Visualization of Sensor Network Data

An emerging area of research in environmental monitoring regards massive networks of motes, small independent sensors ("smart dust"). Deployed in an area of interest, or inside a "smart building," motes self-organize into wireless networks and can gather data, for example about weather, the presence of toxins in the environment, structural soundness of buildings during/after an earthquake, and the progression of hazards such as fires to direct evacuation of building occupants or guide firefighters. Research into smart dust is currently being funded mostly by two sources: The Defense Advanced Research Projects Agency (DARPA), a subsidiary of the Department of Defense who focuses on military and battlefield uses of sensor networks, and the Center for Information Technology Research in the Interest of Society (CITRIS), a collaboration between four University of California campuses sponsored by the state of California. CITRIS focuses on civilian uses of sensor networks, especially the development of smart buildings.

In smart dust, each miniature sensor ("mote") contains a sensor package, a power supply, a CPU, memory, and a wireless network link. Guided by the TinyOS operating system developed at UC Berkeley, motes will self-organize into a hierarchical network upon deployment, and report measurements from their sensor packs towards the root node, where data is downloaded into a database. If one thinks of the measured quantity, e.g., temperature or humidity, as a time-varying mathematical function f(x, y, z, t), then each mote i reports a stream of samples f_i(t) of that function, taken at the mote's position (x_i, y_i, z_i). In order to analyze or visualize the measured function f, one has to reconstruct it from the samples reported by the sensor network.

Scattered data methods are mathematical techniques developed for this purpose. Each one implies an interpolation scheme, using all samples or a selected subset of samples to evaluate the unkown function f at any given position (x, y, z). "Classic" scattered data techniques are typically not geared towards real-time visualization of massive time-varying sets of samples, so it is necessary to adapt existing methods to this purpose or develop new ones. Furthermore, all classic methods assume that the source function for all samples is defined over a simple (convex) domain, typically implicitly taken to be the convex hull of all sample locations. This assumption fails when dealing with data generated by smart buildings; in those, boundary conditions like the presence of walls, the layout of air ducts, etc. have to be taken into account to achieve faithful reconstruction.

Figure 1: Reconstruction of two-dimensional scattered data. The source function (in this case a 400 x 300 pixel RGB color image) was sampled at 1,600 "random" points and then reconstructed using Sibson's interpolation method.

Project Goals

The main goals of this project are to develop scattered data methods suitable for real-time visualization and analysis of sensor network data, and to adapt generic scattered data methods to complex, non-convex domains, e.g., building interiors. This includes the following detail goals:

Design of a generic modular scattered data reconstruction package, with the ability to mix and match different domain representations, distance metrics and interpolation algorithms.
- Implement generic versions of standard scattered data methods:
  - Shepard's method (global and local)
  - Hardy's method (global and local)
  - Sibson's method
  - Triangulation methods (Delaunay and data-dependent triangulations)
- Implement generic domain representation.
- Implement distance metric (used to generate interpolation weights) specific for each domain representation.
Design a domain representation scheme appropriate to represent building interiors, and a compatible distance metric.
Design scattered data visualization methods suitable for overlaying rendering of domain with rendering of data.
Develop interactive viewer for domain representation
- Desktop viewer to visualize overview of important features of measured fields.
- (Virtual reality) building walk-through with embedded field visualization.

Project Status

At this point we are very early in the project. We developed and implemented a generic scattered data reconstruction package, a domain representation for building interiors, and a simple initial distance metric for building interiors. We are currently in the process of implementing interactive viewers for domain representations only, and visualization methods for time-varying scattered data.

On a second front, we are currently implementing a prototype interactive visualization program for a specific time-varying sensor network data set that monitored the nesting burrows of a species of seabird on an island off the coast of Maine without disturbing the bird population. The sensor network project is described in detail on the Great Duck Island home page. More information about the current state of our Great Duck Island visualization project can be found on Valerie Szudziejka's sensor networks web page.

Related Publications

I gave a talk about this project (before it started) in the context of the MEMS seminar at UC Berkeley a while ago. The presentation slides are available for download (PDF format, 2,447KB))

Pages In This Section

Approximation Examples: Examples of reconstruction of a two-dimensional scattered data set using several classic scattered data methods.