Scribd
Upload
113 views

Deep Learning For Sensor-Based Activity Recognition

1) The document discusses using deep learning techniques like CNNs and RNNs with LSTMs and GRUs for sensor-based activity recognition using triaxial accelerometer data from wearable devices. 2) It compares using single or multiple accelerometers placed on different body locations and finds that multiple sensors can improve recognition performance but thigh and wrist sensors performed almost as well. 3) It explores using temporal windows of accelerometer data as inputs to classifiers and finds a tradeoff between window size, recognition delay, and number of training examples.

Uploaded by

Anđela Todorović
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
113 views

Deep Learning For Sensor-Based Activity Recognition

1) The document discusses using deep learning techniques like CNNs and RNNs with LSTMs and GRUs for sensor-based activity recognition using triaxial accelerometer data from wearable devices. 2) It compares using single or multiple accelerometers placed on different body locations and finds that multiple sensors can improve recognition performance but thigh and wrist sensors performed almost as well. 3) It explores using temporal windows of accelerometer data as inputs to classifiers and finds a tradeoff between window size, recognition delay, and number of training examples.

Uploaded by

Anđela Todorović
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 1

Machine learning conference – Subotica, Serbia – Oral presentation, Andjela Todorovic

Deep Learning for Sensor-Based Activity Recognition

A. Todorovic, I. Mihajlovic1,
1
Research team, Createsi GmbH, Horgen, Switzerland

Abstract: acceleration feature values at different sites


were useful for specifying many activities.
Human activity detection has seen immense However, with just two triaxial accelerometers–
growth in the last decade playing an im- thigh and wrist–the recognition performance
portant role in the field of pervasive compu- dropped only slightly.
ting. This popularity can be credited to its Rather than attempt to classify every single data
real-life applications primarily dealing with point it made sense to compute features of the
human-centric problems in general medicine. data that are based on some kind of temporal
Many research attempts with data mining and window. The longer the window length, the
machine learning techniques have been un- more these positive benefits are obtained; how-
dergoing to accurately detect human activities ever if the window length becomes too large, the
for healthcare systems. probability that a given window contains more
than one activity is increased, the delay is in-
In this paper, we are concerned with accelerom- creased, and the number of training examples
eter-based Activity Recognition (AR). Firstly, for the classifier will also be decreased.
we need to clarify the difference between activ- We have considered advanced neural network
ity tracking and activity recognition: whereas based models for AR with CNNs and also inves-
the former is only concerned with estimating tigated the benefit of convolutions and recur-
general levels of activity the latter is attempting rency for feature learning and prediction. Two
to determine the actual activities which are oc- popular methods of incorporating recurrency in
curring. Hereby we discuss only the activity Deep Neural Networks include LSTM and
recognition using the tri-axial accelerometers Gated Recurrent Unit (GRU). These two tech-
providing a high-fidelity measurement of force niques achieve tremendous performance on se-
along the x, y, and z-axes and thus provide a quence prediction tasks since they directly pa-
view into the movement of the person wearing rameterize the influence of nearby data points
the device. Although there is significant poten- on predicting the current time point, likely to
tial for accurately predicting activities of the Markov Models and the neighborhood of influ-
daily routine with accelerometers, many open ence is optimized during the learning phase.
problems exist due to the sheer volume of rea- The CNN configuration is chosen since a com-
sonable configurations available. Accelerome- plete analysis LSTMs over all sensor configura-
ters may, for example, be configured with spe- tions would command particular care in specify-
cific sampling rates, sample resolution and ac- ing the network architectures.
celerometer range, where features can be ex-
tracted from windows of any size, and the selec- Keywords: deep learning; activity recognition;
tion of the ultimate data analysis and classifica- activity tracking; wearable devices
tion pipeline is also non-trivial. Different con-
References:
figurations may have an impact on the predic-
tive performance of an AR classifier, and so 1. J. Wang et al., Deep learning for sensor-
these parameters must be carefully chosen. based activity recognition: A Survey, Pat-
To build a classifier for wearable accelerometer tern Recognition Letters (2018),
data, we need to first collect the data from wear-
able sensors placed on the body of the subject 2. M.A. Alsheikh, A. Selim, D. Niyato, L.
on the hips (belt), upper arms, wrists, ankles, Doyle, S. Lin, H.-P. Tan, Deep activity
and thighs. recognition models with triaxial accelerom-
The results which considered locations of the eters, AAAI Workshop (2016).
above suggested that multiple accelerometers
aided in recognition since connections between

You might also like