Conclusions - Smartphone-Based Human Activity Recognition - page 125

Chapter 8
Conclusions

8.1 Achievements

In this thesis we presented a collection of contributions related to the recognition of
human activities in the AmI framework. We exploited existing commercial hardware
(smartphones) and state-of-the-art ML algorithms (SVMs) in order to contribute with
the design of human-centered services that improve people's QoL. In particular, here
we summarize our most relevant achievements:

￿ We developed a HAR system for the real-time classification of BAs using a sin-

gle waist-mounted smartphone device (PTA-HAR). It can provide activity infor-
mation to other context-aware applications within the same device or externally
located through wireless communications (e.g. for daily monitoring systems for
the elderly and physical activity trackers for athletes). The system also handles PTs
through learning and filtering in order to improve the classification performance
during transitions between BAs. Human bodymotion signals are continuously cap-
tured through the device's accelerometer and gyroscope. These signals are then
processed and segmented into windows to extract relevant activity features in the
time and frequency domain. The features become the input of an ML algorithm: a
multiclass linear SVM (MC-L1-SVM) which allows to make activity predictions.
Finally, consecutive predictions in time are post-processed in order to minimize
classification errors through a filtering module (Table 7.2 ) which is aware of PTs
and considers the relationship between contiguous activities in real-life.

￿ We proposed a novel hardware-friendly SVM intended to predict activities using

only fixed-point arithmetic (HF-SVM) to be applied to mobile devices for lower
power consumption or hardware without floating-point units (e.g. low-cost dispos-
able wearable sensors). We found that the classification of the selected BAs allows
a reduction of the number representation of the sensor data (e.g. from 32 to 6 bits)
without substantially damaging the system performance (Fig. 5.3 ) . In this way, we
can control over model complexity and accuracy in order to improve recognition
speed and reduce battery consumption against typical floating-point based SVM
formulations.