Satellites in Our Pockets: An Object Positioning System

using Smartphones

Justin Manweiler, Puneet Jain, Romit Roy Choudhury

TsungYun 20120827

Outline

• Introduction• Primitives for Object Localization• System Design• Evaluation• Future Work• Conclusion

Introduction

• Augmented Reality (AR)– Location based Query• “Restaurants around me?”

– Distant Object based Query• how expensive are rooms in that nice hotel far away?• is that cell tower I can see from my house too close

for radiation effects?

Introduction

• Wikitude– http://www.youtube.com/user/Wikitude

• out-of-band tagging– Objects in the environment should be annotated

out-of-band – someone visited Google Earth and entered a tag

Introduction

• Problem– Can a distant object be localized by looking at it

through a smartphone?– The problem would have been far more difficult

five years back• SmartPhone !!– Camera, GPS, Accelerometer, Compass, and

Gyroscope

Introduction

• OPS (Object Positioning System)– Computer vision – Smartphone sensors– Mismatch optimization

• Contribution– Localization for distant objects within view– System design and implementation on the Android

Nexus S platform

Introduction

• OPS overview

Primitives for Object Localization

• (A) Compass triangulate

Primitives for Object Localization

• We cannot ask the user to walk too far– distance between camera views is much smaller

than the distance from the camera to the object– Compass precision becomes crucial– Smartphone sensors are not nearly designed to

support such a level of precision• GPS can be impacted– Weather, clock error, …

Primitives for Object Localization

• (B) Visual trilateration– Trilateration is used in GPS, but not in the distant

object positioning

Primitives for Object Localization

• The possible position lies on a curve– Visual angle: Computer vision + accelerometer

Primitives for Object Localization

• (C) Visual Triangulation– Parallax• Multiple views of an object from different angles

produce visual distortions

– The properties of parallax and visual perception in general are well-understood• We can find the interior angle

– Still form a curve

Primitives for Object Localization

Primitives for Object Localization

• Combining Triangulation and Trilateration

Primitives for Object Localization

• We do not obtain a single point of intersection across all curves– Due to errors from GPS, compass, and inaccurate

parameter estimation from the visual dimensions– Increasing the number of camera views will help• it will also increase the number of curves (each with

some error)

– Rely on optimization techniques to find a single point of convergence

System Design

System Design

• Structure form Motion (SFM)– State-of-art computer vision technique– Input: multiple photos from the user– Feature detector, Bundle Adjustment, Levenberg-

Marquardt Algorithm– Output: (a) 3D point cloud of the geometry (b) the relative positions and orientation of the camera

System Design

• Structure form Motion (SFM)

System Design

• The other issues– Capture user Intent• OPS must be able to automatically infer which object in

view the user is most-likely interested• the object-of-interest roughly at the center of the

camera’s viewfinder

– Privacy • user can only upload the keypoints and feature

descriptors

System Design

• We utilize SFM as a “black box” utility• However, GPS/compass readings themselves

will be noisy• Optimization– Minimize the Compass error– Minimize the GPS noise– OPS Optimize on Object Location

System Design

• Triangulation via Minimize the Compass error– this scales to support an arbitrary number of

GPS and compass bearing pairs– We want all C(n, 2) pairs points converge to a

single point – A minimize question• Add a error term to each compass value

System Design

• Triangulation via Minimize the Compass error

System Design

• Minimization of GPS Noise, Relative to Vision– Adjust the GPS reading from the position user take

the photograph– solve a scaling factor λ that proportionally

expands the distances in the SFM point cloud to match the equivalent real-world distances

– +

=

System Design

• Minimization of GPS Noise, Relative to Vision

System Design

• OPS Optimization on Object Location

System Design

• OPS Optimization on Object Location

System Design

• Extending the Location Model to 3D– Pitch• rotational movement orthogonal to the plane of the

phone screen, relative to the horizon

– Adjustment• From our 3D point cloud, there is a unique mapping of

every 3D point back to each original 2D image

Evaluation

• Experiment– More than 50 buildings– Distance from user to the building: 30~150m• far enough away that it makes sense to use the system• limited by the user’s ability to clearly see the object

and focus a photograph

– 4 pictures, each photograph was taken between 0.75m and 1.5m

Evaluation

• Experiment– Processing time: 30~60s• primarily attributable to structure from motion

– Quality of photographs• Lighting• Blur• Overexposed

Evaluation

Evaluation

Evaluation

Evaluation

• Introduce some noise– Gaussian distribution with mean 0 and a varied

standard deviation– Sensitivity to GPS– Sensitivity to compass error

• Sensitivity to Photograph detail– Varied resolution

Evaluation

Evaluation

Evaluation

Future Work

• Live Feedback to Improve Photograph Quality– feedback to the user

• Improving GPS Precision with Dead Reckoning– Dead Reckoning is the process of calculating one's

current position by using a previously determined position

• Continual Estimation of Relative Positions with Video

Conclusion

• Localization for distant objects using the view from camera, without any off-band effort

• Real and implementation on the Android Nexus S platform

• Achieve a promising results– The prime limitation is from GPS error

• The two equation to calculate the “height”