Database Reference
In-Depth Information
lihood classifier, and an SVM classifier. The local decision, together with
the estimated probability of being a correct decision is transmitted to a
local fusion center rather than sending the raw data. The fusion center
can then use maximum aposterior (MAP) estimate to compute the final
decision on the classification of the observation. Extensive experimental
results show that the MAP estimate with the nearest neighbor as the
local classifiers works well in vehicle classification.
Some other important work in this area include the distributed target
classification work by Brooks
et al.
[11], Gu
et al.
[29], and Kotecha
et
al.
[36].
Another area similar to distributed classification in WSN is distributed
event detection. The main goal is to detect frequent event patterns based
on some data mining models while minimizing the need for communi-
cation all the data from all the nodes to the sink. One such method
is the technique based on frequent itemset mining by Romer [49][50].
First local association rules are learned at each node and then these
rules along with the support and confidence are sent to the sink. Ex-
perimental results demonstrate that this method is ecient and detects
correct frequent events. Wittenburg
et al.
[57] present a method for
distributed event detection. Their method consists of sampling the data
in the network, feature selection and then learning a model at each node.
Tavakoli
et al.
[55] consider a scenario in which targets are tracked
using an undersea acoustic sensor network. The sensor nodes report their
local classification result to a cluster head which then in turn performs
an evaluation of the data and may report the outcome to a base station.
As a confidence interval, the method considers the accuracy of these past
reports.
The system proposed by Yang
et al.
[61] is aimed at recognizing
human motions. It is a wearable sensor system consisting of eight sen-
sor nodes attached to the body of a person who may perform one out
of twelve actions. Accelerometer and gyroscope are used to detect the
motions and then features are extracted and classified at each node to
detect events. If a local classification is promising, the data of all nodes
is transmitted to the base station and classified once again. The classifi-
cation process identifies an action by matching the linear representation
of the extracted feature vector to one of several subspaces, each of which
corresponds to one type of action.
4. Outlier Detection in WSN
Outlier detection is one of the most critical tasks performed in WSNs
due to their ability to monitor hostile environments. In general, the
