Human Activity Dataset Generation - Smartphone-Based Human Activity Recognition - page 72

Table 4.9 HAR Competition
References

Approach implemented Accuracy (%)

Romera-Paredes et al. ( 2013 ) OVO Multiclass linear SVM with majority
voting

96.40

Kästner et al. ( 2013 ) Kernel variant of learning vector

quantization with metric adaptation

96.23

Reiss et al. ( 2013 ) Confidence-based boosting algorithm

Conf-AdaBoost.M1

94.33

Test data classification accuracy of the best performing approaches

4.4.2 HAR Competition

A competition targeting the development of novel learning approaches for the clas-
sification of a set of activities was planned as part of a special session in Human
and Motion Disorder Recognition at the European Symposium on Arti?cial Neural
Networks (ESANN) in 2013. Competitors were challenged to submit their proposals
given the HAR training data from

D 2 before its full publication. Participants were
also provided with an unlabeled test set in order to receive from them the predicted
labels for each sample. The performance of their approaches was measured in terms
of classification accuracy using the experiment ground truth.
We received proposals from different universities and research centers in Europe.
The three best contributions are depicted in Table 4.9 . In Romera-Paredes et al.
( 2013 ), a One-Vs-One (OVO) Multiclass SVM with linear kernel was proposed
for the classification task. The method used majority voting to find the most likely
activity for each test sample from an arrangement of 6 binary classifiers. An overall
accuracy of 96.40% was reached on the test data and this method became the com-
petition winner. For comparative purposes, they also evaluated the performance of a
OVA SVM and a k-NNmodel which exhibited poorer accuracies (93.7% and 90.6%
respectively). In the same way, a sparse kernelized matrix Learning Vector Quan-
tization (LVQ) model was employed in Kästner et al. ( 2013 ) for the HAR dataset
classification achieving 96.23% test accuracy, only differing 0.17% against the first
approach. Their method was a variant of LVQ in which a metric adaptation with only
one prototype vector for each class was proposed. Ultimately, a novel confidence-
based boosting algorithm (Conf-AdaBoost.M1.) was presented in Reiss et al. ( 2013 )
and assessed against the traditional decision tree classifier and the AdaBoost.M1
algorithm. The method is a direct multiclass classification approach which exploits
confidence information from weak learners for the classification. They achieved an
accuracy of 94.33% on the test set.
The results of this competition also show evidence of the benefits of using SVMs
for HAR as its winner also employed this algorithm. Their approach was slightly
different than our OVA MC-GK-SVM algorithm but the performance was similar.
In the following chapters, we present variations of the original SVM formulation in
order to solve the recognition problem and adapt it to limited hardware such as the
smartphone.