We provide access to the raw data of our BLE experiments. The data was collected in laboratory conditions between a fixed set of phone models (for the majority of experiments). We do not claim that these experiments are representative of all possible situations that can arise in reality. They serve as an indication and exploration of different design parameters.
Contact tracing requires that exposure (proximity over time) to an infectious subject is measured. Digital contact tracing mechanisms such as DP-3T and GAEN estimate such exposure based on attenuation of BLE signals emitted from two phones. We have to anticipate a large amount of heterogeneity between any two phones:
- Different Bluetooth chips/capabilities (e.g., some phones only transmit at max power);
- Different relative locations (e.g., at ear, outside, on table, in front pocket, in back pocket, in backpack, in purse, in jacket);
- Different obstructions (e.g., people or objects blocking line of sight)
The proximity estimation must take into consideration all these different parameters. Our experiments create a baseline of this parameter space for different phones to estimate proximity together along with an error estimate. We design two sets of experiments: first, laboratory experiments to infer baseline proximity parameters; second, social experiments to test the inferred baseline in a practical setting. The laboratory experiments serve to test individual single aspects through precise measurements and the social experiments test end to end parameters. For both types of experiments we have exact ground truth data to evaluate the precision of our estimation.
The stationary laboratory experiments simulate different static configurations at distances between 1m and 45m.
Here we give a short overview of our dynamic social experiments used to derive basic parameters. We evaluate the following scenarios:
- Lunch environment (with two groups having lunch)
- Train environment (people traveling together in the train)
- Office environment (people interacting in open space setting)
- Queuing at a store (1 or 2 lines with cashiers)
- Free movement of people in crowded space ("party")
- Rapid movement of people similar to a sports environment/gym ("movement")
- People sitting stationary at distance of 0, 1, or 2m with wall or glass between them
Note that, for epidemiological proximity estimation we're not solving the hard problem of assessing exact distance with a few parameters but we are interested in the easier problem of close proximity below a certain threshold over a longer period of time. The current target is shorter than 1.5 meters for 15 or more minutes. These conditions allow us to aggregate over multiple measurements. The experiments we present here are based on individual measurements without aggregation, so serve as an indication of how raw data will look like and will enable us to infer thresholds to assess proximity and, over time, exposure.
While the figures directory contains the full set of figures for the different experiments, we'd like to highlight the lunch, office, and train scenarios, as well as accumulation of all data shown below
These heatmaps show that there is a very low amount of observations with an attenuation below 50dB at a distance of more than two meters. The figures folder also contains precision/recall figures with different thresholds. One example of a precision/recall figure (based on Google's per-model calibration data from 13-08-2020) is as follows:
Based on the data from experiment 34, we also provide an estimation of the GAEN view (version 1.6) and demonstrate that the precision and recall are very close to the precision and recall of the full set of data: the precision is lowered by about 5% while the threshold attenuation values remain approximately the same.
The raw data of the experiments is in the folders of the individual experiments. Each folder contains a description of the individual experiments and their parameters.
We provide the process_raw_data.py script to translate the raw collected
data and ground truth into a set of per-experiment scenario sqlite databases.
These databases can then be used for statistics.
We also checked for linkability of measurements in our datasets. Check out the corresponding report.
After processing the data and combining it with ground truth, we provide
several analysis scripts that process measurement data (sqlite files with
tables containing receiver_id, receiver_model, transmitter_id,
transmitter_model, timestamp, rssi, tx_power, and
ground_truth_distance) to print nice figures for the different experiments
and feature combinations. The resulting figures are in the
figures directory.
The data and contents in this repository is licensed under CC BY-ND 4.0. Copyright of the experimental data and code belongs to EPFL and ETH Zurich. For other uses of the data or code, please reach out via e-mail.