval-based methods, with results covering accuracy, dynamic-scene robustness, efficiency, and component ablations.
4.1 Experimental Setup
- Evaluation uses Oxford RobotCar and NCLT, average translation and rotation error, and APR, SCR, temporal, and retrieval-based comparison methods.The QE-Oxford benchmark is an enhanced Oxford version with refined ground-truth poses.Source: §4.1 · pp. 5–6
4.2 Evaluation
- 0.72m/0.99° versus LightLoc’s 0.83m/1.12° on QE-Oxford; on NCLT, TempLoc cuts translation error to 1.05m from LightLoc’s 1.46m, though LiSA’s rotation error is lower (2.31° vs 2.56°).The reported NCLT translation improvement over LightLoc is 28%.Source: §4.2 · pp. 6–7, Table 3 · p. 6
- On NCLT, TempLoc reduces feature storage to 0 MB versus BEVplace++’s 841 MB and achieves 68 ms latency within the 100 ms real-time limit using a 2.3 GB GPU footprint.Source: §4.2 · p. 7
- On QE-Oxford dynamic sequences, TempLoc remains stable while SGLoc, LightLoc, and RALoc degrade; the reported dynamic-scene evaluation is summarized in Table 5.In a dynamic-vehicle intersection, TempLoc achieves near-perfect point-cloud overlap with ground truth while competing methods show ghosting and misalignment.Source: Table 5 · p. 7, §4.2 · pp. 7–8
4.3 Ablation Study
- 19% (1.30m to 1.05m) NCLT translation gain from PCG+UCF over UE-only; gains are smaller on the structurally more stable Oxford and QE-Oxford datasets.The ablation study reports that each TempLoc component plays a critical role.Source: §4.3 · p. 8, Table 6 · p. 8
5 Conclusion
TempLoc models sequential consistency across LiDAR scans for map-free 6-DoF pose estimation. Experiments on Oxford, QE-Oxford, and NCLT report stronger translation and rotation accuracy than state-of-the-art methods.
5 Conclusion
- TempLoc is designed to address robustness limitations of single-frame methods in challenging dynamic outdoor environments.It models sequential consistency across consecutive LiDAR scans without dense map storage or pre-built databases.Source: §5 · p. 8
- Experiments on Oxford, QE-Oxford, and NCLT show TempLoc consistently outperforms state-of-the-art methods in translation and rotation accuracy.Source: §5 · p. 8
- TempLoc integrates historical and current information, enhancing robustness in complex outdoor scenarios with dynamic objects.Source: §5 · p. 8