Localization and 3D Reconstruction of Urban Scenes Using GPS · OpenStreetMap (OSM) And...

Preview:

Citation preview

Localization and 3D Reconstruction of Urban Scenes Using GPS

Kihwan Kim, Jay Summet, Thad Starner, Daniel Ashbrook,Mrunal Kapade and Irfan Essa

GVU, College of Computing, School of Interactive Computing,Georgia Institute of Technology

{kihwan23,summetj,anjiro,thad,mrunal,irfan}@cc.gatech.edu

Motivation

3D models in Google earth

• Manual modeling and texturing (sketch-up)• 3D model covers limited area/country• Slow update (aerial picture)

• Nice quality of 3D model • Texture is realistic• Aerial map covers almost every area

+

-

Laser scan / Vision based approach for efficient modeling is availableBUT STILL VERY EXPENSIVE

Motivation

OpenStreetMap (OSM)And Crowd-sourcing

• People share their traces and geographic data from OSM

• Million of data is aggregatedthen can be editable for free

• Cell-phone with GPS enables this kind of crowd-sourcing

Motivation

1.Can we make an outline of urban scenery with easier and cheaper way?

2. Can we make use of crowd-sourcing concept for easier update and data acquisition ?

Our work started from two questions

Localize and reconstruct buildings in urban area

By only using off-the-shelf Global Positioning System (GPS)

Goal

GPS

Overview

GPS

Overview

We use these signalsTo detect obstacles

GPS

Overview

Actual signals are reflected based on obstacles.

These multipath reflectors reduce Signal to Noise Ratio(SNR)

Reflected signals

GPS

Overview

Vectors : GPS to SatelliteEach vector has SNR value

GPS

Overview

Example:

Gathering SNR on each step.

GPS

Overview

Then, see what happened in a data chart for recorded SNR on each step

GPS

Overview

Still having 48~50 dB of SNR

GPS

Overview

SNR drops to 30 dB

GPS

Overview

SNR drops to almost 0 dB

GPS

Overview

Now, SNR is back to 40~50 dB

Overview

We can infer that something blocks the signals based on change of SNR!

Overview

Sometimes SNR drops to zero but sometimes less than averages.

Testing environmentSite 1 Site 2

Top View

Bird’s eye

Approach (1) Density map generation

Density of chance that buildings or obstacles exist

Approach (1) Density map generation

Approach (2) Clustering based on Density

Meanshift clustering to find peaks in Density

Approach (3) Region map based on each cluster

Region estimation

If we choose this cluster

Using only signals looking at a cluster

Approach (4) Region estimation

Estimated region (Red) and ground truth region (Yellow)Error from discrepancy is~20% but visually reasonable.

Approach (5) Make a volume

Our vector data does not give distance/depth information ( no way to find attenuation )Make Voxel where the occluded vector passes over the estimated region.

Approach (5) Make a volume

Results

Result Video

Evaluation

Measurement errors and shape similarities of dominant clusters in two testing environments (BOA, OAC areas)

site CentroidDist(ft)

RegionErr(%)

RealHeight(ft)

EstimatedHeight(ft)

HeightErr(%)

BOA 7.398 22.73 1023 862.80 15.65OAC 22.59 22.20 820 705.32 13.98

Conclusion

• We have shown that a GPS receiver can detect and localize buildings by measuring reduction of SNR

• Our approach generated reasonable estimates with around 14~22% errors in region and height.

Conclusion

• Sensor is passive and Inexpensive

the advantages :

• Does not require active aiming

• Automatically self-calibrate

• Using unused signals in conventional GPS system

• Recent cell-phones have GPS capability

Conclusion

: Millions of shared GPS signals from crowd’s cell phone.

(i.e. Open Street Map)

Crowd-sourcing

- 1 hour to gathering 4000 samples.

What if people gather together?

Thank you

Approach

(1) Density map generation

See the paper for more details!

Density map

⎪⎩

⎪⎨

≤−

−=

avgjij

avgjij

jj

ijj

yx jiSV

SVSSVS

L if 0

if ),|( minmax

max

,

satellite j of SNR: thjS

GPS oflocation :| ,yxi

satellite. j thepath to traveled thealong

position sample i at the SNR :

th

thijV

Recommended