Hardware-Friendly Activity Recognition with Fixed-Point Arithmetic - Smartphone-Based Human Activity Recognition - page 84

their movement similarities. Static activities on the other hand performed better,
such as laying , for which we reported 0% classification error. Furthermore, a small
misclassification overlap was found between standing and sitting , which is attributed
to the waist-mounted smartphone physical location and the difficulty to discriminate
between them: this is mainly due to the slight inclination difference of the phone with
respect to the vertical axis when these activities are performed and largely depends
on the user's body type.
These classification errors could be in some way improved by incorporating new
types of features or sensors into the HAR system. For example, the inclusion of
gyroscopes as we will see in the following Chap. 6 or the incorporation of additional
accelerometers in different body parts.

5.4.2 Processing Time and Battery Consumption

Various tests were performed on the smartphone to determine the advantages of using
this novel hardware-friendly approach in terms of recognition speed and battery
consumption. We expected that avoiding the use of the floating-point arithmetic
for complex calculations could lead to energy sparing on stand-alone devices. For
these trials, we used a SGSII smartphone equipped with a Li-Ion 1650 mAh battery
with up to 610 h of stand-by operation and Android OS as the operating system
(Gingerbread version 2.3.4).
Aphone appwas developed for this purpose: measure prediction speed and battery
consumption. All the expensive operations were written in C using the Android NDK
(signal processing and Machine Learning algorithm). Before trials, we turned off all
phone services (e.g. Wi-Fi, Bluetooth and 3 G Network) and, more importantly,
the phone screen. This latter is in general the most energy consuming part in a
smartphone. We aimed to isolate this measurement process as much as possible to
obtain a realistic approximation of the variables under evaluation.
A simulation of the HAR process was implemented on the smartphone with
the possibility of adjusting the number representation. This was achieved for either
fixed-point or floating-point arithmetic using the default data types available in the C
language from 8 to 64 bits. They were selected because the available libraries have
only power of two number representations.
On the first part of the measurements, we decided to continuously make activity
predictions over a fixed period of time (5 min) and obtain an estimate of the average
prediction rate (in number of predictions

s). The time of the activity recognition
process was measured starting from the sensor reading to determining the SVM
FFP output was measured for each approach. Table 5.4 shows the obtained results.
It is worth highlighting the large difference between the rates using the fixed-point
representation instead of the floating-point and also the proportional relationship
between the number of bits used and the processing time. For instance, the 32-bits
integer model outperforms in speed the 32-bit float model by almost 7 times.

/