Online Recognition with Postural Transition Awareness - Smartphone-Based Human Activity Recognition - page 107

Chapter 7
Online Recognition with Postural
Transition Awareness

7.1 Introduction

In this chapter we introduce an online system for the classification of human activities
using smartphones: PTA-HAR. This system is an adaptation of the previously pre-
sented L-HAR system that deals with frequent PTs while sequences of BAs are
carried out. As it will be shown, when PTs are not properly identified in a HAR
system, they can affect its performance by triggering the appearance of false posi-
tives. In most HAR systems, transitions between activities are usually ignored since
their incidence is generally low and duration is shorter when compared against other
activities. Nevertheless, this assumption depends on each application and, therefore,
it should be considered accordingly. For instance, in the design of activity moni-
toring systems for the disabled during rehabilitation practices, or for athletes while
performing fitness/gymnasium activities, it is important to identify PTs because in
these cases is common to do different tasks in short time periods. Incorrect classi-
fication of PTs can significantly affect the performance of the recognition system
because transitions appear more frequently.
The proposed system to deal with PTs employs a probabilistic analysis of con-
secutive activity predictions. More concretely, two different methods to handle these
transitory events are explored. Both follow a similar recognition pipeline although
in one of them PTs are learned by the ML classifier (PTA-7A) along with the BAs,
while in the other they are handled only through filtering (PTA-6A).We employ linear
SVMs whose probability estimates are analyzed in conjunction with the predictions
of its neighboring samples in time and interpreted as activity signals. These signals
are then heuristically filtered in order to clean and suppress unwanted noise since we
assume that, in real-life applications, contiguous events are in general correlated.
We present the benefits of these two methods and show, through experiments over

the D 3 T HARdataset which includes labels of BAs and PTs, that they outperformpre-

viously explored systems which exclude these transitions (e.g. L-HAR). Moreover,
results depict their differences and show their usability is application-dependent. We
also propose the detection of unknown activities (UAs) that the system is not able