The anonymized "shaking" data sets used for finding parameter sets and evaluation of the "Shake well before use" paper are freely available for research purposes. We only ask to reference the original publication (or the upcoming extended journal article) when using the data sets:
R. Mayrhofer and H. Gellersen, “Shake well before use: Authentication based on accelerometer data,” in Proc. Pervasive 2007: 5th International Conference on Pervasive Computing, vol. 4480 of LNCS, pp. 144-161, Springer-Verlag, May 2007
The data sets are split into three experiments:
- Experiment 1: 51 people shaking both devices in their left (*-left-*.log) and in their right hands (*-right-*.log), and each device in one hand (*-both-*.log). This data set further distinguishes between sitting (sitting-*.log) and standing (standing-*.log), exhibiting slightly different motions. Samples are about 5 seconds long.
- Experiment 2: 12 pairs of people shaking one device each, either both in their left (*-both-left-*.log) or the right hands (*-both-right-*.log), or subject 1 in their left and subject 2 in their right (*-s1-left-s2-right-*.log) or vice versa (*-s1-right-s2-left-*.log). Samples are about 15 seconds long.
- Experiment 3: 31 people trying to pair the devices with both protocols in live mode and immediate feedback. In this data set, the samples are continuous and not segmented.
Each of the archives containes the samples (*.log) as well as statistical information about each subject (*.inf) in the format
<left or right handed> <sex> <age>
r m 27
For experiment 2, these files contain two such blocks, one for subject 1, the second for subject 2.
The data format for the samples is line-based ASCII with microsecond accuracy timestamps (format seconds.microseconds) followed by the accelerometer values of all 8 channels. There are 4 channels per device, corresponding to 2x 2D accelerometers with a maximum acceleration of 2g. The accelerometers were mounted at an angle of 90°, so that all 3 dimensions were sampled. For each device, it is therefore sufficient to use the respective first 3 values of each line, while the 4th is redundant. A new line in these data files is written whenever any of the 8 channels changes its sample value, so the sample times corresponding to consecutive lines are not spaced equally. The sample rate for each of the channels is roughly 600 Hz, that is, the values for each channel change at about this rate.
These log files can be imported into most software packages or read easily using custom code. For example, in Matlab or Octave, the data files can be read and re-sampled with
f=dlmread(filename); g=sample_data(f, samplerate);
with filename and samplerate (in Hz) set appropriately. sample_data.m is a small script to down-sample the original data to a given sample rate. Note that the down-sampling is not done precisely, but only by averaging.
Another example code for reading and down-sampling this data format can be found in the current OpenUAT code in form of the ParallelPortPWMReader class.