Multi-LiDAR SLAM Pipeline

• Multi-LiDAR Localization and Mapping Pipeline is a system that integrates multiple LiDAR sensors with inertial and GNSS data to perform robust SLAM and overcome single-sensor limitations.
• The pipeline employs precise temporal and spatial calibration, feature extraction, and real-time sensor fusion to ensure highly accurate mapping in diverse environments.
• Optimization through factor graphs and decentralized architectures enables real-time state estimation and scalability, addressing challenges like occlusion and sensor misalignment.

A Multi-LiDAR Localization and Mapping Pipeline integrates data from multiple LiDAR sensors, optionally fused with inertial, odometric, and GNSS measurements, to achieve robust and accurate simultaneous localization and mapping (SLAM) in complex environments. These systems leverage the complementary coverage and redundancy provided by multiple LiDARs, addressing the field-of-view, occlusion, and degeneration limitations inherent to single-sensor setups. Modern pipelines combine precise temporal/spatial calibration, feature extraction, real-time synchronization, multi-sensor fusion, and optimization-based estimation within scalable architectures.

1. System Architectures and Sensor Integration

Multi-LiDAR SLAM pipelines are architected to process data from spatially distributed and temporally misaligned sensors. Canonical systems, such as GM-Livox, deploy six non-repetitive scanning Livox LiDARs, a high-rate IMU (200 Hz), a wheel encoder, and a GNSS/RTK receiver. Preprocessing steps enforce stringent sensor synchronization, typically via hardware signals (GNSS PPS) and windowed software alignment. Point-level timestamps from all LiDARs are aligned to form temporally coherent "fused sweeps" within fixed intervals $[t_k, t_{k+1})$, with the offset $\Delta t_{L-I}$ for non-primary LiDARs estimated via minimization:

$$\Delta t_{L-I} = \arg\min_{\delta} \|(t^{\text{LiDAR}}_i + \delta) - t^{\text{IMU}}_j\|^2$$

Feature points—edges and planes—are extracted individually per sensor and then merged into the fused sweep $F_k$, preserving geometric integrity across devices (Wang et al., 2021).

Alternative architectures, such as the multi-modal indoor pipeline (Qingqing et al., 2023), incorporate both spinning and solid-state LiDARs. These systems implement spatial-temporal calibration (GICP-based) and undistortion pipelines to handle asynchronous, low-FoV sensors. The resulting point clouds are merged after deskewing using interpolated IMU trajectories.

2. Feature Extraction, Deskewing, and Scan Fusion

Each LiDAR scan undergoes ground segmentation using RANSAC (to extract dominant ground planes $m = [n^\top, d]^\top$), followed by feature extraction. Planar features are identified among neighboring points using

$d_{\text{planar}} = |n^\top (p_i - p_j)|$

while edge features exploit smoothness or curvature metrics, such as

$s_i = \|(p_{i+1} - p_i) \times (p_i - p_{i-1})\|$

After per-sensor feature extraction, alignment of precise timestamps (potentially through interpolation in continuously sampled IMU pose graphs) allows deskewing of each point (Sauerbeck et al., 2023, Wang et al., 2021, Qingqing et al., 2023). The temporally corrected per-LiDAR features (or whole deskewed scans) are then merged in a common vehicle or IMU frame by applying calibrated extrinsics $T_{s \rightarrow \text{base}}$.

Table: Feature Extraction and Fusion Steps | Step | Key Methodology | References | |------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------|--------------------| | Ground segmentation | RANSAC plane fit $m = [n^\top, d]^\top$ | (Wang et al., 2021) | | Edge/plane extraction | Smoothness, curvature, continuity, eigen-analysis | (Wang et al., 2021, Qingqing et al., 2023) | | Deskewing | IMU-based interpolation, pre-integration of $\Delta R$, $\Delta v$, $\Delta p$ | (Wang et al., 2021, Qingqing et al., 2023) | | Time-aligned fusion | Stamp features with $t_k$ (windowed), transform to body/world frame | (Wang et al., 2021, Qingqing et al., 2023) | | Noise/outlier handling | Feature fitness filters, removal of near-sensor points, voxel downsampling | (Nguyen et al., 2021, Sauerbeck et al., 2023) |

Fusion proceeds by merging per-sensor feature sets into a single scan or maintaining distinct scans per sensor and fusing at later pipeline stages (EKF, pose graph, map update). This approach maximizes feature diversity and robustness.

3. State Estimation, Factor Graphs, and Optimization

The pipeline's core is a keyframe-based, sliding-window estimator, often implemented as a factor graph that tightly fuses IMU, LiDAR, odometric, and GNSS/RTK constraints. Each keyframe maintains a state vector:

$$x_k = [q_k,\, t_k,\, v_k,\, b_{a_k},\, b_{\omega_k},\, s_k]$$

where $q_k \in SO(3)$ (orientation), $t_k \in \mathbb{R}^3$ (position), $v_k$ (velocity), $b_a$, $b_\omega$ (IMU biases), and $s_k$ (encoder scale).

The factor graph objective, aggregated over a window $K$, is

$$J(X) = \|r_0\|^2 + \sum_{k=0}^{K-1} \|r_{\mathrm{IMU}}(x_k, x_{k+1})\|^2_{\Sigma_{\mathrm{IMU}}} + \sum_{k=1}^K \|r_L(x_k)\|^2_{\Sigma_L} + \sum_{k=1}^K \|r_g(x_k)\|^2_{\Sigma_g} + \sum_{k \in RTK} \|r_{RTK}(x_k)\|^2_{\Sigma_{GNSS}}$$

with individual residuals for IMU preintegration, LiDAR scan-matching, ground constraints, and GNSS.

Optimization is conducted via nonlinear solvers (e.g., Gauss-Newton, Levenberg–Marquardt, or Ceres/GTSAM), with thread-level parallelization: LiDAR front-end, IMU pre-integration, and factor assembly/solve are dispatched to separate threads, enabling real-time performance ($\approx$10 Hz on Intel i9) (Wang et al., 2021, Nguyen et al., 2021).

4. Online Calibration, Decentralization, and Uncertainty Modeling

Robust calibration of inter-sensor extrinsics is essential. Pipelines such as M-LOAM and decentralized EKF-based frameworks (Jiao et al., 2020, Lin et al., 2020) estimate extrinsic parameters online as part of the state. The full state vector $x$ includes pose, velocity, and for each LiDAR $i$, extrinsic parameters $x^{e\,i} = [r^{e\,i}, t^{e\,i}]$. Pseudo-measurements constrain their sum to zero for observability.

Decentralized architectures assign one EKF per LiDAR, each sharing the full state and covariance; nodes broadcast updated estimates peer-to-peer. LOAM-style feature registration acts as the measurement model for each LiDAR. This design is resilient to node failure and scalable to large sensor arrays.

Uncertainty propagation in mapping leverages per-feature covariances $\bm\Sigma_p$ (sensor, pose, and extrinsic noise) to weight residuals and filter outliers. Probabilistic voxel downsampling further reduces map size without degrading consistency (Jiao et al., 2020).

5. Mapping, Keyframe Management, and Loop Closure

Optimized poses dictate insertion of fused feature clouds into a global point-cloud map, implemented as a voxel grid (typ. $4$ cm–$10$ cm voxels) or sparse PCL point cloud. Keyframe selection criteria are generally:

• Translation threshold (e.g., $\|t_{k+1} - t_k\| > 10$ m in GM-Livox, $>1$ m in MILIOM)
• Rotation threshold (e.g., $\angle(q_k, q_{k+1}) > 2^\circ$)
• Quality metrics (minimum match ratio, voxel occupancy, or fitness) (Wang et al., 2021, Nguyen et al., 2021, Blanco-Claraco, 2024)

Mapping pipelines may include periodic outlier cleaning, probabilistic downsampling, or interactive SLAM loop-closure modules. Loop closure is performed via ICP or Scan Context descriptors, with global pose graph optimization via robust kernels to ensure consistency. In multi-robot or large-scale tunnel scenarios, outlier rejection heuristics (e.g., bounding-box tests for degenerate keyframes) and Pairwise Consistency Maximization (PCM) are critical for loop-edge pruning (Lauro et al., 29 Jul 2025).

6. System Performance and Practical Considerations

Experimental evaluations demonstrate sub-decimeter to sub-meter RMS errors across urban, indoor, and GNSS-denied scenarios when leveraging multiple LiDARs.

Table: Selected Performance Metrics | System / Paper | Sensors | RMS Error | Update Rate | CPU Threads / Utilization | Notable Features | |--------------------|----------------------------|-------------------|------------------|-------------------------------|---------------------------------------------------| | GM-Livox (Wang et al., 2021) | 6 Livox LiDARs, 200Hz IMU | $\approx$0.4 m (5 km) | 10 Hz | 4 threads / 60% | Real-time multi-threaded sliding-window backend | | MILIOM (Nguyen et al., 2021) | 2 spinning LiDARs, IMU | 0.05–0.08 m ATE | 10 Hz | multi-threaded | Fused CFC feature clouds, sliding-window F-G | | M-LOAM (Jiao et al., 2020) | up to 4 LiDARs | $<$0.05 m ATE | 10 Hz | multi-threaded | Online extrinsic calibration, uncertainty-aware | | Autoware (Sauerbeck et al., 2023) | 4 LiDARs, GNSS-IMU | $\leq$ 0.5 m | 20 Hz | $\sim$2 CPU for LiDAR + EKF | Robust per-LiDAR ICP, ROS 2 integration | | Decentralized EKF (Lin et al., 2020) | 5 LiDARs | $\leq$0.2% drift | 10 Hz | 5 nodes/threads | Node-level redundancy, online extrinsic refinement|

These architectures achieve robust performance in feature-poor, dynamic, and large-scale settings. Configuration flexibility (e.g., YAML-driven block-architecture (Blanco-Claraco, 2024)), seamless multi-LiDAR fusion, and real-time constraints are recurring system requirements. Online and scenario-specific tuning (e.g., dynamic ICP thresholds, per-sensor maximum range, or GICP fitness) are leveraged for optimal results, with continuous recommendations for integrating learning-based segmentation or adaptive calibration strategies where appropriate (Sauerbeck et al., 2023, Jiao et al., 2020).

7. Limitations, Extensions, and Open Challenges

Despite their maturity, current multi-LiDAR SLAM pipelines confront challenges related to:

• Temporal/sync misalignments and extrinsic drift in aggressive or high-vibration runs
• Robust handling of degenerate environments (repetitive, low-structure tunnels)
• Data association ambiguities, especially for loop closures in collaborative/multi-robot deployments
• Scalability of back-end optimization and mapping with increasing sensor count
• Integration of deep-learning-based semantic modules (for richer semantic maps (Sauerbeck et al., 2023))
• Dynamic object filtering, sensor dropout mitigation, and redundancy management

Adaptive keyframe selection, automated calibration modules, and hardware-aware computational strategies are actively studied as mitigations (Lauro et al., 29 Jul 2025, Blanco-Claraco, 2024).

A plausible implication is that as LiDAR array size and heterogeneity grow, fusion backends will likely shift towards distributed, resource-aware, and graph-oriented architectures, with data-driven components for map representation, feature abstraction, and data association.

---

References

• GM-Livox: An Integrated Framework for Large-Scale Map Construction with Multiple Non-repetitive Scanning LiDARs (Wang et al., 2021)
• MILIOM: Tightly Coupled Multi-Input Lidar-Inertia Odometry and Mapping (Nguyen et al., 2021)
• Multi-LiDAR Localization and Mapping Pipeline for Urban Autonomous Driving (Sauerbeck et al., 2023)
• Robust Odometry and Mapping for Multi-LiDAR Systems with Online Extrinsic Calibration (Jiao et al., 2020)
• A decentralized framework for simultaneous calibration, localization and mapping with multiple LiDARs (Lin et al., 2020)
• A flexible framework for accurate LiDAR odometry, map manipulation, and localization (Blanco-Claraco, 2024)
• Robust Multi-Modal Multi-LiDAR-Inertial Odometry and Mapping for Indoor Environments (Qingqing et al., 2023)
• Multi-robot LiDAR SLAM: a practical case study in underground tunnel environments (Lauro et al., 29 Jul 2025)