Graphics Reference
In-Depth Information
and discuss the state-of-the-art and challenges in design and implementation of
effective ASM systems for embedded computer vision. ASM systems can be viewed
as real-time data mining systems that operate on streams of data and are constructed
as topologies (directed graphs) of classifiers, where parameters associated with the
topologies and constituent classifiers may be manipulated dynamically based on
changes in data characteristics, operational constraints, and other relevant run-time
considerations.
Intended applications of ASM systems for embedded computer vision are very
diverse, ranging from medical services, to dynamic management of vehicular traf-
fic, to real-time detection of events in home-based health-care, to many kinds of
surveillance and environmental monitoring applications. Each of these applications
requires a topology of classifiers (such as a chain or “pipeline” configuration) that
analyzes streaming data (which dynamically changes over time) from a set of raw
data sources to extract valuable information in real time.
The need for adaptivity in ASM systems is inherent in almost all practical
knowledge extraction application areas as data characteristics and operating con-
ditions often exhibit uncertain or time-varying behavior. Accurate assessment,
understanding, and optimization of ASM systems generally requires extensive
experimentation of how algorithms for data classification and classifier adaptation
interact with the characteristics of input data, and how scheduling and buffer man-
agement for such algorithms should be performed to satisfy real-time constraints
subject to given resource constraints.
Decomposing applications as topologies of distributed processing operators has
merits that transcend the scalability, reliability, and performance objectives of
large-scale, real-time stream mining systems [
1
,
11
,
18
,
27
]. Specifically, many
stream classification and mining applications implement topologies (ensembles such
as trees or cascades) of low-complexity binary classifiers to jointly accomplish the
task of complex classification [
24
]. Such a structure enables the successive identifi-
cation of multiple attributes in the data, and also provides significant advantages in
terms of reduced resource consumption through appropriate dynamic data filtering,
based on the incrementally identified attributes.
It has been shown that using a tree of binary classifiers can achieve better
performance compared to other techniques such as support vector machines or SVMs
(e.g., see [
10
]), rule-based techniques, and neural nets for some applications [
6
,
11
,
19
,
26
,
28
,
31
,
42
]. Furthermore, using classifiers operating in series with the same
model (boosting [
31
]) or classifiers operating in parallel with multiple models (bag-
ging [
19
]) has resulted in improved classification performance.
12.2 ASM System Example
Consider the surveillance application depicted in Fig.
12.1
. A straightforward
approach to dealing with this application requires the cameras to acquire the images
on a continuous basis with the highest resolution, and send them to a central process-
ing unit that is responsible for analyzing the images with complex data analytics.
