PaperlayerView exact revision on arXiv

Source-linked AI summary

Patchwork: Concentric Zone-based Region-wise Ground Segmentation with Ground Likelihood Estimation Using a 3D LiDAR Sensor

Hyungtae Lim, Minho Oh, Hyun Myung

arXiv:2108.05560v2cs.ROcs.CV

TL;DR

Ground segmentation must handle non-flat terrain and nearby objects while limiting false classifications. Patchwork combines a CZM-based representation, region-wise plane fitting, and Ground Likelihood Estimation, achieving promising performance with operation above 40 Hz. Its reported results emphasize robustness against under-segmentation and low recall variation compared with prior approaches.

  • Problem

    Ground segmentation must distinguish ground from non-ground points despite steep slopes, bumpy roads, curbs, flower beds, and other objects that can cause false positives and false negatives.

  • Method

    Patchwork uses a CZM-based nonuniform representation, Region-wise Ground Plane Fitting, and Ground Likelihood Estimation based on uprightness, elevation, and flatness.

  • Results

    Patchwork reports promising performance over state-of-the-art region-wise fitting methods at more than 40 Hz, with little recall variance and improved under-segmentation resistance.

  • Takeaways & Limitations

    The reported ground estimates support robust detection of non-ground objects by mobile robots while remaining fast enough for preprocessing use.

Abstract

from arXiv · show

Ground segmentation is crucial for terrestrial mobile platforms to perform navigation or neighboring object recognition. Unfortunately, the ground is not flat, as it features steep slopes; bumpy roads; or objects, such as curbs, flower beds, and so forth. To tackle the problem, this paper presents a novel ground segmentation method called \textit{Patchwork}, which is robust for addressing the under-segmentation problem and operates at more than 40 Hz. In this paper, a point cloud is encoded into a Concentric Zone Model-based representation to assign an appropriate density of cloud points among bins in a way that is not computationally complex. This is followed by Region-wise Ground Plane Fitting, which is performed to estimate the partial ground for each bin. Finally, Ground Likelihood Estimation is introduced to dramatically reduce false positives. As experimentally verified on SemanticKITTI and rough terrain datasets, our proposed method yields promising performance compared with the state-of-the-art methods, showing faster speed compared with existing plane fitting--based methods. Code is available: https://github.com/LimHyungTae/patchwork

I. INTRODUCTION

Patchwork addresses ground segmentation for non-flat environments with a CZM-based representation, region-wise plane fitting, and Ground Likelihood Estimation. The method targets robust segmentation and reports performance above 40 Hz.

  • Motivation: Ground segmentation supports navigation and moving-object recognition by estimating regions that terrestrial vehicles or humans contact.The paper includes roads, sidewalks, and lawns in its definition of ground.
  • Method: Patchwork extends Region-wise Ground Plane Fitting into a CZM-based region-wise ground segmentation method for 3D point clouds.The paper also analyzes the impact of bin size, which was not covered in the previous work.
  • Method: The CZM-based representation uses nonuniform bin sizes to assign point-cloud density efficiently across zones.Its stated novelty includes analyzing bin-size impact when estimating ground planes in complex urban environments using SemanticKITTI.
  • Method: Ground Likelihood Estimation uses uprightness, elevation, and flatness to reduce false positives in ground estimates.The supplied passage introduces these criteria as part of the method’s contribution.
  • Results: Patchwork shows promising performance over state-of-the-art region-wise fitting methods while operating at more than 40 Hz.The reported result emphasizes low recall variance and overcoming under-segmentation in complex urban environments.

II. RELATED WORKS

Prior ground-segmentation methods address non-flat terrain through grid, probabilistic, learned, and region-wise representations, but face accuracy, computational, or generalization challenges. Region-wise fitting improves robustness by modeling local portions of potentially non-flat ground.

  • The Difficulties of Ground Segmentation: Steep slopes, bumpy roads, curbs, flower beds, and outlier objects can cause under-segmentation when different objects merge into one segment.These conditions complicate simple height filtering and plane-fitting approaches.
  • Related Methods: Gaussian-process and Markov Random Field methods estimate detailed ground points but may require too much computation for preprocessing above 20 Hz.The passage contrasts their detail with the speed requirement of preprocessing algorithms.
  • Elevation Map–based 2.5D Grid Representation: Elevation-map methods are computationally efficient, but steep slopes may be classified as non-ground because of large vertical height differences.These approaches encode point clouds as 2.5D grids using height statistics.
  • Polar Grid Methods: Uniform polar-grid line fitting classifies points using thresholds on parameters such as point-to-line distance, gradient, or y-intercept.Polar representations are commonly used because they match the geometric characteristics of 3D LiDAR sensors.
  • Multiple Plane Fitting–based Ground Segmentation: Region-wise plane fitting improves robustness by modeling local ground regions, while prior assumptions can fail on bumpy roads or complex intersections.Related work also adapts grid size to cloud-point density or incidence angle.

G. Deep Learning-based Methods

Patchwork addresses non-flat ground segmentation by combining a concentric-zone representation with region-wise fitting, while targeting efficient processing. Its design responds to uneven point distributions and the limitations of uniform grids.

  • Region-wise representation: Patchwork models non-flat environments by assuming small point-cloud regions can be treated as locally flat.This motivates region-wise ground-plane estimation rather than a single global plane.
  • Point distribution: More than 90% of SemanticKITTI ground points lie within 20 m of the sensor.This empirical distribution motivates allocating representation density according to range.
  • Uniform-grid limitations: Uniform polar grids suffer from far-range sparsity and near-origin representability problems during ground-plane estimation.Farther bins become too sparse, while overly small nearby bins can impair normal-vector estimation.
  • Concentric Zone Model: CZM assigns different bin sizes across zones to allocate point density appropriately without substantial computational complexity.The representation divides the point cloud into multiple zones containing differently sized bins.
  • Concentric Zone Model: The CZM uses four empirically selected zones, with central and outer-zone bins enlarged to address representability and sparsity issues.The zones are named central, quarter, half, and outer zones.
  • Efficiency: Reducing the bin count from 3,240 to 504 enables Patchwork to operate at more than 40 Hz.The paper connects this reduction to faster processing and improved expressibility for normal-vector estimation.

C. Region-wise Ground Plane Fitting

Region-wise Ground Plane Fitting estimates partial ground surfaces independently within bins, using PCA-based plane estimation and adaptive seed selection. The adaptive initialization targets erroneous low points that can destabilize fitting.

  • R-GPF: Each bin receives an estimated partial ground through R-GPF, after which partial ground points are merged.Patchwork uses PCA instead of RANSAC because PCA provides faster estimation with acceptable performance.
  • Adaptive initialization: Adaptive initial seed selection prevents erroneous points from causing R-GPF to converge to a local minimum.The paper illustrates that mismeasured points can otherwise be selected as seeds and disrupt plane fitting.
  • Computational cost: PCA-based methods are at least two times faster than RANSAC-based methods.This speed advantage supports PCA-based estimation as a preprocessing operation.
  • PCA plane estimation: PCA estimates the ground-plane normal from the eigenvector associated with the smallest covariance-matrix eigenvalue.The corresponding plane coefficient is computed from that normal and the mean point of the unit space.
  • Initial seeds: Bins with sufficient points select their lowest-height points as initial seeds because these points are most likely to belong to the ground surface.The mean seed height is used to initialize the estimated ground set.

n. Next, plane coefficient dl

Patchwork refines its iterative plane estimate using adaptive height filtering for initial seeds. The filtering is stricter in the central zone and becomes less restrictive farther out to preserve downhill ground points.

  • Method distinction: The main difference from the original R-GPF is adaptive initial seed selection to prevent convergence to a local minimum.This modification directly addresses erroneous low measurements during initialization.
  • Failure mode: Erroneous points below the actual ground can arise from multipath or reflected LiDAR signals and impede correct R-GPF estimation.The paper observes this issue primarily in the central zone, where reflections occur under relatively strong signals.
  • Adaptive filtering: Central-zone points are filtered from the initial ground set when their height falls below a sensor-height-based threshold.The threshold uses a height margin below the sensor height.
  • Adaptive filtering: For non-central zones, the adaptive threshold decreases with zone index to avoid filtering downhill points that are actually true positives.This preserves valid ground points as the estimated range from the sensor increases.

D. Ground Likelihood Estimation

Ground Likelihood Estimation (GLE) probabilistically tests region-wise ground-plane estimates using uprightness, elevation, and flatness to reject non-ground planes while recovering some steep-slope ground.

  • GLE performs a region-wise probabilistic binary classification test to determine whether each estimated partial ground belongs to the actual ground.
  • GLE improves precision by excluding unintended planes containing non-ground points, including walls and parked cars rejected by uprightness and elevation.
  • Patchwork models each region’s likelihood using uprightness, elevation, and flatness terms applied to plane orientation and surface statistics.
  • The elevation filter uses mean height and radial distance to reject likely non-ground regions, especially near the sensor, while becoming inconclusive beyond a constant range.
  • Flatness can restore steep-uphill regions rejected by elevation when their surface variable indicates an even plane.

A. Dataset

The evaluation uses SemanticKITTI urban sequences and an additional rough-terrain setting to assess ground segmentation beyond paved vehicle-platform data.

  • SemanticKITTI ground truth includes lane marking, road, parking, sidewalk, other ground, vegetation, and terrain classes.Vegetation points are counted as ground only when their sensor-relative z value is below −1.3 m.
  • The study adds rough-terrain experiments because SemanticKITTI data were acquired from vehicle platforms on pavements only.

B. Error Metrics

Performance is evaluated with point-level precision, recall, F1 score, and accuracy under the reported Patchwork configuration.

  • The evaluation reports Precision, Recall, F1 score, and Accuracy from TP, TN, FP, and FN point counts.
  • The reported CZM configuration uses zone-dependent radial and angular bin counts with Lmin = 2.7m and Lmax = 80.0m.
  • The reported GLE configuration uses Lτ = Lmax,2 and zone-dependent flatness thresholds στ,1 = 0.00012 and στ,2 = 0.0002.

A. Performance Analysis with Different Bin Sizes

Larger bins improve precision and F1 but reduce recall, while uprightness and GLE reduce erroneous ground estimates and support robust performance on complex or rough terrain.

  • A. Performance Analysis with Different Bin Sizes: Larger bin sizes substantially improve precision and F1 score, but reduce recall because each bin has lower spatial resolution.
  • B. Impact of Uprightness: Using uprightness improves precision and reduces variance, although stricter thresholds can reject true ground on steep slopes and road-curb boundaries.The experiments identify θτ = 45◦ as the most reasonable threshold.
  • D. Ground Likelihood Estimation: GLE filters wrongly estimated partial ground and rejects many walls and parked cars through uprightness and elevation.
  • D. Comparison with State-of-the-Art Methods: Competing methods struggle with steep slopes, intersections, curbs, undulated terrain, and rough terrain, producing false negatives or local-minimum estimates.
  • D. Comparison with State-of-the-Art Methods: Patchwork achieves promising performance with little recall variance and addresses under-segmentation in urban corner cases.
  • D. Comparison with State-of-the-Art Methods: Some Patchwork false negatives correspond to the lowest parts of objects and help resolve under-segmentation despite lowering quantitative metrics.

E. Algorithm speed

Patchwork is designed for fast preprocessing, combining robust ground segmentation with speed advantages over other multiple plane fitting–based methods.

  • E. Algorithm speed: More than 40 Hz, Patchwork achieves promising performance over state-of-the-art region-wise fitting–based methods.Its ground estimates also show the least recall variance, addressing under-segmentation in complex urban environments.
  • E. Algorithm speed: 504 bins versus 3,240 for R-GPF reduces plane-fitting computation and contributes to Patchwork’s faster speed.The method also uses PCA rather than the RANSAC-based approach used by CascadedSeg.
  • E. Algorithm speed: Patchwork provides well-segmented ground estimation with smaller performance variations, supporting robust detection of non-ground objects by mobile robots.
Loading 2108.05560v2…