Calibration and Tracking in Sensor Networks

Sensor networks deployed over wide areas make it possible to gather data that was once completely inaccessible to researchers and engineers. Networks of cameras have been common in airports for years, and small wireless sensor nodes equipped with several sensors will be deployed everywhere if some people get their wish. For most tasks, these networks must first be calibrated: their location must be determined, and their sensor parameters must be continually updated in the field.

	Various sensor nodes
Extendible mote with wireless	IPAQ with camera and 802.11 card	MIT Cricket with ultrasound and RF ranging

SLAT and LaSLAT (with Chris Taylor, Jonathan Bachrach)

The SLAT problem (Simultaneous Localization and Tracking) is to calibrate and localize the nodes of a sensor network while simultaneously tracking a moving target. When the sensor nodes were calibrated and localized, tracking the moving target is easy. Alternatively, when the location of the moving target is known, localization and calibrating the sensor nodes is be easy. In most cases, however, we are faced with a chicken and egg problem where none of these things are known beforehand.

This is similar to SLAM (Simultaneous Localization and Mapping), a problem in robotics whereby a robot must build a map of the environment while localization itself within this map.

We adapted SLAM techniques to SLAT by enhancing the internal representation of the uncertainty in the sensor localization and calibration parameters using Laplace's approximation. So we call our method LaSLAT (Laplace-based Simultaneous Localization and Tracking).

A target node travelled around a room littered with cricket sensor nodes (this is a top view of the room from a mezzanine). The trajectory is marked in red. The nodes are circled in blue.	The algorithm updates its estimate of the sensor nodes in real time. This figure shows the estimated sensor locations (red crosses) at the end of the trajectory. The nodes are localized to within 2cm.

This movie(18M) shows that the estimate of the location of the sensors is improved as a target moves around. The target is marked by X. A green line connects the true sensor location (shown in white) with the estimated sensor location (shown in green). The images are above depict the final state of this run.

The method works with fairly large networks. Here's a deployment in the lobby of our computer science building:

And it works in 3D

For details, see this paper:

Simultaneous Localization, Calibration, and Tracking in an ad Hoc Sensor Network, C. Taylor, A. Rahimi, J. Bachrach, and H. Shrobe,
in Information Processing in Sensor Networks (IPSN) 2006. (pdf, Powerpoint presentation)

Non-overlapping Sensor Networks (with Brian Dunagan)

A significant improvement over a traditional tracking sensor network is a sensor network where the field of views of the nodes are not required to overlap. For the same number of nodes, this type of network provides a larger coverage area. Assuming that the calibration target does not change its acceleration abruptly, a technique similar to the above can be applied to non-overlapping sensor networks. This work shows that knowledge about the dynamics of a target can compensate for the lack of overlap in the sensor network.

This figure depicts localization and tracking results from seven non-overlapping wireless IPAQs equipped with cameras, mounted on the ceiling of an office environment. The true location and orientation of each camera is depicted by a dashed square. The recovered location and orientation of each camera is depicted by a gray box. The environment is about 12 meters on each side. The L shape is a wall corner. On average, the recovered location of the sensors was off by about 30 cm, which is about 2.5% of the size of the envionment.

Here is a video (3.4MB) showing the convergence of our system operating on a smaller data set.

For details on how to calibrate networks of non-overlapping sensors, please look at: