In this work, I examine the problem of understanding American football in video. In particular,
I present several mid-level computer vision algorithms that each accomplish a different sub-task
within a larger system for annotating, interpreting, and analyzing collections of American football
video. The analysis of football video is useful in its own right, as teams at all levels from
high school to professional football currently spend thousands of dollars and countless human
work hours processing video of their own play and the play of their opponents with the aim
of developing strategy and improving performance. However, because football is an extremely
challenging visual domain, with difficulties ranging from the chaotic motion and identical appearance
of the players to the visual clutter on the field in the form of logos and other markings,
computer vision algorithms developed towards the end goal of understanding American football
are broadly applicable across a variety of visual problems.
I address four specific football-related problems in this thesis. First, I describe an approach
for registering video with a static model (i.e. the football field in the American football domain)
using a novel concept of locally distinctive invariant image feature matches. I also introduce a
novel empirical registration transform stability test, which we use to initialize our registration
procedure.
Second, I outline a novel method for constructing mosaics from collections of video. This
method takes a greedy utility maximization approach to build mosaics that achieve user-definable
mosaic quality objectives. While broadly applicable, our mosaicing approach accomplishes several
tasks specifically relevant to the analysis of football video, including automatically constructing
reference image sets for our video registration procedure and for computing background
models for initial formation recognition and player tracking algorithms.
Third, I present an approach for recognizing initial player formations. This approach, called
the Mixture-of-Parts Pictorial Structure (MoPPS) model, extends classical pictorial structures
to recognize multi-part objects whose parts can vary in both type and location and for which
an object part's location can depend on its type. While this model is effective in the American
football domain, it is also broadly applicable.
Finally, I address the problem of tracking football players through video using a novel particle
filtering formulation and an associated discriminative training procedure that directly maximizes
filter performance based on observed errors during tracking. This particle filtering framework
and training procedure are also broadly applicable.
For each of these algorithms, I also present a series of detailed experiments demonstrating
the method's effectiveness in the American football domain. As a further contribution, I have
made the data sets from most of these experiments publicly available. / Graduation date: 2013
| Identifer | oai:union.ndltd.org:ORGSU/oai:ir.library.oregonstate.edu:1957/30348 |
| Date | 14 June 2012 |
| Creators | Hess, Robin W. |
| Contributors | Fern, Alan P. |
| Source Sets | Oregon State University |
| Language | en_US |
| Detected Language | English |
| Type | Thesis/Dissertation |
Page generated in 0.0017 seconds