center for robotics and embedded systems University of Southern California Viterbi School of Engineering

We present an online method for automatically segmenting the data stream generated at a sensor node into distinct sub-sequences. The data at each node are assumed to be generated by a stochastic process. The data stream (sequence) sensed at each node is split into a new sub-sequence when the Kullback-Leibler distance between the probability distributions induced by the previously observed data and that induced by the most recently observed data exceeds a threshold. A distributed region growing algorithm is used to cluster spatially proximal sensor nodes that observe similar data streams. Thus, the space being monitored by the sensors can be divided into regions that observe similar activity. Continuous data segmentation and region growing enables changes in the pattern of activity to be observed as shifting spatial regions. We provide an analysis of the time taken for convergence and the power required for transmitting messages when the nodes are arranged in a regular grid and when they are randomly distributed. We present simulation results that show the performance of these algorithms on a shadow tracking application. The data used in these simulations was obtained from a sensor network equipped with light sensors and deployed outdoors.


Go Home
Maintained by webmaster