A Comparison of Drone-Based SfM and Drone-Based Lidar for ...

COLLEGE OF ENGINEERINGSchool of Civil and Construction Engineering

A Comparison of Drone-Based SfM and Drone-Based Lidar for

Dense Topographic Mapping Applications

Chase Simpson, MS, EIT, FS

Instructor of Geomatics

Oregon State University

Motivation

• Large number of surveyors and geospatial professionals using, or interested in using, UAS for topographic mapping

– Reported that UAS can reduce person-hours in surveying by up to 60% (van Rees, 2018)

• 2 main methods of topographic mapping from UAS

1. Structure from Motion (SfM) + Multi View Stereo (MVS) software applied to drone imagery

2. Light detection and ranging (lidar) on UAS

• Confusing and contradictory information on which is “better” 2

UAS for Topographic Mapping

Background on Remote Sensing Systems

3

Active Sensors

3

• Lidar, radar, etc.

• Invariant to ambient lighting conditions

• Typically more expensive

Remote sensor emitting an energy source and

measuring return strength and travel time

UAS-lidar

Background on Remote Sensing

4

Passive Sensors

• Cameras

– RGB, NIR, IR

• Reliant on ambient lighting conditions & environmental factors

• Lower cost

Remote sensor collects energy reflected off of

an object from an external source

UAS-SfM

Background on SfM

• Relatively new photogrammetric approach

– Leverages advanced image matching algorithms from the field of computer vision

• Many requirements are relaxed, as compared with conventional photogrammetry:

– Can work with a wide range of viewing geometries and consumer-grade cameras

– Well suited to UAS imagery!

– Highly automated, easy to use software:

• Agisoft Metashape, Pix4D, etc.

• Typically consists of two different steps:

– Image matching & recovery of camera parameters (SfM)

– Dense reconstruction (MVS)

5

What is Structure From Motion

Specific Goals of Research

Provide information of each platform to aid professionals in selecting the most advantageous technique based on project requirements.

a) Quantitative assessment of UAS-lidar and UAS-SfM in comparison with terrestrial lidar and high accurate check points

b) Qualitative assessment along the following dimensions:

i. Cost

ii. System complexity

iii. Learning curve

iv. Remote aircraft payload requirements

6

Project Location

Summary:

– Large elevation gradients

– Tree Canopy

– Multiple Surface Types

Area: ~8 Acres

7

Port of Skamania – Stevenson, WA

Port of Skamania

• Complex Variety of Surface types

– Grass

– Concrete

– Bare earth

– Gravel

– Asphalt

8

Surface Characteristics

Reference Dataset

• Purpose:- Control for terrestrial lidar & UAS data

- Provide check points for vertical accuracy assessment following ASPRS standards:

- All check points (CP’s) on flat or uniformly sloped open terrain

- Minimum number of 20 points for each surface type

Asphalt: 43 points

Bare Earth: 32 points

Grass: 25 points

• Estimated uncertainties at a 95% Confidence level reported from least square adjustment:

Horizontal: 1.3 cm

Vertical: 2.4 cm

9

Control Survey: Acquisition

Reference Data

• Purpose:

– Provide reference as truth when visually comparing UAS datasets

• Summary:

– Lidar Scanner: Leica P40

– Mounted GNSS: Leica GS14 utilizing ORGN

– Acquired density: 1 cm @ 30 m from scanner

– 10 total scan positions used

– Performed manual ground classification

10

Terrestrial Lidar: Acquisition

After Manual Classification

Before Manual Classification RMSE at 95% Confidence level- Horizontal: 1.1 cm- Vertical: 1.2 cm

UAS-Lidar

11

Platform Specs

Other Phoenix Systems

All images acquired from:

https://www.phoenixlidar.com/

- DJI M600 Pro w/ A3 flight controller- Phoenix Lidar System AL3-32:

- Lidar Scanner: Velodyne HDL 32- GNSS aided INS: NovAtel KVH1725- Camera: Sony A6000

Cost:~$100,000

Note:This platform is ~median cost of UAS-lidar platforms available. Cost increases substantially for increased accuracy

UAS-Lidar

• Mission Planning:

– Software: Phoenix Lidar SpatialExplorer

• Altitude: 180 feet AGL

• Sidelap: 75%

• Flying speed: 8m/s

• Planned Point density: 150 points/m2

• Multiple Returns: (first/last)

• Data Processing

a) NovAtel Inertial Explorer

• Process Trajectory information

b) Phoenix Aerials Inertial Explorer

• Combines lidar and processed trajectory

c) TerraSolid Suite (TerraMatch/TerraScan)

• Maximizes relative accuracy between flight lines

• Point Cloud Classification 12

Data Acquisition

UAS-SfM

• Airframe:

– DJI S900

– Pixhawk Flight Controller

• Camera:

– Sony A6300 (24 MP)

– 30mm lens

– Fixed mount

• Positioning system:

– Piksi Multi GNSS receiver (GPS+GLO)

– Dual Frequency Helical GNSS Antenna

– Records a time stamp for each acquired image

Cost: ~$4,000

(COTS: ~$7,500-$15,000)

13

Platform Specs

UAS-SfM

• Mission Planning:

– Software: Ardupilot Mission Planner

– Altitude: 377 feet AGL

– Planned GSD: 1.5 cm

– Sidelap: 80%

– Overlap: 80%

– Flying Speed: 5 m/s

• Camera Parameters:

– File Format: Raw

– Shutter speed: 1/1250

– Aperture: F5.6

– ISO [min max]: [100 400]

– Focus mode: Auto (center)

– White Balance: Fixed 14

Data Acquisition

UAS-SfM

Software:

• RTKLIB

– Process Trajectory information

– GPS only & GPS+GLONASS

• MATLAB

– Creates .csv file with coordinates of aircraft for each image

– Applies median filter to dense point cloud(s)

• Agisoft Photoscan

– Imagery alignment

– Dense point cloud creation

– Point cloud classification15

Data Processing: Overview

Study completed before the rebranding to “Metashape”

UAS-SfM

• Processing Workflow:

- Held constant for all datasets

- Based on USGS recommended workflow

16

Data Processing: Structure from Motion

Mask out

water from

imagery

Adjust

Camera

Accuracies

Align Photos

(sparse

cloud)

Optimize

cameras

Perform

Gradual

Selection

Create

Dense Point

Cloud

Classify

Ground

Points

Select

GCP(s)

Export Dense

SfM Ground

Points (.las)

Apply

Median Filter

(MATLAB)

Median filtered pointcloud (.las)

Import Data

For GCP sensitivity analysis:

1 GCP vs. 5 GCP(s)

USGSWorkflow

Median Filter Summary:- Binning algorithm- 5 cm x 5 cm bins- Reduce noise- Decrease point density

Summary of Resulting Data

17

Point Clouds

Point CloudAverage Point

Density (pts/m2)Average Point Spacing (cm)

Terrestrial lidar (reference data set) 7000 1.2

UAS-lidar 50 14.5

UAS-SfM raw 5500 1.3

UAS-SfM grid 350 5.4

Qualitative Assessment

Terrestrial Lidar UAS-SfM rawUAS-SfM grid

UAS-lidar

Qualitative Assessment

• Visualized using Cloud Compare

• Benefits:

– Noise visualization

• Key contributions to noise:

– Poor illumination

– Poor Texture

– Decrease in overlap and/or sidelap from large vertical obstructions

– Combinations of above

18

SfM Median Filter

0.00

0.05

0.10

Sta

ndar

d D

evia

tion o

f bin

(m

)

Qualitative Analysis

Profiles showing importance of texture & lighting for SfM

19

Point Cloud Comparisons - Profiles

UAS SfM Raw UAS lidarTerrestrial lidar

(reference data)

UAS SfM Raw UAS lidarTerrestrial lidar

(reference data)

Qualitative Analysis

Asphalts

Poor Texture

Flat/Hard surface

Bare Earth

Good texture

Flat/Soft surface

Rough surface

Grass

Good texture

Rough surface

Dense Vegetation

20

Point Cloud Comparisons

UAS SfM

Raw

UAS SfM

Grid

UAS

lidarGround TruthTerrestrial lidar

Qualitative Comparison

21

Summary of Advantages & Disadvantages

UAS-Lidar UAS-SfM

Sensor type lidar (active) RGB camera (passive)

Can penetrate canopy yes no

Reliant on surface texture no yes

Reliant on lighting conditions no yes

Variables of each data point position & intensity position & RGB

Requirement for georeferencing position and orientation position only

Cost high low

Acquisition time low low

Operational expertise required high moderate

Processing expertise required moderate low

User-input processing time high moderate

Demand on computing resources moderate high

Point density using typical acquisition parameters (pts/m2)

30-250 350-5500

Quantitative Analysis

What’s being compared?

22

Vertical Accuracy Assessment II UAS-SfM vs UAS-lidar

VS

UAS-lidar

Summary: ~$100kActive sensorDirectly georeferenced1 GCP

UAS-SfM

Summary: ~$4kPassive sensorGeoreferenced imagery1 GCPMedian Filtered

Quantitative Analysis

Compared to ground control

– Asphalt: 43 points

– Bare Earth: 32 points

– Grass: 25 points

Summary:

• UAS-lidar excelled over UAS-SfM on poor textured surfaces

• UAS-SfM performed similarly to UAS-lidar on bare earth

• Both techniques performed poorly on grass when compared to the reference dataset. 23

Vertical Accuracy Assessment II UAS-SfM vs UAS-lidar

Terrestrial Lidar(Reference Data)

UAS LidarUAS SfM(Median Filter)

RMSE at 95% confidence level

Conclusions/Recommendations

UAS-SfM should be default system*

– Low cost

– Easy implementation/processing

– Comparable accuracies in many circumstances

24* Except when it shouldn’t

UAS-lidar should be implemented when any of these characteristics are present:

– Homogenous surface texture

– Canopy/vegetation penetration is required

– Poor illumination

– Large quantity of vertical obstructions

QUESTIONS?

More details of this study can be found in my Masters Thesis accessed through the valley Library at OSU

Access to Full Report

26

Addition Content (NOT INCLUDED IN PRESENTATION DUE TO TIME CONSTRAINTS)

Quantitative Analysis

What’s being compared:

• GNSS constellations:

– Does using multiple constellations provide a more accurate trajectory than using a single constellation?

GPS only versus GPS+GLONASS

27

Vertical Accuracy Assessment I SfM Processing

• GCP sensitivity analysis:

– How many GCPs should be used when imagery is georeferenced?

– 1 GCP versus 5 GCPs

– More is probably better, but how much better is it?

“I’m not really sure…”

Quantitative Analysis

28

Vertical Accuracy Assessment I

SfM Processing

Summary:- Bias approached reference data

when more GCPs were used

- Using multiple constellations for the trajectory improved the results in most cases

Expected Results

Measured Results

Bias Standard Deviation

Quantitative Analysis

29

Vertical Accuracy Assessment II UAS-SfM vs UAS-lidar

Summary:- UAS-lidar excelled over UAS-SfM

on asphalts

- UAS-SfM performed similarly to UAS-lidar on bare earth

- Both techniques performed poorly on grass when compared to the reference dataset.

Terrestrial Lidar(Reference Data)

UAS LidarUAS SfM(Median Filter)