PaperlayerView exact revision on arXiv

Source-linked AI summary

PIN-SLAM: LiDAR SLAM Using a Point-Based Implicit Neural Representation for Achieving Global Map Consistency

Yue Pan, Xingguang Zhong, Louis Wiesmann, Thorbjörn Posewsky, Jens Behley, Cyrill Stachniss

arXiv:2401.09101v2cs.ROcs.CV

TL;DR

Neural implicit SLAM systems have struggled to support both online operation and globally consistent loop corrections. PIN-SLAM combines elastic neural points with incremental implicit-map learning, point-to-implicit odometry, and loop closure; experiments report competitive localization alongside compact, consistent maps and accurate mesh reconstruction.

  • Problem

    Many neural implicit SLAM systems lack sensor-frame-rate operation and loop-corrected global consistency, limiting their suitability for online large-scale mapping.

    Source: §I · p. 2, B. Implicit Neural Map Representation · p. 4

  • Method

    PIN-SLAM alternates online learning of a local implicit SDF with correspondence-free point-to-implicit pose estimation, using elastic neural points that transform with loop-corrected poses.

    Source: §I · p. 2, §VI · p. 17

  • Results

    PIN-SLAM achieves better or on-par localization accuracy, builds a more consistent and compact implicit map, reconstructs more accurate and complete meshes, and runs at sensor frame rate.

    Source: §VI · p. 18

  • Takeaways & Limitations

    The implicit map supports arbitrary-position SDF queries for path planning and surface reconstruction.

    Source: §I · p. 2

    The method currently uses a fixed neural-point resolution, which the authors identify as an opportunity to improve reconstruction quality through adaptive point distribution.

    Source: §VI · p. 18

Abstract

from arXiv · show

Accurate and robust localization and mapping are essential components for most autonomous robots. In this paper, we propose a SLAM system for building globally consistent maps, called PIN-SLAM, that is based on an elastic and compact point-based implicit neural map representation. Taking range measurements as input, our approach alternates between incremental learning of the local implicit signed distance field and the pose estimation given the current local map using a correspondence-free, point-to-implicit model registration. Our implicit map is based on sparse optimizable neural points, which are inherently elastic and deformable with the global pose adjustment when closing a loop. Loops are also detected using the neural point features. Extensive experiments validate that PIN-SLAM is robust to various environments and versatile to different range sensors such as LiDAR and RGB-D cameras. PIN-SLAM achieves pose estimation accuracy better or on par with the state-of-the-art LiDAR odometry or SLAM systems and outperforms the recent neural implicit SLAM approaches while maintaining a more consistent, and highly compact implicit map that can be reconstructed as accurate and complete meshes. Finally, thanks to the voxel hashing for efficient neural points indexing and the fast implicit map-based registration without closest point association, PIN-SLAM can run at the sensor frame rate on a moderate GPU. Codes will be available at: https://github.com/PRBonn/PIN_SLAM.

I. INTRODUCTION

PIN-SLAM targets globally consistent neural implicit mapping, using elastic point embeddings to accommodate loop corrections. It combines online map learning, point-to-implicit odometry, and loop closure in a full SLAM system.

I. INTRODUCTION

  • Earlier neural implicit SLAM systems lack direct loop-closure correction for large scenes, largely because regular grid embeddings resist pose corrections.Range-sensor SLAM is a fundamental building block for autonomous mobile robots.Source: §I · pp. 1–2
  • PIN-SLAM uses elastic neural points to support online globally consistent mapping after loop closure, integrating odometry, loop detection, and implicit mapping.The authors describe it as the first full-fledged implicit neural SLAM system with all three capabilities.Source: §I · p. 2
  • Mapping uses voxel-hashed neural feature points and sliding-window data-pool replay, while odometry registers scans to the implicit SDF without point correspondence association.The authors report sensor-frame-rate operation at 10 Hz regardless of mapped-scene scale; a regularity-based robust kernel further improves registration accuracy and robustness.Source: §I · p. 2

A. LiDAR Odometry and SLAM

  • PIN-SLAM detects loops from optimized local-map neural features and transforms neural points with corrected poses to maintain global map consistency.Using local maps for loop detection is described as robust to occlusion and narrow-field-of-view sensors.Source: §I · p. 2, A. LiDAR Odometry and SLAM · p. 3

I. INTRODUCTION

  • PIN-SLAM claims localization accuracy better than or on par with state-of-the-art LiDAR odometry/SLAM, more compact maps, accurate and complete meshes, and sensor-frame-rate operation.The system is also reported to outperform recent implicit neural SLAM methods across datasets and range sensors.Source: §I · p. 2

A. LiDAR Odometry and SLAM

  • Unlike direct LiDAR methods that register scans to local point-cloud maps, PIN-SLAM uses a continuous point-based implicit SDF combining point-map flexibility with neural-map continuity and compactness.The paper identifies its registration as avoiding ICP’s time-consuming, non-differentiable correspondence association.Source: A. LiDAR Odometry and SLAM · p. 3

B. Implicit Neural Map Representation

Prior neural implicit maps combine local features with neural networks, but regular feature grids limit loop-corrected global consistency and online use. PIN-SLAM uses elastic point features in a range-sensor SLAM pipeline, with loop correction transforming the neural map alongside its poses.

B. Implicit Neural Map Representation

  • Neural implicit maps combine explicit local latent features with a neural network, and prior LiDAR systems use octree grids or integrate neural mapping with ICP odometry.The paper places PIN-SLAM among point-based implicit neural approaches rather than grid-based feature maps.Source: B. Implicit Neural Map Representation · pp. 3–4
  • PIN-SLAM stores optimizable local features as neural points, using their elasticity for loop corrections rather than only their spatial flexibility for adaptive scene encoding.Prior hybrid implicit representations place local features in structures including voxel grids, octrees, tri-planes, or point sets.Source: B. Implicit Neural Map Representation · pp. 3–4
  • Previous neural implicit SLAM systems often miss sensor-frame-rate operation and loop-corrected global consistency; regular feature grids require reallocation and map retraining after pose updates.PIN-SLAM addresses these challenges with direct pointwise SDF supervision, second-order odometry optimization, and elastic point features.Source: B. Implicit Neural Map Representation · p. 4

C. Global Consistency in SLAM

  • Unlike submap-based consistency methods, PIN-SLAM avoids submap division and overlap disambiguation while retaining a continuous SDF for efficient 3D reconstruction.Submaps reduce pose-graph size but introduce overlap ambiguity and choices about where to divide submaps; point-based maps can be less suitable for reconstruction and planning.Source: C. Global Consistency in SLAM · p. 4

III. OUR APPROACH TO GLOBALLY CONSISTENT IMPLICIT NEURAL SLAM

  • PIN-SLAM targets large-scale LiDAR SLAM, also applies to RGB-D, and alternates pose estimation with online mapping before loop detection and pose-graph correction.After pose-graph optimization, it transforms neural points and the training pool with corrected frames; the implicit SDF can be queried for mesh reconstruction.Source: §III · pp. 4–5
  • The pipeline downsamples scans separately for registration and mapping, learns local-map features from sampled SDF data, and uses local polar-context descriptors to propose loop closures.A candidate loop is verified by registering the current scan against the candidate frame’s local map before adding a pose-graph edge.Source: §III · p. 5, Figure 2 · p. 5

A. Neural Point-based Map Representation

PIN-SLAM represents the map with optimizable neural points that encode local geometry. A shared decoder predicts SDF values from each point’s features and the query position in that point’s coordinate frame, enabling elastic map adjustment and weighted predictions from neighboring points.

A. Neural Point-based Map Representation

  • Each neural point stores a local geometry feature and coordinate frame, along with creation, update, and stability information.The map is a set of neural points defined in the world frame; each point is linked to a sensor pose for map manipulation.Source: 1) Map Representation · p. 5
  • A shared shallow decoder predicts each nearby neural point’s SDF from its latent feature and the query position in that point’s coordinate system.The query position is transformed into the neural point’s local coordinates before decoding.Source: 1) Map Representation · pp. 5–6
  • Relative coordinates make SDF predictions invariant to local translation and rotation, allowing neural points to move with pose corrections after loop closure.This coordinate choice supports the map’s elasticity during loop-closure correction.Source: 1) Map Representation · p. 6
  • The final SDF prediction is an inverse-distance-weighted interpolation of the predictions from the K nearest neural points.The map also defines query-position stability as a distance-weighted mean of the nearby neural points’ stability values.Source: 1) Map Representation · p. 6
  • Neighboring neural points act as local DeepSDF-like models whose features represent local geometry, while a shallow shared decoder interprets those features.Unlike the deeper decoder used in DeepSDF [53], this decoder need not be pretrained to represent an object globally.Source: 1) Map Representation · p. 6

2) Map Data Structure:

PIN-SLAM uses voxel hashing to index active neural points and find nearby points for SDF prediction. It adds or replaces neural points when a voxel is empty, a hash collision places a stored point far away, or that point becomes inactive.

2) Map Data Structure:

  • Voxel hashing maps positions to hash-table entries holding neural-point indices, with at most one active neural point per voxel.The structure supports fast indexing and neighborhood search at a fixed voxel resolution.Source: 2) Map Data Structure · p. 6
  • Neighborhood search inspects voxels around the query, then selects the K nearest neural points for SDF prediction.A larger voxel neighborhood increases receptive field but also increases search computation; interpolation can use fewer than K available points.Source: 2) Map Data Structure · p. 6
  • Fig. 3(a) shows each neighboring neural point predicting the query’s SDF from its feature and the query’s local coordinate through a shared decoder.The surrounding points contribute individual predictions at the gray query position.Source: Figure 3 · p. 6

3) Map Initialization and Update:

  • A new neural point is initialized when its hash entry is empty, a collision leaves the stored point far away, or the stored point is inactive.An inactive point has exceeded a travel-distance threshold since its last update; replaced points remain in the global map for loop-closure correction and map adjustment.Source: 3) Map Initialization and Update · p. 6

4) Local Map:

The local map combines spatial and travel-distance windows to avoid registration against inconsistent historical observations. Training samples are drawn along scan rays, replayed from a bounded local pool, and filtered to support incremental map learning.

4) Local Map:

  • Registration uses an active local map to avoid alignment with inconsistent historical observations caused by odometry drift.The local map includes neural points within both a spatial radius and a travel-distance window, rather than using only a spatial or temporal window.Source: 4) Local Map · pp. 6–7

5) Map Training Samples:

  • Each scan ray supplies its measured endpoint, Gaussian-sampled near-surface points, free-space samples in front, and truncated free-space samples behind the surface.The projective SDF target is measured range minus sample depth; it is an overestimate but a quick approximation near the surface.Source: 5) Map Training Samples · p. 7
  • Training samples retain their sensor-frame timestep so each sample can later be transformed into the world frame.At each timestep, samples are collected along rays from the voxel-downsampled mapping scan.Source: 5) Map Training Samples · p. 7
  • A replay pool retains historical training samples to address catastrophic forgetting and removes samples outside a local sliding window.If the pool exceeds its sample cap, samples are randomly retained, including when the robot stops; Fig. 4(c) depicts the pool with points colored by SDF target.Source: 5) Map Training Samples · p. 7, Figure 4 · p. 7

6) Map Training Losses:

PIN-SLAM incrementally trains local neural-point features from a replay pool using BCE and Eikonal losses. The BCE term fits projective SDF targets, while the Eikonal term regularizes the SDF gradient.

6) Map Training Losses:

  • The BCE loss provides fast convergence, fits projective SDF values through logistic regression, and softly truncates distances near surfaces.Its truncation smoothness is adjusted with the scaled sigmoid’s σt parameter.Source: 6) Map Training Losses · p. 7
  • The Eikonal loss enforces the SDF property that the gradient norm is 1.PIN-SLAM estimates numerical gradients by perturbing positions, enabling updates beyond the local hash grid and producing smoother gradients for the loss.Source: 6) Map Training Losses · p. 7
  • The final training objective combines BCE and Eikonal terms, and optimization updates the latent features of involved local-map neural points.Training samples are randomly drawn in batches from the pool, and nearby neural points’ update times and stability are also updated.Source: 6) Map Training Losses · pp. 7–8

B. Preprocessing

Preprocessing creates separate point clouds for mapping and registration and deskews LiDAR scans. Odometry aligns the registration cloud to the local implicit map without point correspondences, then checks the solution before using it for mapping.

B. Preprocessing

  • Separate voxel-downsampled clouds serve mapping and registration, while LiDAR scans are deskewed before registration and after odometry, respectively.The mapping cloud uses a smaller voxel size; registration points are deskewed using constant-velocity motion prediction.Source: B. Preprocessing · p. 8

C. Odometry Estimation

  • Correspondence-free scan-to-implicit-map registration minimizes SDF error using second-order optimization, avoiding explicit point associations.A Levenberg–Marquardt solver starts from a constant-velocity estimate.Source: C. Odometry Estimation · p. 8
  • Geman–McClure weights down-weight SDF residuals and gradient anomalies associated with noise, dynamic objects, and underfitted map details.The two pointwise weights scale the diagonal elements of the optimization weight matrix.Source: C. Odometry Estimation · p. 8
  • Odometry is accepted only when residual, valid-point ratio, and minimum Hessian-eigenvalue checks pass; otherwise the constant-velocity pose is used and mapping is skipped.The correction threshold and maximum iteration count also control optimization termination.Source: C. Odometry Estimation · pp. 8–9

D. Mapping and Local Bundle Adjustment

PIN-SLAM updates its implicit map from static measurements and periodically refines recent poses and neural point features together. This local bundle adjustment uses the implicit SDF rather than extracted geometric features.

D. Mapping and Local Bundle Adjustment

  • Dynamic points are filtered before mapping by treating measurements in stable free space as dynamic, leaving static points for map updates.The filter uses SDF and stability predictions with distance and stability thresholds.Source: D. Mapping and Local Bundle Adjustment · p. 9
  • Incremental mapping samples rays to static points, initializes nearby neural points, and optimizes pooled training samples; after Fmlp timesteps, the decoder is frozen while neural point features continue training.The decoder and local-map features are jointly optimized during the first Fmlp timesteps.Source: D. Mapping and Local Bundle Adjustment · p. 9
  • Every Fba timesteps, local bundle adjustment jointly optimizes recent sensor poses and local neural point features using endpoint samples and an L2 SDF loss.The current state initializes the joint optimization, which aims to make the local map and poses more consistent.Source: D. Mapping and Local Bundle Adjustment · p. 9
  • After map optimization and local bundle adjustment, the updated local map is assigned back to the global map..Source: D. Mapping and Local Bundle Adjustment · p. 9

E. Loop Closure Detection

PIN-SLAM searches first for local loop candidates and then uses neural-point descriptors for global place recognition and an initial relative pose. Verified loops drive pose-graph optimization, with neural points transformed alongside their associated frames.

E. Loop Closure Detection

  • Local loop candidates are screened using a distance-based criterion on current and sufficiently separated historical scan positions..Source: E. Loop Closure Detection · p. 9
  • Global loop retrieval encodes average neural-point features in polar bins, uses a rotation-invariant ring key to retrieve candidates, and estimates yaw by sector shifts.A cosine-distance threshold determines whether a candidate is accepted.Source: E. Loop Closure Detection · p. 9
  • Polar-context augmentation generates descriptors at virtual positions along a direction perpendicular to motion to address limited translational invariance.The most similar historical descriptor supplies the translation component of semi-metric localization.Source: E. Loop Closure Detection · pp. 9–10

F. Globally Consistent Implicit Neural Map Adjustment

IV. EXTENSION TO RGB-D OR METRIC-SEMANTIC SLAM

The evaluation reports localization and mapping performance across range-sensor settings, while the paper also demonstrates extensions for color and semantic information. PIN-SLAM’s reported results include comparisons with LiDAR and neural implicit SLAM approaches and sensor-frame-rate operation on an NVIDIA A4000 GPU.

IV. EXTENSION TO RGB-D OR METRIC-SEMANTIC SLAM

  • The extension predicts color and semantic values from nearby neural points, adds weighted L1 color and cross-entropy semantic losses, and can use color differences as an additional odometry constraint.Reconstructed SDF meshes can also be colorized or annotated by querying values at mesh vertices.Source: §IV · p. 10

V. EXPERIMENTAL EVALUATION

  • PIN-SLAM reports localization accuracy better than or on par with state-of-the-art LiDAR odometry and SLAM, greater accuracy than recent implicit neural SLAM, compact maps, and sensor-frame-rate operation on one NVIDIA A4000 GPU.The experiments also report large-scale globally consistent mapping with loop closure and accurate, complete mesh reconstruction.Source: §V · p. 11

A. Experimental Setup

The evaluation spans varied LiDAR and RGB-D datasets and tests PIN-LO against established odometry methods on KITTI, accounting for variation across repeated runs.

1) Datasets:

  • Evaluation datasets cover driving LiDAR, handheld indoor and outdoor LiDAR, and synthetic RGB-D, with reference poses or maps drawn from GNSS-INS, TLS, or tracking systems.Source: 1) Datasets · p. 11

2) Parameters and Implementation Details:

  • Length-based parameters scale with each sensor’s maximum range; online map training uses 15 iterations, increased to 600 for first-frame initialization.The approach uses Adam with learning rate 0.01 and batch size 16384 for online map training.Source: 2) Parameters and Implementation Details · p. 11

1) LiDAR Odometry Evaluation on KITTI:

  • The KITTI odometry comparison uses average relative translational error and includes feature-based, dense point-map, other explicit-map, and learning-based methods.Supervised learning baselines train on sequences 00–06 and test on 07–10; PIN-LO and Nerf-LOAM use no pre-training.Source: 1) LiDAR Odometry Evaluation on KITTI · pp. 11–12
  • 0.51% average translation error (standard deviation 0.02%) puts PIN LiDAR odometry on par with KISS-ICP and CT-ICP and ahead of all learning-based methods.Results are the mean and standard deviation over 10 runs with different random seeds.Source: 1) LiDAR Odometry Evaluation on KITTI · p. 12, Table III · p. 12
  • Against Nerf-LOAM, PIN-LO performs better, which the authors attribute to improved close-to-surface free-space SDF training and more robust Levenberg–Marquardt registration with robust kernels.Source: 1) LiDAR Odometry Evaluation on KITTI · p. 13
  • PIN LiDAR odometry ranks as the best learning-based method on KITTI hidden sequences 11–21 and performs among the leaderboard’s best LiDAR odometry/SLAM systems.Source: 1) LiDAR Odometry Evaluation on KITTI · p. 13

2) LiDAR SLAM Evaluation:

PIN-SLAM’s loop-closure system improves global trajectory consistency across LiDAR datasets, while its RGB-D evaluation shows accurate camera tracking and high-fidelity mesh reconstruction.

2) LiDAR SLAM Evaluation:

  • 1.0 m average ATE RMSE on KITTI sequences with loops and 1.2 m across all eleven sequences make PIN-SLAM competitive with leading loop-closing systems, including an offline post-processing method.The loop-sequence RMSE falls from 3.2 m with PIN LiDAR odometry to 1.0 m with PIN-SLAM; results average 10 runs.Source: 2) LiDAR SLAM Evaluation · p. 13
  • The KITTI qualitative comparison shows loop corrections removing duplicated structures from the locally consistent PIN-LO mesh to form a globally consistent PIN-SLAM mesh.Source: 2) LiDAR SLAM Evaluation · p. 13
  • Across nine MulRan and four IPB-Car sequences, PIN-SLAM corrects PIN LiDAR odometry drift and achieves the best overall localization accuracy against the compared approaches.Source: 2) LiDAR SLAM Evaluation · p. 14
  • PIN-SLAM yields the smallest localization error on Newer College and Hilti-21; on the challenging Newer College stairs sequence, its RMSE is 6 cm while half the compared methods fail.Hilti-21 sequences are short and mostly confined, without explicit loops.Source: 2) LiDAR SLAM Evaluation · p. 14

3) RGB-D SLAM Evaluation:

  • 2.4 mm average RMSE gives PIN-SLAM the best camera-tracking accuracy across eight Replica RGB-D sequences, where loop closure is disabled because each camera trajectory remains within the local map.Local bundle adjustment and color information reduce localization RMSE by 40% and 31%, respectively; PIN-SLAM also reconstructs high-fidelity colorized meshes.Source: 3) RGB-D SLAM Evaluation · p. 14, Table IX · p. 14, Figure 6 · p. 14

4) Additional Challenging Scenarios:

PIN-SLAM is evaluated across challenging environments, loop detection, reconstruction quality, map consistency, design ablations, and memory use. The results support globally consistent mapping, strong reconstruction quality, and a compact representation.

4) Additional Challenging Scenarios:

  • PIN-SLAM builds a globally consistent cave map from quadruped data and filters moving pedestrians to reconstruct a static mesh in highly dynamic ETH DOALS scenes.The cave example uses the Nebula dataset.Source: 4) Additional Challenging Scenarios · p. 14

1) 3D Reconstruction Quality:

  • PIN-SLAM achieves the best completeness, Chamfer distance, and F-score in both Newer College scenes, outperforming the compared reconstruction methods on these measures.The evaluation uses Quad and Math Institute, with TLS reference models and a 20 cm mesh voxel size.Source: Table XI · p. 15, 1) 3D Reconstruction Quality · p. 16

C. Evaluation of the Loop Closure Detection Performance

  • Without offline pre-training, PIN-SLAM achieves the second-best average KITTI loop-detection recall, and online neural-point features improve descriptor distinctiveness.Its polar context descriptor also improves recall over the original scan context.Source: C. Evaluation of the Loop Closure… · p. 15

2) Map Consistency:

E. Ablation studies

  • A 0.4 m neural-point resolution yields the smallest ARTE on all three tested sequences, while the ablations also favor BCE mapping loss and the stated odometry design choices.The reported choices include numerical gradients for Eikonal loss, analytical gradients for fitted-map odometry, and GM robust kernels for SDF residuals and gradient anomalies.Source: Table XIV · p. 16, E. Ablation studies · p. 16, Table XII · p. 16

1) Memory Requirement:

  • The PIN map averages about 0.7% of raw point-cloud memory and is about 15% more memory efficient than SHINE at the same resolution and feature dimensions.Loop correction and map updates reduce PIN map memory by about 20% versus PIN LiDAR odometry by eliminating duplicated neural points.Source: Table XV · p. 16, 1) Memory Requirement · pp. 16–17

2) Computational Requirement:

The light version of PIN-SLAM runs at approximately 11 Hz, while the conclusion reports localization accuracy on par with or better than previous methods and compact, consistent maps. The paper also demonstrates metric-semantic mapping, with semantics not yet used for odometry or loop closure.

2) Computational Requirement:

  • The light version runs at approximately 11 Hz and is 38% faster at a 10% localization-accuracy cost; Nerf-LOAM takes more than 4 seconds per frame, 30× slower.Odometry and map optimization each take about 40% of light-version runtime; loop detection and pose-graph optimization can occasionally push per-frame time above 200 ms.Source: Table XVI · p. 17, 2) Computational Requirement · p. 17, Figure 13 · p. 17

G. Extension on Semantic Mapping

  • PIN-SLAM builds a metric-semantic map on SemanticKITTI using an additional shallow MLP, but semantics do not currently contribute to odometry or loop closure.The authors identify using semantics in those tasks as a possible future extension.Source: G. Extension on Semantic Mapping · p. 17, Figure 14 · p. 18

VI. CONCLUSION

  • PIN-SLAM achieves better or on-par localization accuracy than previous methods and builds more consistent, compact implicit maps that reconstruct into more accurate and complete meshes.The conclusion also reports operation at sensor frame rate using a moderate GPU.Source: §VI · pp. 17–18

Related papers

Loading 2401.09101v2…