PaperlayerView exact revision on arXiv

Source-linked AI summary

Patchwork++: Fast and Robust Ground Segmentation Solving Partial Under-Segmentation Using 3D Point Cloud

Seungjae Lee, Hyungtae Lim, Hyun Myung

arXiv:2207.11919v2cs.ROcs.CV

TL;DR

Ground segmentation methods face labeling and domain-sensitivity limitations in learning-based approaches, while conventional methods retain under-segmentation problems on bumpy ground. Patchwork++ extends Patchwork with adaptive estimation, temporal recovery, region-wise fitting, and reflected-noise removal; SemanticKITTI experiments report promising accuracy, robustness, and speed.

  • Problem

    Ground segmentation remains limited by labeling and domain sensitivity in learning-based methods and under-segmentation in conventional methods on bumpy ground.

  • Method

    Patchwork++ extends Patchwork with A-GLE, TGR, R-VPF, and RNR for adaptive parameters, temporal ground recovery, layered-ground fitting, and virtual-noise removal.

  • Results

    Patchwork++ achieves the highest F1 score among compared methods and higher precision and recall than Patchwork.

  • Takeaways & Limitations

    Patchwork++ is reported as faster and more robust than state-of-the-art methods while requiring fewer environment-dependent parameters to fine-tune.

Abstract

from arXiv · show

In the field of 3D perception using 3D LiDAR sensors, ground segmentation is an essential task for various purposes, such as traversable area detection and object recognition. Under these circumstances, several ground segmentation methods have been proposed. However, some limitations are still encountered. First, some ground segmentation methods require fine-tuning of parameters depending on the surroundings, which is excessively laborious and time-consuming. Moreover, even if the parameters are well adjusted, a partial under-segmentation problem can still emerge, which implies ground segmentation failures in some regions. Finally, ground segmentation methods typically fail to estimate an appropriate ground plane when the ground is above another structure, such as a retaining wall. To address these problems, we propose a robust ground segmentation method called Patchwork++, an extension of Patchwork. Patchwork++ exploits adaptive ground likelihood estimation (A-GLE) to calculate appropriate parameters adaptively based on the previous ground segmentation results. Moreover, temporal ground revert (TGR) alleviates a partial under-segmentation problem by using the temporary ground property. Also, region-wise vertical plane fitting (R-VPF) is introduced to segment the ground plane properly even if the ground is elevated with different layers. Finally, we present reflected noise removal (RNR) to eliminate virtual noise points efficiently based on the 3D LiDAR reflection model. We demonstrate the qualitative and quantitative evaluations using a SemanticKITTI dataset. Our code is available at https://github.com/url-kaist/patchwork-plusplus

I. INTRODUCTION

Patchwork++ addresses limitations in non-learning-based LiDAR ground segmentation, including parameter tuning, partial under-segmentation, and uneven environments. It adds adaptive and outlier-rejection mechanisms, and experiments on SemanticKITTI report promising performance and speed.

  • Non-learning-based ground segmentation avoids the labeling burden and scene or sensor-configuration sensitivity associated with learning-based methods.
  • Ground segmentation must be fast, preserve precision and recall, and handle uneven outdoor environments because it precedes downstream perception algorithms.
  • Conventional methods can suffer under-segmentation, while Patchwork occasionally fails partially because deterministic parameters may filter out solid ground.
  • Patchwork++ introduces A-GLE and TGR to reduce false negatives through adaptive parameters and temporal ground information.
  • RNR and R-VPF reject reflected noise or non-ground points when the usual ground assumptions do not hold.
  • SemanticKITTI experiments report promising performance compared with state-of-the-art methods, including Patchwork, together with the fastest speed.

II. RELATED WORKS

Related work spans learning-based and conventional LiDAR ground segmentation, but both families retain important limitations. Patchwork++ builds on this context with additional modules aimed at robustness and efficiency.

  • Conventional methods include local-convexity, line-based, PCA-based, and multi-region approaches designed for fast or non-flat ground segmentation.
  • Learning-based methods can require laborious point-wise human labeling and may degrade on scenes or sensor configurations unlike their training data.
  • Single-plane assumptions are insufficient for broad, partially bumpy regions, leaving conventional methods vulnerable to under-segmentation.
  • Ground segmentation supports object clustering by removing roughly half of outdoor LiDAR points and also supplies ground features for LiDAR odometry.
  • Patchwork++ differs from Patchwork through RNR, R-VPF, A-GLE, and TGR, while retaining the earlier concentric-zone and region-wise fitting components.

III. PATCHWORK++: FAST, ROBUST, AND ADAPTIVE GROUND SEGMENTATION

Patchwork++ defines ground segmentation as classifying LiDAR points into ground and non-ground sets, then addresses failure sources through reflected-noise handling and adaptive processing modules. Its RNR component targets low, weak-intensity virtual points that can corrupt ground initialization.

  • Problem Definition: The task partitions a measured point cloud into estimated ground points and estimated non-ground points, with the goal of maximizing true positives while rejecting false positives and negatives.
  • RNR: Reflected Noise Removal: Low points beneath the actual ground can become R-GPF initial seeds and trigger under-segmentation because seed selection assumes the lowest bin points are ground.
  • RNR: Reflected Noise Removal: Filtering points below a fixed z threshold assumes sufficiently flat surroundings and can erase true ground points on steep slopes.
  • RNR: Reflected Noise Removal: Reflected virtual noise tends to lie along sensor-to-incidence directions and have low intensity after additional reflections from vehicles or glass.
  • RNR: Reflected Noise Removal: RNR removes low-height, low-intensity points from selected bottom rings using height and intensity thresholds, limiting loss of actual ground points.

C. R-VPF: Region-wise Vertical Plane Fitting

R-VPF preprocesses each concentric-zone bin by identifying and rejecting dominant vertical points, enabling R-GPF to estimate elevated ground planes more accurately. It also addresses reflected and downhill noise that can otherwise trigger under-segmentation.

  • Motivation: R-GPF can select retaining-wall points as ground seeds when the desired ground lies above a vertical structure.These outliers can corrupt PCA-based ground fitting because the estimated plane is sensitive to them.
  • Noise handling: Reflected noise can appear below the ground because bottom-laser reflections create virtual points that violate lowest-point seed assumptions.The associated RNR module uses the LiDAR reflection model to remove such noise, while downhill filtering examples show reduced false negatives.
  • Motivation: Elevated ground remains classified as ground because objects such as humans may stand on it, despite its separation from lower ground.The paper therefore treats layered elevated surfaces as actual ground rather than non-ground.
  • R-VPF method: R-VPF rejects dominant vertical points within each concentric-zone bin before R-GPF estimates the ground plane.The method divides the point cloud into polar-coordinate bins and estimates vertical points through four iterative steps.
  • R-VPF method: R-VPF selects seeds from the lowest points, estimates a PCA direction, samples candidate vertical points, and accumulates them across iterations.Previously identified vertical points are rejected in subsequent iterations.

D. A-GLE: Adaptive Ground Likelihood Estimation

A-GLE adapts Patchwork++ thresholds from previously estimated ground planes because elevation and flatness distributions vary across environments. It updates ground-likelihood parameters and the noise-removal height threshold using stored estimation states.

  • Ground likelihood: The prior ground-likelihood function uses uprightness, elevation, and flatness indicators derived from PCA orientation, mean height, and local surface variation.These indicators determine whether region-wise estimated planes are ground or non-ground.
  • Motivation: Patchwork’s elevation and flatness thresholds require environment-specific adjustment because true-ground distributions vary across scenes.The paper illustrates similar distributions for some sequences but substantially different distributions for others.
  • A-GLE overview: A-GLE adaptively updates elevation, flatness, and noise-removal height parameters from previously obtained ground-segmentation results.It stores prior ground states and uses them to update eτ,m, fτ,m, and hnoise for the next estimation.
  • Elevation: A-GLE defines definite ground as previously estimated planes satisfying uprightness and elevation conditions in each concentric-zone ring.The properties of these planes update the elevation parameter for the next estimation.
  • Flatness: Patchwork’s flatness measure can be inconsistent because concentric-zone binning changes λ1,n and λ2,n even when the overall ground is unchanged.Patchwork++ revises the flatness calculation and updates its threshold from the distribution of definite-ground planes.

E. TGR: Temporal Ground Revert

TGR addresses temporary partial under-segmentation by reverting rejected ground points using ground-plane properties estimated at the current time. It complements A-GLE’s temporally updated parameters and reduces false negatives.

  • Motivation: A-GLE can miss unusual ground points because its updates over time behave like a low-pass filter when flatness temporarily becomes large.Rough terrain with grass may have higher elevation and flatness values than the self-updated parameters.
  • TGR mechanism: TGR reverts under-segmented ground planes into segmented bins using definite ground planes identified at time t.The method uses temporary ground properties to revisit rejected regions.
  • TGR mechanism: For each under-segmented bin, TGR compares the bin’s flatness value with the temporary threshold f^t_τ,m.Rejected points can be reverted when their flatness satisfies the temporary threshold condition.
  • Integration: A-GLE stores estimated ground-plane variables and updates parameters from previous results, while TGR double-checks under-segmented grounds at time t.This describes the complementary coarse-to-fine temporal flow of the two modules.

A. Dataset and Error Metrics

The evaluation uses SemanticKITTI point-wise labels to define ground and non-ground classes for comparing ground segmentation methods. Vegetation is excluded from both evaluation classes.

  • Ground-label definition: SemanticKITTI labels road, parking, sidewalk, other ground, lane marking, and terrain as ground points.The evaluation is based on point-wise dataset labels.
  • Ground-label definition: All other points are regarded as non-ground points, while vegetation is evaluated as neither ground nor non-ground.This differs from the authors’ previous work.

B. Parameters of Patchwork++

Patchwork++ is evaluated with fixed implementation settings, including initialized A-GLE and TGR thresholds and specified gains for each module. Its overall comparison reports the highest F1-score among the listed methods.

  • Parameter settings: Initial e_τ,m and f_τ,m are set to 0 for all four concentric-zone rings.Other parameters are identical to the authors’ previous work.
  • Parameter settings: RNR uses N_noise = 20 and I_noise = 0.2, while R-VPF uses d_v = 0.1, θ_v = 0.707, and K_v = 3.These settings define the outlier-rejection modules used in the evaluation.
  • Parameter settings: A-GLE uses a_m = 1, b_m = 3 for m = 1, b_m = 2 otherwise, and δ = −0.5; TGR uses c_m = 1.5 for all rings.These are the reported gains and offset for adaptive updating and temporal revert.
  • Comparison: Patchwork++ exhibits the highest F1-score among RANSAC, LineFit, GPF, R-GPF, and CascadedSeg.The comparison is summarized in Table I and Fig. 8.
  • Comparison: Patchwork++ achieves higher precision and recall than Patchwork and a lower standard deviation of recall.The authors characterize the resulting segmentation as accurate and robust in complex urban environments.

B. Effect of R-VPF

R-VPF enables correct ground-plane estimation when elevated terrain is confused with structures such as fences or flower beds. The reported visual evidence also supports the validity of A-GLE’s self-updated thresholds for separating ground and non-ground planes.

  • R-VPF effect: R-VPF detects vertical-plane points during preprocessing so R-GPF can estimate the correct ground plane when fence points interfere with terrain.Without R-VPF, terrain points are under-segmented; with it, fence points are extracted as non-ground points first.
  • Quantitative comparison: The supplied table caption identifies Table I as a performance comparison whose metrics are reported as mean ± standard deviation.No table-row values are provided in the passage.
  • A-GLE effect: A-GLE’s estimated elevation threshold yields D_m containing approximately 95.8% true ground points.The green D_m points also include 85.3% of true ground points, supporting updates to the flatness threshold.
  • A-GLE effect: The self-updated flatness threshold f_τ,m appears suited to distinguish blue ground-plane points from red non-ground-plane points.The paper treats this separation as evidence that the updated threshold is valid.

D. Effect of TGR

TGR uses the distribution characteristics of currently estimated ground planes to recover missed ground points, increasing recall with negligible precision loss. The method also adapts its parameters to surroundings and improves segmentation speed over Patchwork.

  • TGR effect: TGR double-checks bins using currently estimated ground-plane distributions and reverts numerous false negatives into true positives.This coarse-to-fine strategy addresses partial under-segmentation.
  • TGR effect: TGR increases recall with a negligible decrease in precision.
  • Adaptive parameters: Self-updated eτ,m and fτ,m become smaller on relatively flat highways and higher in country scenes with steeper or bumpier terrain.Figure 12 compares parameter distributions across country, urban, and highway SemanticKITTI sequences.
  • Algorithm speed: Bin-wise sorting reduces the sorting complexity from O(N log N) to approximately O(LM log M) compared with Patchwork's global sorting.Under N = LM, the reduction is LM log L because N ≫ L and N ≫ M.
  • Overall effect: Patchwork++ was verified to be faster and more robust than other state-of-the-art methods while requiring fewer environment-dependent parameters.The authors state that these changes make ground segmentation easier to use in practice.
Loading 2207.11919v2…