Yang, Jiaolong
Description
Motion estimation is one of the fundamental problems in computer
vision. It has broad applications in the fields of robot
navigation, mixed and augmented reality, visual tracking, image
and video processing, intelligent transportation systems and so
on. Up until now, motion estimation is far from a solved problem,
and it is still one of the active research topics in and beyond
the computer vision community. This thesis is dedicated to both
camera motion...[Show more] estimation -- including motion estimation for 3D
and 2D cameras -- and dense image motion for color images. We
push the limits of the state of the art in various aspects such
as optimality, robustness, accuracy and flexibility. The main
contributions are summarized as follows.
First, a globally optimal 3D point cloud registration algorithm
is proposed and applied to motion estimation of 3D imaging
devices. Based on Branch-and-Bound (BnB) optimization, we
optimally solve the registration problem defined in Iterative
Closest Point (ICP). The registration error bounds are derived by
exploiting the structure of the SE(3) geometry. Other techniques
such as the nested BnB and the integration with ICP are also
developed to achieve efficient registration. Experiments
demonstrate that the proposed method is able to guarantee the
optimality, and can be well applied in estimating the global or
relative motion of 3D imaging devices such as 3D scanners or
depth sensors.
Second, a globally optimal inlier-set maximization algorithm is
proposed for color camera motion estimation. We use BnB to seek
for the optimal motion which gives rise to the maximal inlier set
under a geometric error. An explicit, geometrically meaningful
relative pose parameterization -- a 5D direct product space of a
solid 2D disk and a solid 3D ball -- is proposed, and efficient,
closed-form bounding functions of inlier set cardinality are
derived to facilitate the 5D BnB search. Experiments on both
synthetic data and real images confirm the efficacy of the
proposed method.
Third, a scene constraint based method for relative pose
estimation between a 2D color camera and a 3D sensor is
developed. We formulate the relative pose estimation as a 2D-3D
registration problem minimizing the geometric errors from the
known scene constraints. Our method takes only a single pair of
color and depth images as input, and is correspondence-free. In
addition, a new single-view 3D reconstruction algorithm is
proposed for obtaining initial solutions. The experiments show
that the method is both flexible and effective, producing
accurate relative pose estimates and high-quality color-depth
image registration results.
Fourth, a highly-accurate optical flow estimation algorithm based
on piecewise parametric motion model is proposed. It fits a flow
field piecewise to a variety of parametric models where the
domain of each piece (i.e., shape, position and size) and its
model parameters are determined adaptively, while at the same
time maintaining a global inter-piece flow continuity constraint.
The energy function takes into account both the piecewise
constant model assumption and the flow field continuity
constraint, enabling the proposed algorithm to effectively handle
both homogeneous motions and complex motions. The experiments on
three public optical flow benchmarks show that the proposed
algorithm achieves top-tier performances.
At last, we propose a robust algorithm for optical flow
estimation in the presence of transparency or reflection. It
deals with a challenging, frequently encountered, yet not
properly investigated problem in optical flow estimation: the
input two frames contain one background layer of the scene and
one distracting, possibly moving layer due to transparency or
reflection. The proposed algorithm performs both optical flow
estimation and image layer separation. It exploits a generalized
double-layer brightness consistency constraint connecting these
two tasks, and utilizes the priors for both of them. The
experiments on synthetic and real images confirm its efficacy.
Items in Open Research are protected by copyright, with all rights reserved, unless otherwise indicated.