Database Reference
In-Depth Information
Edge detection methods.
In edge detection based event detec-
tion and tracking, the challenge is to devise a method for nodes to be
identified as “edge nodes” that are near the boundary of a region, and
from that, calculate an approximate boundary for the region in ques-
tion. Three methods guided by statistics, image processing techniques,
and classifier technology are developed and compared in [5]. A novel
method for edge detection of region events makes use of the duel-space
principle [6, 7]. The algorithm is fundamentally centralized, but it can
be distributed among backbone nodes in a two-tier architecture. This
approach, like [5], does not accomplish event labeling.
Existing research on point event detection includes various protocols
such as Distributed Predictive Tracking [8], Dynamic Convoy Tree-based
Collaboration (DCTC) [9] and theoretical contributions [10]. One of the
most notable contributions is DCTC [9]. It uses a “Dynamic Convoy
Tree” protocol to accomplish both event tracking and communication
structure maintenance. DCTC essentially forms and maintains a span-
ning tree over the nodes which senses the event. This is perhaps the most
obvious and straightforward method of detecting events within the net-
work. Moreover, many of the existing high level event detection services
either cite DCTC directly or at least assume a spanning tree structure
like it as part of the middleware needed for their query support.
2.1 Event Models for Sensor Streams
Next we consider two application scenarios of event models for sensor
streams. The first is an o
ine variant in which event detection happens
at the database that stores the measurements collected by the network.
This detection method is used to automatically identify “interesting”
regions within the swaths of data acquired by the sensor network. In
the other online application, motes in the network use events and models
to alter their behavior.
O
ine event detection.
The oine event detection provides a
model suitable for querying events from noisy and imprecise data. Both
database systems [12, 13] and sensor networks [14-16] have explored
model-based queries as a method for dealing with irregular or unreliable
data. Models in these systems include Gaussian processes [14], inter-
polation [17, 18], regression [14-19] and dynamic-probabilistic models
[13-15]. PCA (Principal Component Analysis) based model is specifi-
cally suited to event detection [11]. MauveDB [13] provides a user-view
interface to model-based queries, which greatly extends the utility and
usability of models.
