Hardware-Friendly Activity Recognition with Fixed-Point Arithmetic - Smartphone-Based Human Activity Recognition - page 86

of the relationship between the battery savings and processing time more experimen-
tal tests with different devices and operational conditions would be required.
In current scenarios, even small savings in battery consumption make a big
difference in deciding whether or not to use a mobile app, such as in cases where
the HAR application is required to deliver activity information to other higher-level
decision applications (e.g. phone apps for maintaining a healthy lifestyle through
HAR Lane et al. 2012 ), thus implying sharing system resources. In general, we aim
to build a device able to operate at least during a full day in order to be able to recharge
battery during idle times such as at nights. These results are a good indicator of the
benefits that this method can offer for saving battery life and the possibility of being
integrated into devices for everyday life.
5.5 Summary
In this section we presented a novel energy efficient system for the classification of
BAs using smartphones (HF-HAR). It has been constructed based on amodified SVM
model that works with fixed-point arithmetic (MC-HF-SVM). This approach brings a
significant reduction of the processing time in the prediction of activities, attributable
to the change of arithmetic, leading to a lower use of system resources. Moreover,
results showed the method provides energy savings while maintaining compara-
ble recognition performance when compared with the traditional SVM approach
(MC-LK-SVM).
The proposed model was supported in terms of Structural Risk Minimization
principles, where simpler models are preferred if they have equivalent ability to learn
when compared to more complex approaches. This work is relevant to AmI and AAL
applications where energy consumption becomes a critical issue such as in long term
smartphone-based activity monitoring systems. Similarly, it could be explored the
possibility of using this approach in low-cost devices (e.g. with fixed-point hardware)
for applications including body sensor networks with local prediction of events and
disposable wearable sensing.
The experimental results confirmed that it is possible to substitute the stan-
dard Multiclass SVMmodel with more efficient fixed-point representations. Further
experimentation is required to evaluate the system in more realistic conditions when
the smartphone system shared resources are allocated for different applications, and
also using different smartphones for its evaluation.
References
D. Anguita, A. Ghio, L. Oneto, S. Ridella, in A learning machine with a bit-based hypothesis space ,
European Symposium on Artificial Neural Networks, Computational Intelligence and Machine
Learning, 2013
D. Anguita, A. Ghio, L. Oneto, S. Ridella, in Selecting the hypothesis space for improving the gener-
alization ability of support vector machines , International Joint Conference on Neural Networks,
2011