|
|
|
Camera Tracking by Linearizing the Local
Appearance Manifold | A set
of n-pixel images captured by a moving camera lie on a 6D manifold
in Rn. While appearance manifold is usually nonlinear and
numerous samples are required to learn it globally. It can be
linearized around a local region from no more than 7 samples.
Based on this idea, we propose to
incrementally tracking camera motion through sampling and
linearizing the local appearance manifold. At each frame time, we
use a cluster of calibrated and synchronized small baseline cameras
to capture scene appearance samples at different camera poses. We
compute a first-order approximation of the appearance manifold
around the current camera pose. Then, as new cluster samples are
captured at the next frame time, we estimate the incremental camera
motion using a linear solver. By using intensity measurements and
directly sampling the appearance manifold, our method avoids the
commonly-used feature extraction and matchingprocesses, and does not
require 3D correspondences across frames. Thus it can be used for
scenes with complicated surface materials, geometries, and
view-dependent appearance properties, situations where many other
camera tracking methods would fail. |
|
|
Appearance-Based 3D Motion
Segmentation | Motion
segmentation is a fundamental aspect of tracking in a scene with
multiple moving objects. Various feature-based 3D motion
segmentation methods have been developed that cluster feature
trajectories associated with different motions. However no
counterpart has been proposed for appearance-based 3D tracking
methods. Here we propose to cluster individual image
pixels associated with different 3D rigid motions. The basic idea is
that under the brightness constancy assumption, the change of the
intensity of a pixel can be locally approximated as a linear
function of the motion of the corresponding imaged surface. To
achieve appearance-based 3D motion segmentation we capture a
sequence of local image samples at nearby poses, and assign for each
pixel a vector that represents the intensity changes for that pixel
over the sequence. We call this vector of intensity changes a pixel
``intensity trajectory''. We show that, similar to 2D feature
trajectories, the intensity trajectories of pixels corresponding to
the same motion span a local linear subspace. Thus the problem of
motion segmentation can be cast as that of clustering local
subspaces. We have tested this novel segmentation approach using
some real image sequences. We present results that demonstrate the
expected segmentation, even in some challenging cases. |
|
|
|
|
|
|
|
