Online Geometric Consistency Assessment of LiDAR Measurements via Recursive Bayesian Comparison
Authors: Mingqiao Han, Qigeng Duan, Xuchen Liu, Jinqiang Cui, Xi Chen, Ben M. Chen
Organizations: Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Shatin, N.T., Hong Kong, China. · Pengcheng Laboratory, Shenzhen, Guangdong, China.
LiDAR-based localization and mapping systems produce registered point clouds that serve as geometric measurements for downstream estimation and navigation. Registration errors introduce duplicated surfaces and structural deformation into these measurements. In the absence of an independent reference, online assessment relies on geometry accumulated from earlier measurements. Such assessment requires tolerance to sampling variation and an accumulation process in which observations enter the history only after evaluation. We therefore formulate a recursive Bayesian comparison that couples geometric assessment with history maintenance. A Normal-Inverse-Wishart model represents historical voxel geometry and its uncertainty, while a weak isotropic reference adapts to current sampling. Comparing the marginal predictive support of the same current points under both models yields Normalized Drift Evidence (NDE). Assessment precedes assimilation: positive NDE triggers an alarm and preserves the preceding historical state; otherwise, the current observations update the history. Across 14 real-world sequences totaling approximately 73 min from four platform types, alarms identified all nine corrupted sequences and were absent from all five consistent sequences. In four representative scenarios with event references, first alarms occurred within 0.9 s of the reference times, with no preceding alarms. Cross-backend evaluation demonstrates consistent responses to reconstructed geometry. Ablation and controlled perturbation results establish the roles of uncertainty modeling, recursive history, displacement, and geometric support. Mean processing time is 4.2--6.1 ms per update at 1 Hz. In ten UAV trials, alarms initiated odometry isolation, landing, relocalization, and mission continuation.
Figures & tables
Fig. 1: Geometric inconsistency from misregistration. (a) A pose error at frame k+1 misaligns a revisited surface. (b) Voxelization reveals local mismatch. (c) Voxel centroid, orientation, and shape change. Green/red denote nominal/failure-affected observations.
Fig. 2: Local geometry during nominal mapping and corruption. (a)–(b) Compatible voxel distributions persist while motion extends the observed region. (c)–(d) Corruption changes revisited distributions in centroid, orientation, and thickness. Ellipsoids represent sampled geometry; blue, purple, and red indicate increasing thickness.
Fig. 3: Geometric consistency monitoring and recursive updating. (A) Current observations are scored under historical model Mv,k−1c and weak, non-propagated reference Mv,kr . NDE negates the mean contrast sv,k over eligible voxels. Positive NDE triggers an alarm and retains the preceding state. (B) Otherwise, NIW updates incorporate eligible observations for subsequent comparisons.
Fig. 4: UAV monitoring and recovery. SLAM supplies nominal odometry; NDE monitors registered geometry. After an alarm, odometry isolation precedes landing with the flight controller’s own sensing and control. SLAM restart and relocalization against the retained pre-failure map restore odometry for mission continuation.
Fig. 5: Experimental platforms and environments. Top: handheld, truck-mounted, and aerial LiDAR platforms. Bottom: a feature-rich indoor scene, a long tunnel with weak geometric constraints, and an outdoor scene with limited geometric structure.
ID
Platform
LiDAR
Failure manifestation
Main cause
Dur. (s)
tref (s)
tdet (s)
Δt (s)
NDEmax
Npre
S1
Handheld
MID-360
Abrupt mis-registration
Severe vibration
16.7
6.1
6.1
0.0
+18.7
0
S2
Truck
Livox Avia
Abrupt mis-registration
Geometric degeneracy
1107.5
1078.3
1079.2
0.9
+450.6
0
S3
Truck
Livox Avia
Cumulative drift-induced failure
Geometric degeneracy
182.3
151.0
151.2
0.2
+63.1
0
S4
Handheld
MID-360
Cumulative drift-induced failure
Large-scale traversal
263.3
238.3
238.7
0.4
+17.7
0
TABLE I: Event-level results for S1–S4. Reference times tref follow Section IV-A ; tdet is the first NDE alarm in each sequence. Δt reports the reference offset, and Npre counts preceding alarms.
Fig. 6: Scenarios S1–S4: abrupt corruption under vibration and degeneracy (top two rows), then drift under degeneracy and large-scale traversal (bottom two). Rows pair registered geometry at frames k−1 through k+2 with NDE. Green/red labels denote observations before/after the reference time; dashed lines mark zero. Blue, purple, and red ellipsoids indicate increasing thickness.
Fig. 7: Geometry and NDE for ten GEODE runs: (a) corrupted G1–G5 and (b) consistent G6–G10. Map and trajectory inspection determines these groups without using NDE. All runs share the monitoring configuration and zero boundary. Blue, purple, and red ellipsoids indicate increasing thickness.
Fig. 8: Cross-backend comparison on identical input. PV-LIO, VoxelMap, and LIO-SAM rows pair four consecutive registered-map snapshots with NDE under fixed settings. Blue, purple, and red ellipsoids indicate increasing thickness.
ID
Sequence
Platform
Scenario
Dur. (s)
Geometry status
tresp (s)
NDEmax
Outcome
G1
flat_surfaces_smooth
Handheld
Flat surface
79
Corrupted
14.8
+20.4
TP
G2
flat_surfaces_aggressive
Handheld
Flat surface
86
Corrupted
17.7
+105.6
TP
G3
stairs_gamma
Handheld
Stairs
193
Corrupted
115.2
+112.3
TP
G4
shield5
UGV/Handheld
Shield tunnel
35
Corrupted
30.7
+28.0
TP
G5
Tunneling_tunnel5_gamma
UGV
Mine tunnel
232
Corrupted
138.401
+785
TP
G6
offroad4
UGV
Off-road
402
Geometrically consistent
–
−1.6
TN
TABLE II: Sequence-level GEODE validation with fixed settings. Visual inspection of backend maps assigns geometry status without using NDE. “Response” denotes any zero crossing within the sequence; the summary gives sequence-level counts.
NDE (ours)
NIS-like
ICP-fit
X-ICP-inspired [ 10 ]
Hessian-based [ 20 ]
Seq.
Dur.
tref
tdet
Nal
Npre
rpre
O.W.
tdet
Nal
Npre
rpre
O.W.
tdet
Nal
Npre
rpre
O.W.
tdet
Nal
Npre
rpre
O.W.
tdet
Nal
Npre
rpre
O.W.
S1
16.7
6.1
6.1
1
0
0%
✓
2
5
3
60%
✓
4
3
3
100%
✓
—
0
0
—
×
—
0
0
—
×
S2
1107.5
1078.3
1079.2
1
0
0%
✓
22
617
589
95%
×
26
184
155
84%
✓
15
396
368
93%
✓
20
285
258
91%
✓
S3
182.3
151.0
151.2
1
0
0%
✓
2
94
80
85%
✓
26
27
19
70%
×
5
61
46
75%
✓
12
53
38
72%
✓
S4
263.3
238.3
238.7
1
0
0%
✓
14
17
16
94%
×
12
12
12
100%
×
32
20
15
75%
×
—
1
1
100%
×
Summary
—
4
0
0%
4/4
—
733
688
94%
2/4
—
226
189
84%
2/4
—
477
429
90%
3/4
—
339
297
88%
2/4
TABLE III: Monitoring responses on S1–S4. The comparator columns report custom implementations of residual, fitting, localizability, and degeneracy indicators alongside NDE.
Fig. 9: NDE and custom residual-indicator responses on S1–S4. NIS-like scores use normalized squared point-to-plane residuals; ICP-fit uses mean absolute residuals. Dashed lines mark the zero NDE boundary and causal rolling thresholds for residual indicators.
Fig. 10: Dynamic-scene and cumulative-drift ablation: full NIW with recursive updating, NIW scoring without posterior propagation, and direct comparison of adjacent-frame voxel statistics. Red arrows identify the moving object.
Fig. 11: Persistent chair rearrangement: (a) original room; (b) rearranged room. The sensor leaves during rearrangement and returns afterward. FAST-LIO2 and the monitor process one continuous recording with unchanged settings.
Fig. 12: Intersection traffic: (a) vehicles queued at a red light; (b) vehicles moving through the junction. Monitoring uses original FAST-LIO2 registered clouds, 1 m voxels, 1 s accumulation, and the zero NDE boundary.
Condition
omin
Emed
n+/nv
nu
τ+
(%)
(s)
Unperturbed
96.3
313
0/19
0
–
x=0.10 m
94.5
308
0/19
0
–
x=0.25 m
87.9
294
19/19
0
0.70
x=10 m
8.1
66
18/19
0
1.70
x=20 m
0.0
224
1/14
5
16.50
TABLE IV: Representative controlled pose and support conditions, each with 19 exposure updates.
Condition
Defined
Positive
NDE
NDE
/total
/defined
median
max
Before rearrangement
46/46
0/46
−23.813
−15.464
Transit/intervention
79/79
7/79
−27.430
15.445
After rearrangement
50/50
0/50
−22.547
−1.114
Complete chair replay
195/197
7/195
−23.799
15.445
Intersection traffic
213/–
0/213
−35.296
−9.582
TABLE V: Responses to chair rearrangement and moving traffic.
Tacc (s)
Lv (m)
0.2
0.5
1.0
2.0
5.0
0.3
N/D
N/D
N/D
N/D
N/D
0.5
7.1 (+1.0)
6.9 (+0.8)
N/D
N/D
N/D
1.0
6.6 (+0.5)
6.3 (+0.2)
6.1 (0.0)
6.2 (+0.1)
N/D
2.0
F/A (1.3)
6.3 (+0.2)
6.3 (+0.2)
6.2 (+0.1)
F/A (0.0)
5.0
F/A (0.3)
F/A (0.0)
F/A (0.0)
F/A (0.0)
F/A (0.0)
TABLE VI: Voxel-size and accumulation-time sensitivity.
Parameter
Value
NDEmin
NDEmax
Npre
Scan of κ0r with ν0r=5 fixed
κ0r
0.001
−21.4
18.7
0
κ0r
0.005
−18.9
19.5
0
κ0r
0.010
−16.2
22.3
0
κ0r
0.050
−15.8
37.1
0
κ0r
0.100
−17.2
56.9
0
TABLE VII: Sensitivity to weak-reference hyperparameters.
Input
Alarms
Mean NDE
Furniture
Wall
Unperturbed
0/19
−23.29
−15.59
−54.95
0.25 m translation
19/19
+60.04
+11.16
+563.01
5∘ yaw
19/19
+71.23
+13.18
+422.98
TABLE VIII: Detection and regional evidence in the cluttered room.
Fig. 13: Indoor and outdoor UAV missions with vibration-induced SLAM failures. The indoor mission contains two injections. NDE alarms initiate odometry isolation and landing; relocalization against the retained map restores odometry for takeoff and mission continuation.
Scenario
Trials
Faults/trial
Alarmed
False alarms
tiso−talarm (s)
tland−talarm (s)
treloc−tland (s)
ttakeoff−treloc (s)
Continued
Indoor
5
1
5/5
0
0.31±0.06
5.50±0.23
37.40±2.60
3.18±0.27
5/5
Outdoor
5
1
5/5
0
0.30±0.02
3.64±0.58
49.38±4.32
4.36±0.27
5/5
TABLE IX: Repeated UAV response tests from failure to recovery. Values are means ± standard deviations.
FAST-LIO2
Proposed monitor
Seq.
CPU (%)
Mem. (MB)
CPU (%)
Mem. (MB)
Monitoring rate (Hz)
Processing time (ms)
Decision latency (ms)
S1
3–4
120–140
0.1–0.2
15–25
1.0
4.2
7.2
S2
12–15
130–350
0.1–0.2
10–20
1.0
6.1
6.5
S3
8–12
130–155
0.1–0.2
10–15
1.0
5.9
8.5
S4
3–4
120–180
0.1–0.2
15–25
1.0
4.7
6.9
TABLE X: Runtime overhead and timing in representative scenarios.
Camera localization in a prior LiDAR map provides a persistent geometric reference for long-term robotic navigation, yet remains challenging because of the substantial modality gap between camera images and point-cloud maps. We present a unified localization framework that makes the LiDAR map visually addressable rather than relying on a dedicated image-LiDAR correspondence model. Map geometry and reflectivity are rendered into LiDAR-derived quasi-images with explicit 2D-3D provenance, enabling camera observations and rendered map views to share mature visual features and matchers for both global localization and continuous pose tracking. Point and line correspondences are established through this common visual interface, while the retained provenance recovers metric LiDAR geometry and line-supported planar constraints for pose estimation. To improve robustness under ambiguous associations and weak geometry, we further introduce a distribution-aware, observability-complementary optimization strategy. Instead of reducing matching ambiguity to a scalar confidence, candidate association distributions are propagated into directional pose-information uncertainty, and reliable structural factors are selectively reinforced according to their ability to complement the currently weak pose directions. Experiments on the EuRoC MAV benchmark and self-collected real-world sequences demonstrate accurate global localization and robust continuous 6-DoF tracking using only a pre-built LiDAR map as the persistent prior, including under severe illumination variations and dynamic occlusions.
Wentao Zhao, Zikun Chen, Yihe Niu +2
School of Automation and Intelligent Sensing, Institute of Medical Robotics, Shanghai Jiao Tong University · School of Mathematical Sciences, Shanghai Jiao Tong University · School of Electronic and Information Engineering, Beijing Jiaotong University
LiDAR relocalization has attracted increasing attention as it can deliver accurate 6-DoF pose estimation in complex 3D environments. Recent learning-based regression methods offer efficient solutions by directly predicting global poses without the need for explicit map storage. However, these methods often struggle in challenging scenes due to their equal treatment of all predicted points, which is vulnerable to noise and outliers. In this paper, we propose LEADER, a robust LiDAR-based relocalization framework enhanced by a simple, yet effective geometric encoder. Specifically, a Robust Projection-based Geometric Encoder architecture which captures multi-scale geometric features is first presented to enhance descriptiveness in geometric representation. A Truncated Relative Reliability loss is then formulated to model point-wise ambiguity and mitigate the influence of unreliable predictions. Extensive experiments on the Oxford RobotCar and NCLT datasets demonstrate that LEADER outperforms state-of-the-art methods, achieving 24.1% and 73.9% relative reductions in position error over existing techniques, respectively. The source code is released on https://github.com/JiansW/LEADER.
Jianshi Wu, Minghang Zhu, Dunqiang Liu +5
Fujian Key Laboratory of Urban Intelligent Sensing and Computing · Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, School of Informatics, Xiamen University, China · School of Engineering Mathematics and Technology, University of Bristol
We propose SEAM, a LiDAR-based lifelong mapping framework. Instead of relying on a single anchor spanning the entire session, SEAM generates evidence based on a trajectory optimized with submap-level anchors, and performs dynamic object removal and change detection. Through submap-level reprojection, the generated evidence remains usable even if the trajectory is subsequently modified by a new session, eliminating the need to recompute the entire process from scratch. SEAM suppresses geometrically unreliable inter-session loop edges using a DOP-based confidence measure. Suppressing unreliable loop edges prevents alignment errors. SEAM also uses a directional voxel-wise evidence model. The model accounts for occupancy patterns that vary with ray direction. Direction-aware evidence separates dynamic objects from environmental changes more precisely. Experiments on a real construction-site dataset and a long-term multi-session dataset show that SEAM achieves higher accuracy and faster processing than existing methods.
Kyuwon Kim
HMG Construction R&D Division, Hyundai Engineering & Construction, Seoul, Republic of Korea