Visual Odometry for Vehicles in Urban Environments

1
Visual Odometry for Vehicles in Urban Environments

• CS223B Computer Vision, Winter 2008
• Team 3 David Hopkins, Christine Paulson, Justin
  Schauer

2
Goal Determine Vehicle Trajectory from Video
Cameras Mounted on a Vehicle

• 2 calibrated cameras forward-looking
  side-looking with non-overlapping field of view
• Compare visual odometry results to GPS and
  inertial sensor ground-truth data

3
Approach SIFT features, RANSAC, derive rotation
and translation from essential matrix

• 1. Identify corresponding SIFT features between
  image pairs
• 2. Estimate the fundamental matrix that satisfies
  the epipolar constraint for uncalibrated cameras
  
• using adaptive RANSAC to refine F and reject
  outliers
• 3. Compute the essential matrix from the
  fundamental matrix and the camera calibration
  matrix
• 4. Recover rotation and translation components
  from the essential matrix using singular value
  decomposition (SVD)

4 solutions Pick one where world points are in
front of both cameras
4
Selecting reliable features is key
3067 SIFT candidate features
276 feature correspondences after mutual
consistency check
69 feature correspondences after RANSAC
5
Example Trajectory Animation
Car turns left, then right onto a street with
oncoming traffic Mean Absolute Error 6 m Total
Distance 322 m
Link 2 3
Web 2 3
6
Mean Absolute Error 1 3 percent
Mean Absolute Error 2.7m Total Distance 312
m
Car driving backwards Mean Absolute Error 2.2
m Total Distance 141 m
Straight road with lots of traffic
Mean Absolute Error 0.6 m Total Distance 23
m
Mean Absolute Error 0.3 m Total Distance 27
m
Mean Absolute Error 1.7 m Total Distance
90 m
7
Conclusions / Issues

• Cumulative error is extremely sensitive to
  orientation
• Adaptive RANSAC was helpful in reducing effects
  of moving vehicles
• Visual odometry is not a replacement for GPS, but
  could be used as an alternate or complementary
  method to GPS (i.e. tunnels, parking structures,
  Mars rovers)

8
Equal Team Efforts