PaperlayerView exact revision on arXiv

Source-linked AI summary

PointLoc: Deep Pose Regressor for LiDAR Point Cloud Localization

Wei Wang, Bing Wang, Peijun Zhao, Changhao Chen, Ronald Clark, Bo Yang, Andrew Markham, Niki Trigoni

arXiv:2003.02392v3cs.RO

TL;DR

LiDAR relocalization needs to estimate absolute poses from point clouds, whose unordered structure complicates direct regression. PointLoc uses a self-attention point-cloud network to predict 6-DoF poses from a single scan, and the paper reports improved relocalization accuracy over prior LiDAR and visual approaches. The authors state that it can support meter-level localization and complement GNSS where GNSS is absent.

  • Problem

    Point-cloud retrieval relocalization requires a reference database and has O(n) search time and storage, while unordered point clouds are difficult for traditional deep regression models to process.

    Source: §I · p. 3

  • Method

    PointLoc uses an end-to-end PointNet-style network with self-attention to predict 6-DoF poses directly from a single LiDAR point cloud.

    Source: §I · p. 3, C. Self-Attention Module · p. 6

  • Results

    PointLoc achieves better relocalization accuracy than previous LiDAR and visual sensor approaches, and improves on PointNetVLAD by 46.47% in translation and 70.45% in rotation.

    Source: C. Results on the Oxford Radar RobotCar · p. 8, §VII · p. 10

  • Takeaways & Limitations

    The authors state that PointLoc can be applied to meter-level relocalization and can complement GNSS in indoor environments or urban areas with high-rise buildings.

    Source: §VII · p. 10

Abstract

from arXiv · show

In this paper, we present a novel end-to-end learning-based LiDAR relocalization framework, termed PointLoc, which infers 6-DoF poses directly using only a single point cloud as input, without requiring a pre-built map. Compared to RGB image-based relocalization, LiDAR frames can provide rich and robust geometric information about a scene. However, LiDAR point clouds are unordered and unstructured making it difficult to apply traditional deep learning regression models for this task. We address this issue by proposing a novel PointNet-style architecture with self-attention to efficiently estimate 6-DoF poses from 360{\deg} LiDAR input frames.Extensive experiments on recently released challenging Oxford Radar RobotCar dataset and real-world robot experiments demonstrate that the proposedmethod can achieve accurate relocalization performance.

I. INTRODUCTION

The introduction motivates LiDAR relocalization as a map-free alternative to approaches that require reference databases or rely on image inputs, and presents PointLoc as an end-to-end regressor for absolute 6-DoF poses. It also reports evaluation on the Oxford dataset and new indoor data, with results favoring PointLoc over state-of-the-art DNN-based LiDAR and visual relocalization methods.

I. INTRODUCTION

  • Point-cloud retrieval approaches require a reference database, with O(n) search time and storage, and are reported to have insufficient recall.The introduction notes that map-based approaches can also be difficult to build and computationally costly as maps grow.Source: §I · pp. 2–3
  • The introduction motivates LiDAR inputs by their 360° geometric coverage, while noting that unordered, unstructured point clouds complicate feature learning.It contrasts these properties with image-based methods, whose inputs have a narrow field of view and are sensitive to environmental change.Source: §I · p. 3
  • PointLoc directly regresses absolute 6-DoF poses from LiDAR point clouds in an end-to-end neural network without a pre-built map.The introduction describes the single-timestamp input and cites results across different Oxford Radar RobotCar environments.Source: Figure 1 · p. 2, §I · p. 3
  • The listed contributions include a self-attention architecture and real-world indoor robot experiments with the released vReLoc dataset.The dataset is described as supporting study of indoor relocalization.Source: §I · p. 3
  • PointLoc is reported to outperform state-of-the-art DNN-based LiDAR and visual relocalization methods on experiments using two new datasets.The introduction also points to its Oxford Radar RobotCar results and indoor robot experiments.Source: Figure 1 · p. 2, §I · p. 3

II. RELATED WORK

The related work reviews learning-based localization, LiDAR odometry, and deep learning on point clouds. It places PointLoc among global pose-regression methods while distinguishing its LiDAR input and relocalization task from prior image-based and relative-pose approaches.

II. RELATED WORK

  • The section surveys learning-based localization, LiDAR odometry, and point-cloud DNN architectures as the related areas framing the paper.It distinguishes odometry, which estimates motion between scans, from the paper's focus on global pose regression.Source: §II · p. 3, A. Visual Sensor Relocalization · p. 3

A. Visual Sensor Relocalization

  • Earlier global pose-regression methods use single or sequential RGB images, while the paper positions its work as improving deep global pose regression with LiDAR.The review notes that translation and rotation accuracy of prior DNN relocalization methods was not yet satisfactory for real-world scenarios.Source: A. Visual Sensor Relocalization · p. 3

B. Learning-based Localization Systems

  • The reviewed learning-based localization systems span radar, cellular, wireless, Kinect, and WiFi applications, whereas this work applies deep learning to LiDAR global localization.The examples include recurrent, kernel, transfer-learning, and domain-adaptation approaches.Source: B. Learning-based Localization Systems · pp. 3–4

C. DNN-based LiDAR Odometry

This section distinguishes LiDAR odometry from relocalization and frames PointLoc as absolute-pose estimation within previously visited environments. It also situates the task in point-cloud learning, where unordered point sets can be processed directly.

D. Deep Learning on Point Clouds

  • PointNet++-related methods process unordered point sets directly, while other reviewed point-cloud networks use voxel embeddings or learn features for detection and segmentation.The cited work covers object detection, part segmentation, and semantic segmentation.Source: D. Deep Learning on Point Clouds · p. 4

III. PROBLEM STATEMENT

  • PointLoc is scoped to absolute 6-DoF relocalization in previously visited areas using one LiDAR frame rather than sequential inputs.The task is defined as estimating an agent's pose within an environment it has already traversed.Source: §III · p. 4, Figure 2 · p. 4

C. DNN-based LiDAR Odometry

  • LiDAR odometry estimates relative poses between consecutive scans, whereas relocalization estimates absolute poses in world coordinates.The section notes that relative-pose estimates accumulate drift and that relocalization assumes agents have previously traversed the scene.Source: Figure 2 · p. 4, C. DNN-based LiDAR Odometry · p. 4

III. PROBLEM STATEMENT

  • Each input is an N-by-3 point cloud, and the regressor maps it to a 6-DoF pose comprising 3-D translation and a 4-D quaternion rotation.The pose is expressed with respect to world coordinates.Source: §III · p. 4

IV. DEEP POINT CLOUD RELOCALIZATION

PointLoc processes fixed-shape LiDAR point clouds with a PointNet-style encoder and a sequence of modules that produces a global pose. Its encoder uses hierarchical point features and can theoretically learn a critical subset relevant to relocalization.

IV. DEEP POINT CLOUD RELOCALIZATION

  • The system combines point-cloud preprocessing, a PointNet-style encoder, self-attention, group-all layers, and a pose regressor.Its input is a down-sampled point cloud of fixed shape (N, 3).Source: §IV · p. 4, Figure 3 · p. 5

A. Point Cloud Pre-Processing

  • Random sampling fixes each input at (N, 3), with N set to 20,480 to retain information near the dataset's average of 21,000 points.The varying number of points in raw LiDAR frames motivates this preprocessing.Source: A. Point Cloud Pre-Processing · p. 5

B. Point Cloud Encoder

  • PointNet-style max pooling can identify a critical subset of points that determines the encoded feature representation.The paper argues PointLoc's PointNet++ structure can learn such a subset from LiDAR point clouds in theory.Source: B. Point Cloud Encoder · pp. 5–6
  • Four consecutive PointNet++ set-abstraction layers use sampling, grouping, and PointNet operations to learn hierarchical point features.The encoder also uses multi-scale grouping and farthest-point sampling to form regions and extract local features.Source: B. Point Cloud Encoder · p. 6

C. Self-Attention Module

PointLoc's self-attention module learns a feature mask intended to suppress noisy features, including those from moving objects, before pose regression.

C. Self-Attention Module

  • A shared MLP and sigmoid produce a 1×C mask from point features, which is broadcast across points and multiplied element-wise with those features.The module is intended to remove outlier features such as those from moving objects.Source: C. Self-Attention Module · p. 6

D. Group All Layers Module

The group-all layers aggregate point features into an embedded vector that is passed onward for pose prediction.

D. Group All Layers Module

  • Group-all layers apply an MLP to the point features, then max-pool them into a C5-dimensional vector for pose regression.The input feature set has size N4 × C4, and the MLP first updates its feature dimension to C5.Source: D. Group All Layers Module · p. 6

E. Pose Regressor

PointLoc's regressor uses separate branches to predict translation and rotation. Its training loss jointly accounts for translation and quaternion-based rotation errors with learnable balancing factors.

E. Pose Regressor

  • Two branches of four fully connected layers separately regress translation and rotation from the embedded feature vector.Leaky ReLU follows each layer except the last, and the fully connected layer sizes decrease gradually.Source: E. Pose Regressor · p. 6

F. Loss Function

  • The training loss combines L1 translation error with L1 error between logarithmic quaternion representations of rotation.Learnable factors β and γ balance the translation and rotation terms.Source: F. Loss Function · p. 6

V. INDOOR LIDAR SENSOR DATASET FOR RELOCALIZATION

The authors introduce the released vReLoc indoor dataset and evaluate PointLoc on it and Oxford Radar RobotCar, using mean outdoor errors and median indoor errors.

V. INDOOR LIDAR SENSOR DATASET FOR RELOCALIZATION

  • vReLoc adds a released indoor dataset combining LiDAR and vision sensors on a mobile robot.The dataset was collected to support research in indoor LiDAR localization.Source: §V · pp. 6–7
  • The experimental robot is Turtlebot 2, mounted with a Velodyne HDL-32E LiDAR featuring 32 lasers, a 40° vertical and 360° horizontal field of view, and a 10 Hz operating frequency.Each point cloud contains about 60,000 points, and a Vicon Motion Tracker system acquires ground-truth 6-DoF poses with approximately 1 cm precision.Source: §V · p. 7
  • The 18 vReLoc sequences cover static scenes, one-person walking, and two-person walking, with sensor streams synchronized to share timestamp poses.The authors selected challenging sequences for training.Source: §V · p. 7

VI. EXPERIMENTS; A. Implementation Details

  • Experiments compare PointLoc with state-of-the-art methods on Oxford Radar RobotCar and vReLoc, training each model for 100 epochs with a 0.001 learning rate.Mean error is reported for outdoor data and median error for indoor data.Source: §VI · p. 7, A. Implementation Details · p. 7

B. Baselines

The evaluation compares PointLoc with LiDAR and camera relocalization baselines, including retrieval, single-image, and sequential approaches.

B. Baselines

  • PointLoc is compared with LiDAR methods PointNetVLAD and DCP, plus camera methods PoseNet17, AtLoc, LSTM-Pose, and MapNet.The authors implemented and tuned the baselines for best performance.Source: B. Baselines · p. 8
  • PointNetVLAD serves as a point-cloud retrieval baseline using a reference database built from validation sequence FULL 5.Its triplet dataset and point count were adjusted while other hyperparameters remained unchanged.Source: B. Baselines · p. 8
  • DCP is adapted from point-cloud registration by using its feature extractor to compare global-pose relocalization performance.The camera baselines include single-image methods and sequential methods that can use temporal constraints.Source: B. Baselines · p. 8

C. Results on the Oxford Radar RobotCar

On Oxford Radar RobotCar, PointLoc improves on LiDAR and camera relocalization baselines, while indoor vReLoc tests show successful real-world localization.

C. Results on the Oxford Radar RobotCar

D. Results on a Real-World Indoor Robot

E. Ablation Study

Ablations indicate gains from self-attention, denser point sampling, and the two-branch pose regressor; the conclusion identifies scope and future directions.

E. Ablation Study

  • Without self-attention, PointLoc's performance decreases by 7.77% in translation and 19.75% in rotation on Oxford Radar RobotCar.The authors attribute improved relocalization accuracy to the self-attention module.Source: E. Ablation Study · p. 10
  • Using PointLoc rather than 4096 Points improves translation by 1.4% and rotation by 9.09%.The comparison tests the effect of sampling more points from raw point clouds.Source: E. Ablation Study · p. 10
  • PointLoc improves over the two-layer Pose Regressor by 73.83% in translation and 84.88% in rotation.The authors link this result to their two-branch MLP pose-regressor design.Source: E. Ablation Study · p. 10

VII. CONCLUSION

  • The conclusion identifies noisy point-feature removal and intensity information as future avenues for improving relocalization accuracy.It frames PointLoc for large-scale relocalization and robot navigation with meter-level localization requirements, including as a complement when GNSS is absent.Source: §VII · p. 10
Loading 2003.02392v3…