RFID Data Overview

We have publicly released here an RFID dataset collected using our building-wide RFID deployment. We will add to this dataset as we continue to generate new data. In addition to the raw RFID traces, this release also includes the Markovian stream data inferred from the raw readings; ground-truth location labels; and animations showing the raw readings, inference process, and ground truth overlaid on a map of the RFID environment.

RFID Environment

The data we publish here were collected in an RFID (Radio Frequencey IDentification) deployment called the RFID Ecosystem in the Computer Science \& Engineering building at the University of Washington. The Ecosystem is a deployment of nearly 150 RFID antennas spread throughout the hallways of all six floors of the building. These antennas detect RFID tags that pass near them, and log the sightings along with their timestamp in a database. There are over 100 tags currently active in the system, associated with individuals who regularly visit the building, as well as their keys, mugs, notebooks, etc. For more information about the RFID deployment, please visit the RFID Ecosystem website. Floor plans of the building as well as the locations of the RFID antennas can be downloaded at the bottom of this page.

Data Scenarios

While our RFID deployment is constantly collecting data, the data we publish here was collected by volunteers following pre-scripted paths meant to simulate normal daily activities. This simulation allows us to 1) collect ground truth location information through manual annotation.

Understanding The Trace Data

For each data trace that we have made available, we provide an animation of the trace as well as a tarball of all relevant data.

The data tarballs contain the following (download detailed formatting explanations here):

  1. Raw, timestamped RFID tag readings.
  2. Filtered marginal distributions over the tag location at each second, inferred from an HMM via particle filtering.
  3. Smoothed marginal distributions over the tag location at each second, along with correlations between the marginal distributions at adjacent seconds.
  4. Raw, timestamped ground truth annotations (latitude/longitude/floor) as recorded by our volunteers.
  5. Ground truth annotations snapped onto the building's connectivity graph.
  6. Interpolated ground truth annotations at each second, computed via linear interpolation between the timestamps for which ground truth location is provided.
The animations show the following:
  1. Raw ground truth annotations collected by the study participants.
  2. Interpolated ground truth tag location for each second.
  3. Smoothed marginal distributions over tag location at each second.
  4. Active and inactive RFID antennas at each second.
  5. Discrete locations within the building and connectivity between them.
Below is a key identifying the different elements of each animation:

Trace Downloads

Please see the above notes for an explanation of the data below.

"Good" Traces

These traces contain some noise, but are largely clear about the room that the tag is in at any given time. Traces 1-5 contain several entered/exit room events each, while Trace 26 reflects a tag entering/exiting a room many many times in sequence.

"Ambiguous" Traces

These traces contain some noise and each also contains at least one fundamentally ambiguous situation. Usually this situation involves uncertainty about a tag's location because two rooms are symmetric with respect to the sensor deployment. Traces 12-16 contain several entered/exit room events each, while Trace 27 reflects a tag entering/exiting a room many many times in sequence.

Auxiliary Information