Spectrum analysis of motion parallax:
Heading computation in a 3D cluttered scene
(Joint work with Mike Langer, McGill University)
Previous methods for estimating observer motion in a rigid 3D scene
assume that image velocities can be measured at isolated points.
When the observer is moving through a cluttered 3D scene such as a
forest, however, point-wise measurements of image velocity are more
challenging to obtain because multiple depths and hence multiple
velocities are present in most local image regions.
This work introduces a method for estimating egomotion which avoids
point-wise image velocity estimation as a first step. In its
place, the direction of motion parallax in local image
regions is estimated, using a spectrum-based method, and these
directions are then combined to directly estimate 3D observer motion.
Synthetic Sequences
Three different sequences were used: random (textured) squares, two
layer (additive) transparency, and random oriented (untextured)
cylinders. Three types of motion were used.
Click on the images below to see the (MPEG compressed) movies.
The sequences we used (uncompressed) are available on the links
below. We performed twenty runs with each sequence. Only
the first one is given here.
Case (i): Forward motion
Case (ii): Forward motion + pan (rotation about Y axis)
Case (iii): Lateral motion (left) + roll (rotation about Z axis)
Real sequences