Fast and Extensible Building Modeling from Airborne LiDAR Data Qian-Yi Zhou Ulrich Neumann University of Southern California.

Fast and Extensible Building Modeling from Airborne LiDAR Data

Qian-Yi Zhou Ulrich NeumannUniversity of Southern California

Introduction

• Toward automatic 3D building model reconstruction …– 3D models are useful in several applications– Manual creation is slow and expensive– New instruments provides more and more data– Problem

How to fulfill the gap between data and 3D building models?

Introduction

• Approach overview– Data source: airborne LiDAR data

• Without other data sources• Working directly on point cloud, without rasterization

– Experiments• On 3 different data sets

– Automatic vs. interaction• Fully automatic for flat roofs• Use interaction to acquire the extensibility for non-flat

roof patterns

Pipeline

Input LiDAR data Classified points

ClassificationClassification Planeextraction

Planeextraction

Planes of one patch

BoundarydetectionBoundarydetection

Boundary pointsBuilding model

ModelingModeling

Pipeline

Planeextraction

Planeextraction

Planes of one patch

BoundarydetectionBoundarydetection

Boundary pointsBuilding model

ModelingModeling

Input LiDAR data Classified points

ClassificationClassification

Classification

• Algorithm– Use a SVM (Support Vector Machine) algorithm

• Machine learning method• All the weight parameters could be trained from a

small area of labeled data

– Take several differential geometry properties as features

• Features are defined locally• Same solution works for different data sets, even their

global variants (e.g. absolute height) varies

Classification

• Features: Building (ground) vs. vegetation– Distribution of neighbor points

• Regular vs. Irregular

– Normal direction• Vertical vs. Unordered

– Flatness• Flat vs. Non-flat

– Normal distribution• Regular vs. Irregular

1F p p

2 1 p zF n e

03

0 1 2

F

4 1 5 0,n nF F

Classification0.0 1.00.5

Input LiDAR data F1 (regularity of distribution) F2 (normal direction)

F3 (flatness) F4 (regularity of normal dist.) F5 (regularity of normal dist.)

Classification

• Experiments for classification– Weights for features learned from data:

• W1~5 = (2.5, 0.1, 1.7, 5.2, 20.4)• Regularities of normal distribution are important

– We get around 95% accuracy rate on 3 different data sets

• Sample rate of different data sets varies from 6 points per sq.m. to 17 points per sq.m.

• Global variants (e.g. average height, average intensity) varies between different data sets

Classification

• Post-processing– Refinement

• Use the intuition that points of same category usually occur together in space

• Let points in a local neighborhood vote for the final classification result

– Segmentation• Apply region growing algorithm to find roof patches• Largest patch is assigned as ground

Classification

Input LiDAR data

ClassificationClassification

Pipeline

BoundarydetectionBoundarydetection

Boundary pointsBuilding model

ModelingModeling

Classified points

Planeextraction

Planeextraction

Planes of one patch

Plane extraction

• Algorithm– For each roof patch, apply a region growing

algorithm but based on normal similarity– Iteratively find all plane patterns in one roof patch

Classified points

Planeextraction

Planeextraction

Input LiDAR data

ClassificationClassification

Pipeline

Building model

ModelingModeling

Planes of one patch

BoundarydetectionBoundarydetection

Boundary points

Boundary detection

• Algorithm– Apply a uniform grid onto the point cloud P– Correspondence:

• Boundary line in grid Boundary point of P• Boundary corner in grid Boundary edge of P

– Run two passes, to construct a watertight manifold polygonal boundary from boundary points and edges

Boundary detection

• Algorithm (continue)Boundary

lines

Boundary cornersFirst passFirst passSecond passSecond pass

Boundary detection

• Advantages– Topology guarantee: watertight manifold

boundary– Easy to implement– Efficient and robust– Can be used with

morphological operations

• Limitation– Cannot guarantee geometry completeness

Planes of one patch

BoundarydetectionBoundarydetection

Classified points

Planeextraction

Planeextraction

Input LiDAR data

ClassificationClassification

Pipeline

Boundary pointsBuilding model

ModelingModeling

Building modeling

• Principal directions– Intuition: most boundary line segments in a local

area fall into a couple of directions, known as principal directions

1

2 3

4 5

6 7

Building modeling

• Snapping– Snap to principal directions

• Find as much as possible boundary segments along principal directions

– Snap between neighbor segments• Patches may share same boundary segments• Patches may be connected via a vertical wall

Building modeling

Planes of one patch

BoundarydetectionBoundarydetection

Classified points

Planeextraction

Planeextraction

Input LiDAR data

ClassificationClassification

Pipeline

Boundary pointsBuilding model

ModelingModeling

Non-flat roof extension

Non-flat roof extension

User interactionUser interaction

Non-flat roofs

Non-flat objects extension

• Algorithm– Define a few pattern types, e.g. cone, sphere,

cylinder– Let user specify pattern type of certain roof patch– Apply RANSAC algorithm to get the parameters

for this shape pattern

Experiments

• Oakland (17 samples/sq.m.; 600m x 600m)

Experiments

• Denver (6 samples/sq.m.; 1km x 1km)

Experiments

• Industrial site (14 samples/sq.m.)

Conclusion

• We provide an automatic building modeling pipeline with novel classification, boundary extraction and modeling algorithms; in addition, we show the extensibility to non-flat roofs with the help of a few user interactions.

• Future works– How to quantitatively analysis errors?– How to handle data sets with billions of points?

• An out-of-core implementation is possible!

The end

• Q & A• Acknowledgement

– Data• Airborne 1 Corp., Sanborn Corp., Chevron Corp.,

Sentinel AVE LLC

– Discussion and comments• Suya You, Yuan Li, and anonymous reviewers

– Support• Provost’s Fellowship from USC