![](bg2.png)
as, i.e. either accelerometers, EMG sensors or elec-
trogoniometers. However, the combination of sensors
and thus the fusion of biosignals may improve the ro-
bustness of the system or the accuracy of recognition.
(Rebelo et al., 2013) studied the classification of iso-
lated human activities based on all of the three above
described sensors (acceleration, EMG, goniometer)
attached to the knee. They successfully recognized
seven types of activities, i.e. sit, stand, sit down,
stand up, walk, ascend and descend, with an accu-
racy of about 98% in person dependent recognition.
While these results are very encouraging, it still re-
mains challenging to robustly recognize a large va-
riety of human everyday activities in the real world.
The first step to achieve this goal is to build a frame-
work for recording, annotating and processing biosig-
nals and applying the processed biosignal data to ac-
tivity recognition and strain prediction. Such a frame-
work includes:
• a high-quality wireless biosignal data recording
device with appropriate sensors,
• a mobile data acquisition and archiving system,
• an alignment method to semi-automatically anno-
tate the biosignal data,
• a real-time system to robustly recognize everyday
activities,
• a strain prediction system based on recognized ac-
tivities and biomechanical as well as medical ex-
pert knowledge.
In this paper we focus on a proof-of-concept vali-
dation of everyday activity recognition. For this pur-
pose we collected a pilot dataset consisting of biosig-
nals captured by two accelerometers, four EMG sen-
sors and one electrogoniometer. Subsequently, the
dataset was used to evaluate the recognition system.
To model human activities we followed the ap-
proach as described in (Rebelo et al., 2013) and used
Hidden-Markov-Models (HMM). HMMs are widely
used to a wide range of activity recognition tasks,
such as in (Lukowicz et al., 2004) and (Amma et al.,
2010). In the latter, the authors present a wearable
system that enables 3D handwriting recognition based
on HMMs. In this so-called Airwriting system the
users write text in the air as if they were using an
imaginary blackboard, while the handwriting gestures
are captured wirelessly by accelerometers and gyro-
scopes attached to the back of the hand (Amma et al.,
2010).
In this paper we focus on a seven basic daily ac-
tivities to validate our framework, i.e. the activities
”sit”, ”stand”, ”sit-to-stand”, ”stand-to-sit”, ”walk”,
”curve-left” and ”curve-right”. Five of these activi-
ties correspond to those described in (Rebelo et al.,
2013), while the remaining two activities ”ascend”
and ”descend” in (Rebelo et al., 2013) were replaced
by ”curve-left” and ”curve-right”. Thus, the total
number of seven activities is the same in both stud-
ies.
2 THE ACTIVITY SIGNALS KIT
(ASK)
The goal of our pilot study is to validate the end-to-
end activity recognition system of the ASK frame-
work. We first selected a suitable device and sen-
sors to continuously capture the activity data. Subse-
quently the are automatically archived and processed
for the recognition and validation steps.
2.1 Equipment and Setup
2.1.1 Device
We chose the biosignalsplux Research Kits
1
as
recording device. One PLUX hub
2
records signals
from 8 channels (each up to 16 bits) simultaneously.
We used two hubs for recording data and connected
the hubs with a synchronization cable which syn-
chronizes signals between the hubs at the beginning
of each recording. This procedure ensures the time-
alignment of sensor data during the entire recording.
2.1.2 Biosignals and Sensors
Similar to (Mathie et al., 2003), we used two triaxial
accelerometers
3
, four bipolar EMG sensors
4
(instead
of six) and both channels of one biaxial electrogo-
niometer
5
together. Instead of using only one channel
of the electrogoniometer as in (Mathie et al., 2003),
we used both channels to measure both the frontal and
sagital plain since we need to recognize rotational ac-
tivities like ”curve-left” and ”curve-right”.
One channel on the hub1 was plugged with a
pushbutton
6
. It is used for the semi-automatic anno-
tation mechanism (See Section 2.3.1). The signals of
all sensors were recorded wirelessly at different sam-
pling rate. Table 1 shows the sampling rate of each
sensor.
1
biosignalsplux.com/researcher
2
store.plux.info/components/263-8-channel-hub-820201701
.html
3
biosignalsplux.com/acc-accelerometer
4
biosignalsplux.com/emg-electromyography
5
biosignalsplux.com/ang-goniometer
6
biosignalsplux.com/pushbutton
ASK: A Framework for Data Acquisition and Activity Recognition
263