GeoLocated Sparse Reconstruction Using GPS and Video GPS SLAM

1
Geo-Located Sparse Reconstruction Using GPS and
VideoGPS SLAM

• Brian Clipp
• December 13, 2006

2
Combine GPS and Vision Measurements to create a
Geo-Located Camera Trajectory
3
Strengths and Weaknesses of Vision and GPS

• Vision is consistent locally but drifts globally
• GPS data is generally globally correct but
  contains large outliers
• GPS is locally inaccurate (3 meters RMS error)
  and errors are time correlated (non-white)

4
A Change of Approach

• Initial Proposed Approach
• Sliding fit of GPS to vision followed by bundle
  adjustment
• Final Approach
• Kalman filter estimator. Builds on Urbanscape
  work.

5
Three Step Technique

• Compute scale and rotation rate using three
  camera pose estimation
• Track vision only through a turn, gathering GPS
  measurements.
• Calculate linear fit of GPS measurements to
  vision track to find pose of camera in world
  coordinates.

6
The Ideal Path for This Algorithm
C2
C3
Direction of Travel
C1
Awful Perspective Drawing of Building
C0
t0
7
Three Pose Estimate

• x1TEx3 0
• X1 2D Feature Tracks From First Image
• X3 2D Feature Tracks From Third Image
• E Essential Matrix
• E -R13TT13xR13T
• R13 Rotation from Image 1 to 3
• T13 Translation from Image 1 to 3
• Camera 1 to 2 found similarly
• RANSAC used to find determine inliers

8
Use of 3 Pose Estimate

• Allows computation of scale and rotation rate.

C2
C3
Tvis
C1
Tgps
9
Vision Only Tracking

• By processing over a turn the rotation and
  translation can be found between the camera
  centers and GPS measurements.

10
KF Smooth Motion Model

• T1 T0 Vdt N(0,sT)
• T1 T0 Odt N(0,sT)

11
Vision and GPS Processing

• GPS constrains position drift
• Rotation is taken only from the feature tracks.
• So long as there are static features to track
  rotation can be estimated quite accurately.
• Possibility exists for long term rotational
  drift, may need to update rotation estimate using
  poses when in turns.

12
Completed Project Tasks

• Vision only tracker implemented.
• Initial scale estimate and rotation rate
  calculation in place.
• Linear fit of GPS points to camera centers
• Transforms vision coordinate camera centers to
  world coordinates.
• GPS and Vision processed fused to estimate camera
  position.

13
Data processed so far

• Generated GPX file from SBET positions
• Using this for development because GPS outlier
  detection is then not an issue.
• Franklin street at Top of the Hill, 200 frames

14
Difficulties with Consumer Grade GPS Measurement
Devices

• Multi-path

15
Difficulties with Consumer Grade GPS Measurement
Devices

• Time-correlated bias

16
GPS Path Processed
17
Initialization- Vision only path
18
Vision GPS After Initialization
19
Vision only After Initialization
20
Transformation Results Processing GPS
North
East
21
Transformation Results Processing GPS
Up
East
22
Transformation Results Processing GPS
Up
North
23
Sparse Model Processing GPS

• Shows both initial path and path after
  transformation.
• VRML Model With GPS after initialization
• VRML Model Vision only after initialization

24
Necessity to Apply GPS Measurement to Point
Estimates
Expectation of Pose After GPS Measurement
Expected Feature Projection
Feature Projection from Video
Expectation of Pose Before GPS Measurement
Residual
Motion Direction
Image Plane
25
Remaining Challenges

• KML generation to view models in Google Earth.
• Add outlier detection to handle noisy GPS
  measurements.
• Smoothly Integrate GPS measurements into camera
  path estimate.

26
Questions?