Activity
recognition dataset - BodyAttack Fitness dataset
Daniel Roggen, Wearable Computing Laboratory, ETH Zurich
droggen@gmail.com
Initial documentation: 06.09.2010
Last updated: 06.09.2010
Description
This dataset contains 6 fitness activity classes, done mostly with the
legs.
This dataset was collected by Kilian Förster to investigate the
effect of sensor displacement on activity recognition performance
[Förster09].
Availability
This dataset can be freely used in publications provided the following
paper is cited: [Förster09].
Sensors
This dataset contains 6 activity classes, recorded 10 sensors placed on
the right leg of the subject, at regular intervals.
Sample rate: 64Hz.
Activities
- 6 fitness activities
- 1 subject
- 5 runs
- About 30 seconds per activity class and run (i.e. total of
2.5mn/activity class, and total recording length of 15mn)
Files
baclass_20090317 contains the segemented dataset. Load into matlab with
the load command.
Variables loaded: datasetall.
Format: acceleration = datasetall{sensor}{activity}{run}
sensor number (30 sensors: 10 3-axis sensors = 10x3 = 30
axis). With sensor=0...29, the sensor number is mod(s,3). The sensor
number corresponds to the
above figure.
Acceleration is calibrated in milli-g units (1000 =
earth gravity vector).
activity is:
- 1: flick kicks
- 2: knee lifts
- 3: jumping jacks
- 4: superman jumps
- 5: high knee runs
- 6: feet back runs
run is the recording number.
Caveats
Due to data acquisition issue, some data intervals were lost. The lost
inverals were replaced by data of same length from the same sensor,
same activity and same run (unless otherwise noted). This has no
influence on a window-based activity recognition. It may however lead
to discontinuities in the signal where a data interval is repeated.
Essentially, however, this dataset may be used as-is for many activity
recognition problems.
The following sensors / activities / runs are concerned (indicated the
samples that were replaced):
- Sensor 6, run 3, activity 4: samples 2800->end
- Sensor 6, run 3, activity 5: samples end-1000->end
- Sensor 6, run 3, activity 6: all samples (data copied from sensor
2)
- Sensor 7, run 4, activity 1: samples 1720->end
- Sensor 7, run 4, activity 2: samples end-620->end
- Sensor 9, run 3, activity 1: samples end-1000->end
References
- [Förster09] K. Förster, D. Roggen, G. Tröster. Unsupervised Classifier Self-Calibration
through Repeated Context Occurences: Is there Robustness against Sensor
Displacement to Gain? In Proc.
13th {IEEE} Int. Symposium on Wearable Computers (ISWC 2009),
pages 77-84, 2009
