Improving Robustness of Homography Estimation for Ice Rink Registration
Loading...
Date
2023-08-28
Authors
Shang, Jason
Journal Title
Journal ISSN
Volume Title
Publisher
University of Waterloo
Abstract
Hockey analytics involves obtaining information from games so that coaches, managers,
and teams can make better decisions in training, playing, and hiring. As there is a large
amount of information available in each game, manual analysis is difficult and tedious, so
automated computer vision techniques have been developed to acquire and process data
more efficiently.
One key component to such analysis is the location information of players and events.
This information can be obtained using a technique called rink registration, which involves
estimating the homography matrix needed to warp an overhead template of the rink onto
video frames, or vice versa. By doing this, we can obtain the location of objects in video
with respect to the fixed reference frame of the overhead template. Current methods focus
on NHL rinks, which have a standardized size and have similar appearances. However, the
quality of results drop when other types of rinks are used, because the existing methods
are not trained to work on non-NHL rinks.
This work seeks to improve the rink registration process by making it more robust to
differences in rinks, while maintaining good accuracy. It also tries to develop a generalized
system that can work on a variety of rink types, such as NHL, Olympic, and European,
without the need for additional rink-specific training or expensive annotations. By developing
this rink-agnostic system, it can provide rink registration results regardless of rink,
making analysis more equitable for smaller groups. It also reduces the cost needed as it only
requires broadcast video and the overhead rink template, without the need for additional
technology to be installed or annotations to be made. The results of this rink-agnostic
system are competitive with the results of an NHL-only baseline on NHL rinks and are
noticeably better than the baseline on non-NHL rinks. The rink-agnostic system achieves
a 1.1% IOUpart improvement on the Olympic 2014 rink and a 8.8% IOUpart improvement
on the Berlin Mercedes-Benz Arena rink.
Description
Keywords
machine learning, homography, computer vision, deep learning, image and video processing, segmentation