993. Desert
iversity College
tructure of red-
Iscapes, IEEE
sing, 27(4): 441-
nse to soil
irectional
Remote Sensing
id-infrared Data
essment, by Sun
irces Bulletin.
idscape. The
ication series),
SUMMARIZING THE CONTENT OF MOTION IMAGERY DATASETS
Anthony Stefanidis, Peggy Agouris, Panos Partsinevelos
Dept. of Spatial Information Science and Engineering
National Center for Geographic Information and Analysis
University of Maine
348 Boardman Hall
Orono, ME 04469-5711, USA
(tony, peggy, panos} @spatial.maine.edu
Commission V, WG: V/5
KEY WORDS: Video analysis, neural networks, snakes, metadata.
ABSTRACT:
In this paper we present a framework and algorithms for the summarization of motion imagery content to model geospatial
information depicted in it. More specifically, we proceed towards modeling this information by detecting breakpoints in the
trajectories of objects captured in a video dataset, and changes in the outlines of these objects. Our approach to motion imagery
indexing and queries is based on the concept of spatiotemporal lifelines, defined as the trajectories of objects in space and time
during a motion imagery feed. Here we introduce the spatiotemporal helix as a model of spatiotemporal lifelines, providing explicit
yet concise descriptions of object behavior. The helix model is highly suitable for spatiotemporal analysis and offers a powerful
abstraction mechanism for to create brief summaries of motion imagery datasets. These summaries can in turn be exploited to support
content-based motion imagery retrieval.
1. INTRODUCTION
The transition from static to spatiotemporal analysis is
becoming increasingly evident in the geospatial community.
Regarding image analysis this signifies an evolution from single
images to collections of time-varying imagery. Time-varying
imagery collections may range from continuous video segments
to sequence of static images that differ by seconds, minutes, or
even days, depending on the temporal resolution of the event
that they are used to describe. We use the term motion imagery
(MI) to refer to these multitemporal image datasets. The
processing and analysis of spatiotemporal datasets is
introducing interesting data handling challenges, mostly
associated with the large volumes of datasets, the corresponding
processing times, and the diverse nature of information
contained in them.
The efficient modeling of spatiotemporal events is a major
research challenge and an important step towards the analysis
and management of large spatiotemporal datasets. Relevant
research includes the work of [Smith & Kanade, 1995] on the
analysis of visual and speech properties to construct “skim”
video synopses by merging select segments of the original
video. The extraction of select key frames for the generation of
video summaries has also been addressed in [Yeung & Yeo,
1997]. [Pfoser and Theodoridis, 2000] provide a spatio-
temporal synthetic dataset generator to simulate movement
trajectories, analyze novel index schemes for moving points
using tree structures. The indexing and querying of moving
points is also addressed in [Vazirgiannis & Wolfson, 2001],
while [Sistla et al., 1997; Wolfson et al., 1999] discuss the use
of future temporal logic for modeling and querying moving
objects. Work on indexing animated objects is reported in
[Kollios et al., 2001], while [Tao & Papadias, 2001] propose a
framework for indexing and querying spatiotemporal data by
constructing new tree structures.
In [Stefanidis et al., 2001] we introduced a general framework
for the summarization of spatiotemporal trajectories considering
point datasets (a point changing its position over time). This
was in accordance to the above-mentioned relevant works,
where moving objects are reduced to a point representation,
ignoring spatial extent of objects and the variations of their
outlines. In this paper we move beyond this simplification,
extending the framework introduced in [Stefanidis et al., 2001]
to accommodate the spatial extent of objects. This allows us to
consider not only the movement but also the deformation of
spatial objects, introducing a more comprehensive
spatiotemporal model than the currently existing ones. At the
core of our work is the concept of the spatiotemporal helix, a
novel spatiotemporal object model. It allows us to model object
movements and deformations, supporting complex
spatiotemporal analysis.
This is a key development to support the analysis of
spatiotemporal phenomena that have certain spatial extent and
change their position and/or extent over time. Such phenomena
can be slowly moving (e.g. urbanization trends depicted in a
series of monthly satellite images) or rapidly evolving (e.g.
hurricanes depicted in hourly or daily datasets), and they take
place over a fixed area (e.g. flooding) or may be constantly
changing their location (e.g. a moving fire front).
The rest of the paper is organized as follows: In Section 2 we
introduce the concept of the spatiotemporal helix as a concise
representation of spatiotemporal events. The automated
generation of spatiotemporal helixes makes use of two
automated techniques we have developed: a technique based on
self-organizing maps for the generalization of point trajectories
(Section 3) and differential snakes, an extension of the model of
deformable contour models to perform outline comparison
(Section 4). Section 5 addresses spatiotemporal analysis issues
using helixes with final comments following in section 6.
—515—
