Fault tolerant event boundary detection and target tracking in sensor networks.

摘要

In the last decade, sensor network has been emerging as an indispensable application in the area of biological observation [31, 72], security surveillance [33, 59], traffic monitoring [5, 13], earth activity recording [71, 74] and others. Detecting event frontline or boundary sensors and tracking dynamically moving events in a complex sensor network environment are critical problems for sensor network applications. By considering the nature of sensor data, general data mining techniques are not directly applicable, which motivates us to investigate collaborative, distributed data mining methods that enables efficient distributed computation by individual sensor nodes with limited computation power and memory storage. In this thesis, we provide two classes of distributed in-network processing schemes, based on different task requirements, for outlier sensor detection, fault-tolerant event boundary detection and target tracking in sensor networks.;We first propose robust Median estimator based approaches for identification of outlying sensors and detection of the reach of events in sensor networks. To identify outlying sensors, median is used as the estimation of the observation in a close proximity. Accordingly, an outlier is detected by collaborative in-network comparison in the close proximity. As for event frontline detection, a special proximity is chosen such that the measurements of a sensor node close to the real event boundary significantly differentiate with the local sensing estimation in this special neighborhood.;We next introduce our exploration of using statistical clustering methods with model selection analysis [1, 2, 28, 67] for distributional sensor data modeling and event frontline sensor detection [25]. A Boundary sensor is considered as being associated with a multimodal local neighborhood of (univariate or multivariate) sensing readings, and each Non-Boundary sensor is treated as being with a unimodal sensor reading neighborhood. Furthermore, the set of sensor readings within each sensor's spatial neighborhood is formulated using Gaussian Mixture Model. Two classes of Boundary and Non-Boundary sensors can be effectively classified using the model selection techniques for finite mixture models. We further propose its temporally adaptive version for dynamic target tracking in changing environments, under a unified statistical mixture modeling framework. The proposed algorithms can be implemented within each purely localized sensor neighborhood and scale well to large-range sensor networks. The computational complexity is moderate and comparable to our previous Median based approaches. Our extensive experimental results demonstrate that our algorithms effectively detect the event boundary with a high accuracy under moderate noise levels. Desirable quantitative target tracking results are also achieved under challenging background conditions.