PaperlayerView exact revision on arXiv

Source-linked AI summary

KISS-ICP: In Defense of Point-to-Point ICP -- Simple, Accurate, and Robust Registration If Done the Right Way

Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, Cyrill Stachniss

arXiv:2209.15397v2cs.RO

TL;DR

LiDAR odometry systems often gain capability through added complexity, but KISS-ICP asks whether a carefully designed point-to-point ICP pipeline can generalize without extensive tuning. It combines adaptive correspondence thresholds, robust estimation, motion compensation, and downsampling, and reports state-of-the-art-level performance across diverse platforms and conditions.

  • Problem

    Existing LiDAR odometry approaches are often coupled to motion, environment, sensor, or tuning assumptions, limiting out-of-the-box use across diverse scenarios.

  • Method

    KISS-ICP uses point-to-point ICP with adaptive correspondence thresholding, a robust kernel, constant-velocity motion compensation, and voxel-based point-cloud subsampling.

  • Results

    The same configuration performs on par with state-of-the-art odometry across varied environments, motion profiles, platforms, and LiDAR sensing technologies.

  • Takeaways & Limitations

    A minimal point-to-point ICP system can provide broadly applicable LiDAR odometry without IMU or wheel-odometry integration.

Abstract

from arXiv · show

Robust and accurate pose estimation of a robotic platform, so-called sensor-based odometry, is an essential part of many robotic applications. While many sensor odometry systems made progress by adding more complexity to the ego-motion estimation process, we move in the opposite direction. By removing a majority of parts and focusing on the core elements, we obtain a surprisingly effective system that is simple to realize and can operate under various environmental conditions using different LiDAR sensors. Our odometry estimation approach relies on point-to-point ICP combined with adaptive thresholding for correspondence matching, a robust kernel, a simple but widely applicable motion compensation approach, and a point cloud subsampling strategy. This yields a system with only a few parameters that in most cases do not even have to be tuned to a specific LiDAR sensor. Our system using the same parameters performs on par with state-of-the-art methods under various operating conditions using different platforms: automotive platforms, UAV-based operation, vehicles like segways, or handheld LiDARs. We do not require integrating IMU information and solely rely on 3D point cloud data obtained from a wide range of 3D LiDAR sensors, thus, enabling a broad spectrum of different applications and operating conditions. Our open-source system operates faster than the sensor frame rate in all presented datasets and is designed for real-world scenarios.

I. INTRODUCTION

KISS-ICP argues that a deliberately minimal point-to-point ICP system can provide accurate, robust LiDAR odometry across diverse sensors, environments, and robot motions. Its design combines a few carefully selected components rather than sophisticated features, learning, or extensive tuning.

  • Existing 3D LiDAR odometry methods commonly depend on assumptions about robot motion and environmental structure, limiting generalization across scenarios.
  • The proposed system returns to classical point-to-point ICP and reduces the odometry pipeline to essential components grounded in robot kinematics and sequential sensing.
  • The same parameter set is reported to work across highway driving, drone flights, handheld devices, segways, and other challenging real-world scenarios.
  • Its core components include motion prediction, spatial scan downsampling, a robust kernel, and adaptive ICP correspondence thresholds.
  • The authors claim performance on par with state-of-the-art odometry while accurately handling varied environments and motion profiles without IMUs or wheel odometers.

II. RELATED WORK

The paper positions KISS-ICP against increasingly complex LiDAR odometry and SLAM systems by emphasizing minimal point-to-point registration, adaptive correspondence control, and broad applicability.

  • ICP foundations: ICP alternates between finding point correspondences and minimizing a transformation objective until convergence.
  • KISS-ICP design: KISS-ICP adapts the correspondence threshold to changing scenarios using system kinematics and data characteristics, while retaining a robust kernel.
  • Contrasts with prior systems: Unlike feature-based, surfel-based, mesh-based, and implicit-map systems, the approach relies on a point-to-point metric without pose graph optimization.
  • Motion compensation: The system uses a constant-velocity model for motion distortion instead of more sophisticated compensation requiring a known motion profile.
  • Pipeline scope: The pipeline incrementally estimates a moving LiDAR’s trajectory by sequentially registering point clouds with a minimal component set.

A. Step 1: Motion Prediction and Scan Deskewing

KISS-ICP predicts motion with a constant-velocity model and uses it to deskew each LiDAR scan before registration. This avoids additional sensors and is intended to remain effective across common robotic platforms.

  • The method reframes consecutive-scan registration as estimating deviation from expected robot motion rather than aligning arbitrary point-cloud pairs.
  • The constant-velocity model uses the previous translational and rotational velocities to provide a generally applicable motion estimate without wheel encoders or an IMU.
  • Previous pose estimates define the predicted relative pose and corresponding translational and angular velocities for each time step.
  • Each point’s relative acquisition timestamp is used to compensate motion-induced distortion within the LiDAR sweep.
  • Constant-velocity deskewing is easy to implement, broadly applicable, and requires neither additional sensors nor high-precision cross-sensor synchronization.
  • The approach can outperform more complex compensation systems when motion between sweep endpoints is small, as in many robotics applications.

B. Step 2: Point Cloud Subsampling

KISS-ICP avoids environment-dependent keypoint extraction by using voxel-grid downsampling, preserving original point coordinates and applying a second reduction for efficient ICP alignment.

  • Keypoint extraction can improve convergence or robustness but adds complexity and parameters that often require environment-dependent tuning.
  • The method instead computes a spatially downsampled deskewed scan using a voxel grid and retains one point per voxel.
  • A second voxel-grid reduction creates a lower-resolution scan for ICP, reducing processed points and enabling fast alignment.
  • Unlike approaches that select voxel centers, KISS-ICP preserves original coordinates and chooses a point already present in the deskewed scan.

C. Step 3: Local Map and Correspondence Estimation

The local map stores a bounded voxelized subset of prior scans, while adaptive correspondence thresholds use observed deviations from motion prediction to limit nearest-neighbor matches.

  • Local map: Deskewed and subsampled scans are registered against a local map rather than aligned frame-to-frame.The local map is built from previously registered scans.
  • Local map: The local map uses a voxel grid that stores up to Nmax points per voxel.A hash table provides memory-efficient voxel storage and fast nearest-neighbor search.
  • Local map: After registration, transformed points update the voxel grid, while full voxels and measurements beyond rmax are excluded so map size remains bounded.The update uses the current global pose estimate.
  • Correspondence estimation: Nearest-neighbor associations are treated as outliers when their distance exceeds the maximum correspondence threshold.This threshold is integrated into the adaptive search rather than selected solely heuristically.
  • Correspondence estimation: Correspondences are restricted by a threshold τ determined from the expected point displacement caused by deviations from constant-velocity motion prediction.The displacement accounts for rotational and translational deviation, with rotational displacement bounded using the maximum sensor range rmax.
  • Correspondence estimation: The threshold is estimated from a Gaussian model of past displacements exceeding δmin, using the three-sigma bound τt = 3 σt; experiments set δmin to 0.1 m.This prevents σt from shrinking excessively during stationary or constant-velocity motion.

E. Step 4: Alignment Through Robust Optimization

Alignment uses classic point-to-point ICP with nearest-neighbor correspondences, robust optimization, and iterative pose correction until the correction becomes sufficiently small.

  • Registration choice: Classic point-to-point ICP avoids scanner- or environment-dependent features such as normals, curvature, and descriptors.This design targets generalization across different sensor resolutions and noisy or sparse LiDAR data.
  • Alignment pipeline: The predicted scan is transformed into the global frame using the previous pose estimate before correspondence search against the local map.This produces the source points used by ICP.
  • Alignment pipeline: Nearest-neighbor correspondences are retained only when their point-to-point distance is below τt, after which robust optimization estimates the pose correction.The optimization minimizes point-to-point residuals.
  • Robust optimization: The Geman-McClure kernel provides strong outlier rejection during robust minimization.The kernel is applied to residuals from the accepted correspondence set.
  • Termination: ICP repeats correction estimation until convergence, without imposing a maximum iteration count.Termination is based on the applied correction becoming smaller than γ.
  • Pose update: The resulting pose combines the ICP correction with the previous pose and motion prediction, then applies the correction to the point cloud.The correction is accumulated across ICP iterations.

F. Parameters

The implementation uses seven parameters, with the same parameter set applied across all experiments.

  • Parameter count: The system has seven parameters: two for correspondence search, four for map representation and scan subsampling, and one for ICP termination.The maximum scanner range is treated as sensor-dependent rather than a system parameter.

IV. EXPERIMENTAL EVALUATION

Experiments support the claims that the approach matches more complex odometry systems, generalizes across environments and motion profiles, and handles motion distortion without IMUs or wheel odometers.

  • Experimental claims: Experiments support performance on par with more complex state-of-the-art odometry systems.The evaluation also supports accuracy across varied environments and motion profiles using one configuration.
  • Experimental claims: The same system configuration accurately handles varied environments and motion profiles without relying on IMUs or wheel odometers for motion-distortion compensation.These claims are reported as outcomes supported by the experiments.

A. Experimental Setup

The evaluation spans multiple datasets, robot platforms, and LiDAR settings, with KITTI used for comparison against state-of-the-art systems. Some trajectory and runtime results are omitted from the manuscript, while KITTI results use open-loop registration.

  • A. Experimental Setup: The study evaluates KITTI, MulRan, NCLT, segway, and Newer College datasets across autonomous-driving, aerial, handheld, and mobile-robot scenarios.The datasets cover different sensors, platforms, and operating conditions.
  • A. Experimental Setup: Trajectory plots and detailed runtime evaluations are omitted because of space limitations and referred to the official project page.The project page provides plots and per-sequence runtime evaluations.
  • A. Experimental Setup: KITTI comparisons use motion-compensated scans for fairness because most competing systems do not perform motion compensation.Motion compensation is disabled for KISS-ICP and CT-ICP in this initial comparison.
  • A. Experimental Setup: KISS-ICP ranks second among open-source approaches and ninth overall on the official KITTI Benchmark.Only CT-ICP ranks ahead among the compared submissions.
  • A. Experimental Setup: The KITTI evaluation reports average relative translational error in percent and separates SLAM methods from odometry methods.Relative rotational error is omitted from the manuscript but available through the official benchmark.

C. Comparison to State-of-the-Art Systems on Other Datasets

Across MulRan, Newer College, and NCLT, KISS-ICP uses one configuration across varied platforms and achieves competitive or superior reported results, while ablations examine motion compensation and adaptive thresholding.

  • C. Comparison to State-of-the-Art Systems on Other Datasets: KISS-ICP uses the same parameter values across MulRan, handheld, and segway experiments, unlike approaches that employ IMUs or different configurations.The comparison covers different datasets, scenarios, and robot types.
  • C. Comparison to State-of-the-Art Systems on Other Datasets: KISS-ICP outperforms all evaluated state-of-the-art approaches by a large margin on MulRan in both relative and absolute error.The evaluation uses all publicly available MulRan sequences and averages repeated runs per sequence.
  • C. Comparison to State-of-the-Art Systems on Other Datasets: KISS-ICP achieves results similar to CT-ICP on the short Newer College experiment and reports similar results to CT-ICP on NCLT.The long Newer College gap is attributed to CT-ICP’s additional loop-closing module.
  • C. Comparison to State-of-the-Art Systems on Other Datasets: The NCLT evaluation is uncertain because of GPS-ground-truth misalignments, missing frames, and data inconsistencies.The authors discourage using NCLT as an odometry evaluation tool and present the numbers only as estimates.
  • C. Comparison to State-of-the-Art Systems on Other Datasets: The ablations evaluate the motion-compensation scheme and adaptive threshold using KITTI-based experiments.Motion compensation is studied with raw, uncompensated KITTI point clouds.
  • C. Comparison to State-of-the-Art Systems on Other Datasets: KISS-ICP’s velocity estimation is on par with or slightly better than the IMU, and its constant-velocity compensation is reported sufficient for motion distortion.The authors conclude that more sophisticated techniques are unnecessary for most robotic odometry estimation.

2) Adaptive Data-Association Threshold:

The study finds that adaptive correspondence thresholds outperform fixed thresholds across differing motion profiles, while robust-kernel support remains part of the evaluated system. Together with point-cloud-only ICP, these design choices support the paper’s broader claim of effective LiDAR odometry with few parameters.

  • Adaptive threshold evaluation: 0.5 m for KITTI sequence 00 and 1.0 m for sequence 04 are the best fixed thresholds, so one fixed value requires tuning.The sequences represent different average accelerations and motion profiles.
  • Robust kernel: The ablation results use a robust kernel; without it, translational error is 0.67% and rotational error is 0.25%.
  • System implications: Point-to-point ICP, adaptive thresholding, robust kernels, motion prediction, and spatial downsampling form the paper’s compact odometry design.The system is evaluated as a point-cloud-only approach across challenging operating conditions with the same parameter set.
Loading 2209.15397v2…