PaperlayerView exact revision on arXiv

Source-linked AI summary

RadarSLAM: Radar based Large-Scale SLAM in All Weathers

Ziyang Hong, Yvan Petillot, Sen Wang

arXiv:2005.02198v1cs.RO

TL;DR

Robust SLAM in extreme weather remains an open problem. RadarSLAM builds a full radar-based graph SLAM system with geometry-aware tracking, probabilistic point-cloud generation, and graph optimization, and experiments show reliable localization and mapping across adverse weather conditions.

  • Problem

    Robust large-scale SLAM in extreme weather conditions remains an open research problem across sensor modalities.

  • Method

    RadarSLAM uses radar geometry and graph SLAM with feature matching, probabilistic point-cloud generation, pose tracking, local mapping, loop closure detection, and pose graph optimization.

  • Results

    RadarSLAM reliably operates and localizes in dark night, dense fog, heavy snowfall, and other adverse-weather sequences, while achieving 2.1854% translation error and 0.0071 deg/m rotation error in the reported comparison.

  • Takeaways & Limitations

    Radar can support reliable large-scale online SLAM and globally consistent mapping in adverse weather conditions.

Abstract

from arXiv · show

Numerous Simultaneous Localization and Mapping (SLAM) algorithms have been presented in last decade using different sensor modalities. However, robust SLAM in extreme weather conditions is still an open research problem. In this paper, RadarSLAM, a full radar based graph SLAM system, is proposed for reliable localization and mapping in large-scale environments. It is composed of pose tracking, local mapping, loop closure detection and pose graph optimization, enhanced by novel feature matching and probabilistic point cloud generation on radar images. Extensive experiments are conducted on a public radar dataset and several self-collected radar sequences, demonstrating the state-of-the-art reliability and localization accuracy in various adverse weather conditions, such as dark night, dense fog and heavy snowfall.

I. INTRODUCTION

Robust large-scale SLAM in extreme weather remains challenging. RadarSLAM addresses this with radar-geometry-based matching, probabilistic point-cloud generation, and graph optimization, supported by experiments in adverse conditions.

  • Large-scale outdoor SLAM remains difficult in adverse weather, motivating radar-based localization and mapping for conditions such as heavy snowfall.
  • RadarSLAM combines radar geometry with graph SLAM for robust localization and mapping in large-scale outdoor environments.
  • Its feature-matching and pose-tracking algorithm uses radar geometry and graph representation to improve reliability.
  • A probabilistic point-cloud generation method reduces speckle noise in radar-derived point clouds.
  • The full graph-optimization SLAM system is designed to operate in adverse weather conditions, including dense fog and heavy snowfall.

II. RADAR SENSOR BACKGROUND

FMCW radar provides metric range information and can represent scans in polar or Cartesian form, but its measurements contain substantial noise that complicates SLAM.

  • FMCW radar measures target distance and can estimate velocity from the frequency difference between transmitted and received signals.
  • Radar readings suffer from range and angular errors, false detections, clutter, sidelobes, multipath reflections, and receiver saturation.
  • A 360° FMCW radar sweeps Ns azimuth angles, with azimuth step size 2π/Ns, and collapses each return into a range distance.
  • Radar scans can be transformed from polar coordinates into Cartesian grayscale images, preserving absolute metric range information.
  • For a polar point (a, r), the Cartesian coordinate uses θ = 2π · a/Ns and scaling factor γ; this work uses Cartesian representation.

III. RADAR BASED SLAM

RadarSLAM estimates radar poses online with keyframe-based tracking, using radar geometry and graph-based outlier rejection to obtain reliable feature correspondences and pose updates.

  • Pose Tracking: RadarSLAM estimates each current radar pose relative to a known keyframe pose using a relative transformation in SE(2).The system continuously tracks the current frame pose in the world coordinate system.
  • Pose Tracking: Radar geometry restricts correspondence searches using motion priors and rejects inconsistent matches through pairwise consistency constraints.The search-radius restriction reduces incorrect matches and computation, while pairwise consistency filters outliers based on similar motion tendencies.
  • Pose Tracking: The maximum mutually consistent match set is represented as a graph maximum clique and used to compute the relative transformation with SVD.The rejected matches are outliers outside the maximum clique.
  • Pose Tracking: The final tracked pose minimizes reprojection errors over successfully matched current-frame and keyframe feature pairs.Map-point coordinates are expressed in the world coordinate system.
  • Pose Tracking: New keyframes are generated using criteria involving the minimum number of matches and translation and rotation relative to the current keyframe.These criteria follow a similar strategy to visual SLAM.

B. Local Mapping

Local mapping refines poses and map consistency through local bundle adjustment, while radar-derived peak processing supports structure-preserving map information by suppressing noise.

  • Local Mapping: Local mapping jointly optimizes nearby keyframe poses and observable map-point locations after a new keyframe is created.It runs in parallel with the pose-tracking thread and uses local Bundle Adjustment.
  • Radar Point-Cloud Processing: The proposed radar point-cloud extraction preserves environmental structure while suppressing multipath-reflection and speckle-noise detections.Compared with local-maxima detection, the proposed extraction avoids the large number of noise peaks shown in the radar scan.
  • Local Mapping: The optimization minimizes a weighted sum of squared residual errors using the Levenberg–Marquardt method.The state contains keyframe poses and map-point locations, and the weighting matrix is symmetric positive definite.
  • Local Mapping: Map points created by the latest keyframe are culled when they are not observable by more than two keyframes.This limits the computation required for local mapping.

C. Loop Closure Detection

RadarSLAM detects loop closures from radar point-cloud structure rather than relying on image-style Bag-of-Words descriptors, using probabilistic peak selection and a global point-cloud descriptor.

  • Loop Closure Detection: Bag-of-Words is considered inadequate for radar loop closure detection because radar images have repetitive descriptors and multipath reflections.These properties make radar image pixels less distinctive than optical-image pixels.
  • Point-Cloud Generation: The radar point-cloud conversion procedure takes a radar image and outputs a point cloud using minimum peak-prominence and peak-distance parameters.It processes each azimuth row, detects peaks, transforms them into points, and adds them to the output cloud.
  • Point-Cloud Generation: Radar images are converted into point clouds to capture the spatial signature of reflection density.The resulting point clouds are used to describe scene structure for loop closure detection.
  • Point-Cloud Generation: The probabilistic extraction method selects peaks beyond one standard deviation and above the mean power in each azimuth scan to separate true detections from false positives.Peak power is modeled with a normal distribution whose mean and standard deviation are computed per azimuth scan.
  • Loop Closure Detection: M2DP describes each generated point cloud using a rotation-invariant global density signature for loop closure detection.The descriptor uses singular vectors of the point-cloud density signature.

D. Pose Graph Optimization

RadarSLAM adds geometrically verified loop closures to a growing pose graph and optimizes all keyframe poses when the loop constraint is accepted.

  • Pose Graph Optimization: A detected loop is geometrically tested by estimating the relative transformation with ICP and RANSAC before adding a loop-closure constraint.The constraint is added only after the geometric verification process.
  • Pose Graph Optimization: When ICP converges, RadarSLAM performs pose-graph optimization over all keyframes using the g2o library.The optimized keyframe poses are then used to update the map.

IV. EXPERIMENTAL RESULTS

Both quantitative and qualitative experiments evaluate RadarSLAM’s pose estimation accuracy and robustness in large-scale and adverse-weather environments.

  • The evaluation combines quantitative pose-accuracy measurements with qualitative robustness tests in large-scale environments and adverse weather.The experiments assess the proposed RadarSLAM system using both numerical and visual evidence.

A. Quantitative Evaluation

RadarSLAM is evaluated on the Oxford Radar RobotCar Dataset using trajectory accuracy metrics and comparisons with radar, camera, and LiDAR methods. It achieves accurate localization while maintaining stronger robustness when competing methods lose tracking or drift.

  • A. Quantitative Evaluation: The quantitative evaluation computes mean translation and rotation errors over trajectory segments from 100 to 800 meters.The evaluation follows KITTI odometry criteria with 100-meter increments.
  • A. Quantitative Evaluation: The Oxford Radar RobotCar Dataset provides 32 radar sequences traversing the same Oxford route with ground-truth poses for benchmarking.The dataset uses a 4Hz FMCW scanning radar with a 163-meter maximum range.
  • A. Quantitative Evaluation: The estimated trajectories and ground truth are compared visually across five Oxford Radar RobotCar sequences.The comparison is presented in Fig. 6.
  • A. Quantitative Evaluation: 2.1854% translation error and 0.0071 deg/m rotation error are achieved by RadarSLAM, outperforming its pure odometry method.RadarSLAM methods outperform the approach in but are inferior to [10] on the reported mean errors.
  • A. Quantitative Evaluation: RadarSLAM achieves comparable or better localization accuracy with enhanced robustness than camera- and LiDAR-based methods.ORB-SLAM 2 stereo, LOAM, and SuMa fail to complete whole sequences or drift rapidly under severe dynamics.

B. Qualitative Evaluation

Qualitative evaluation examines self-collected radar sequences in fog/rain, snow, countryside, and night conditions. RadarSLAM remains reliable across these conditions, whereas camera- and LiDAR-based methods are affected by sensor degradation.

  • B. Qualitative Evaluation: Five self-collected radar sequences cover fog/rain, snow, countryside, and night conditions using a vehicle equipped with GPS/IMU, LiDAR, stereo camera, and radar.The dataset is designed to compare sensor robustness under challenging weather conditions.
  • B. Qualitative Evaluation: Heavy snowfall covers substantial portions of the camera, LiDAR, and radar sensors, making the Snow sequence particularly challenging.The snow conditions create severe occlusion during data collection.
  • B. Qualitative Evaluation: The Snow, Countryside, and Night trajectories remain close to ground truth, while the two Fog/Rain trajectories drift over time without loop closures.The trajectories are plotted on Google maps in Fig. 7.
  • B. Qualitative Evaluation: RadarSLAM is the only one of the three sensor modalities reported to operate and localize reliably in all tested weather conditions.LiDAR methods are affected by fog and snow, while vision methods fail with water drops or motion blur.

C. Computation Time

RadarSLAM runs online without a GPU and demonstrates real-time performance on a long radar sequence. The system is designed to operate reliably in large-scale environments while constructing globally consistent maps.

  • C. Computation Time: The C++ implementation runs at approximately 6Hz on an Intel i7 2.60GHz laptop with 16GB RAM and no GPU.This runtime is reported for the complete RadarSLAM system.
  • C. Computation Time: A 37-minute sequence containing about 9000 images is used to demonstrate the system’s real-time performance.The runtime is presented in Fig. 10.
  • C. Computation Time: The full FMCW-radar SLAM system is intended to operate online in large-scale environments and construct globally consistent maps.The conclusion identifies future IMU-radar fusion as a direction for better accuracy.
Loading 2005.02198v1…