Energy Efﬁcient On-Sensor Processing for Online Activity Recognition

Florian Gr

utzmacher

, Albert Hein

, Benjamin Beichler

, Polichronis Lepidis

, Rainer Dorsch

Thomas Kirste

and Christian Haubelt

University of Rostock, 18051 Rostock, Germany

Bosch Sensortec GmbH, Gerhard-Kindler-Straße 9, 72770 Reutlingen, Germany

Keywords:

Energy Efﬁciency, Wearable Sensors, Activity Recognition, Energy Trade-Off, On-Sensor Computations.

Abstract:

In sensor-based online activity recognition, the communication of sensor samples at high data rates has a

great impact on the energy consumptions of wearables. In our work we investigate the idea of calculating

data reducing stages of activity recognition systems on wireless sensor nodes in order to reduce the amount

of transmitted data and thus the overall energy consumption. In our experiments, this approach could reduce

the energy consumption of a wireless sensor node by up to 27%. Since the beneﬁt of this approach highly

depends on design parameters of the activity recognition, we introduce an energy trade-off model for wireless

sensor nodes to estimate energy-savings of application speciﬁc conﬁgurations at design time. By calibrating

this model for our wireless sensor node, we could achieve an accuracy of more than 99% in our experiments.

1 INTRODUCTION

Online Activity Recognition with wearables has

gained increasingly importance, especially in the do-

mains healthcare, e-health, and ambient assisted liv-

ing (AAL). Usually, activity recognition is done in

multiple consecutive stages, which are known as the

Activity Recognition Chain (ARC) like introduced by

Bulling et al. in (Bulling et al., 2014). This chain

is composed of the stages data acquisition, prepro-

cessing, segmentation, feature extraction, and ﬁnally,

classiﬁcation. For a more detailed description we re-

fer to (Bulling et al., 2014).

Common setups involve wearable sensors from

which raw data is sent via wireless channels to per-

form online activity recognition. When using wear-

able sensors, its energy consumption plays a crucial

role, as it directly inﬂuences the long term usability of

the system. Thus, it is of great importance to keep the

energy consumption as small as possible, to prolong

the lifetime of the batteries. A considerable amount

of energy is used for the wireless transmission of raw

sensor data from the wearable sensors (Gr

utzmacher

et al., 2017; Rault et al., 2017). In order to reduce

the overall energy consumption, the amount of wire-

lessly transmitted sensor data should be kept as low as

possible. In order to achieve application speciﬁc data

reduction the feature extraction stage of most ARCs

could be utilized. By extracting the features on the

wireless sensor node, the amount of communicated

data can be reduced.

While some researchers proposed calculating fea-

tures on wireless sensor nodes for energy efﬁciency

reasons (Van Laerhoven and Aronsen, 2007; Laer-

hoven et al., 2006; Lorincz et al., 2009), the actual

energy consumption highly depends on the design pa-

rameters of the activity recognition and does not allow

for generalized statements. There exists a trade-off

between the added computational effort on the sen-

sor side and the reduced wireless transmissions. This

trade-off needs to be estimated at design time in order

to substantiate early design decisions. As a solution,

this paper introduces an energy trade-off model con-

sidering the important design parameters of the activ-

ity recognition system. Our contributions are:

• examining important design parameters of fea-

ture extraction inﬂuencing the energy consump-

tion when performed on wireless sensor nodes,

• an energy trade-off model for application speciﬁc

feature extraction conﬁgurations,

• an evaluation of our model by four feature set im-

plementations on a wireless sensor node, and

• a delay analysis of our implementations.

The remainder of the paper is structured as fol-

lows: In Section 2 the related work is discussed,

which is followed by a concept description in Sec-

tion 3. In Section 4 our experimental setup is de-

Grützmacher, F., Hein, A., Beichler, B., Lepidis, P., Dorsch, R., Kirste, T. and Haubelt, C.

Energy Efﬁcient On-Sensor Processing for Online Activity Recognition.

DOI: 10.5220/0006860100850092

In Proceedings of the 8th International Joint Conference on Pervasive and Embedded Computing and Communication Systems (PECCS 2018), pages 85-92

ISBN: 978-989-758-322-3

scribed, followed by the discussion of our results in

Section 5. In Section 6 conclusions are drawn.

2 RELATED WORK

As summarized in (Rault et al., 2017), the energy-

efﬁcient design of wearable sensor devices has been

subject of research in the wireless sensor networks,

activity recognition, healthcare, and AAL domains

for many years. One general approach to reduce

power consumption is the reduction of communi-

cation overhead using efﬁcient protocols (Ye et al.,

2002) or sensor network topologies (Younis and

Fahmy, 2004).

Several software architectures for activity recog-

nition systems have been studied to improve power

consumption: disabling hardware sensors using pre-

diction of activities has been proposed in (Gordon

et al., 2012) and (Wang et al., 2009). Saving com-

putational effort at a conceptual level can be done us-

ing feature selection (Yan et al., 2012), classiﬁer se-

lection (Liang et al., 2014), or ﬁxed-point arithmetic

(Anguita et al., 2013). These approaches are orthog-

onal to the approach investigated in the paper at hand

and can be combined for additional energy savings.

The principal of communication reduction by

performing on-board computations of wireless sen-

sor nodes has been introduced by several re-

searchers. Their works either focus on ofﬂine scenar-

ios (Van Laerhoven and Aronsen, 2007; Laerhoven

et al., 2006), sensor selection strategies (Lorincz et al.,

2009; Van Laerhoven and Aronsen, 2007; Laerhoven

et al., 2006; Zappi et al., 2008), or signal compres-

sion and abstraction approaches (Berlin and Van Laer-

hoven, 2012; Berlin and Van Laerhoven, 2010; Mam-

aghanian et al., 2011; Marcelloni and Vecchio, 2008).

However, except (Mamaghanian et al., 2011)

which focuses on sensor data compression, none of

the aforementioned works provide a quantiﬁed reduc-

tion of energy consumption of their approaches as no

explicit comparison to reference scenarios transmit-

ting raw data is done.

Rault et al. already indicated that a trade-off be-

tween communication reduction and added compu-

tational load has to be considered when performing

sensor-side computation (Rault et al., 2017). While

they propose to perform ”light” computations on the

sensor to keep the energy consumption for the addi-

tional computations as low as possible, we can show

that even calculating Fast Fourier Transform (FFT)

based features on the wireless nodes sensor subsys-

tems can lead to an overall energy reduction.

In (Gr

utzmacher et al., 2017) we have shown for

one conﬁguration, i.e. a window-based zero-crossing

rate on accelerometer data, that calculating it on the

wireless sensor node can reduce the energy consump-

tion by a considerable amount. Besides the evaluation

of more sophisticated feature sets, the paper at hand

studies the important design parameters of different

conﬁgurations inﬂuencing the energy consumption of

wireless sensor nodes.

3 CONCEPT

In online activity recognition with wireless sensor

nodes, the prevalent setup involves sensors sampled

at high sampling rates, e.g. 100Hz, which are trans-

mitted wirelessly to a host system for further process-

ing. As modern sensors already perform signal cor-

rection as a preprocessing step, the wireless commu-

nication usually takes place between the preprocess-

ing and segmentation stages of an ARC (see Figure 1).

Especially in settings which require energy efﬁcient

designs like smartphones, wireless sensor nodes or

wearables, highly frequent wireless communication is

not desired, when targeting a low energy consump-

tion.

Examining the ARC, data reducing stages can be

identiﬁed. The feature extraction stage for example is

mainly implemented as a sliding window based fea-

ture extraction in most of the literature (Huynh and

Schiele, 2005; Kr

uger et al., 2014; Atallah et al.,

2011; Capela et al., 2015). It often drastically re-

duces the data rate, as a relatively small number of

features is calculated from a high number of samples

of a window. Thus, the calculation of features, which

is included in an ARC anyway, could be shifted to

the wireless sensor node. Figure 1 shows this princi-

pal. While this approach can reduce a considerable

amount of energy, it highly depends on the design

parameters of the activity recognition (e.g. a sliding

window with 50% overlap halfs the data rate, com-

pared to a sliding window with 75% overlap). On the

other hand, the added computational effort has to be

taken into account as well, as it increases the energy

consumption of the processing units. The trade-off

between reduced data rate and added workload has to

be considered in order to estimate the resulting energy

savings.

3.1 Estimation of Energy Savings

As different features are useful to detect different

kinds of activities, the number and composition of

features varies amongst different applications.

PEC 2018 - International Conference on Pervasive and Embedded Computing

Data

Acquisition

Preprocessing Segmentation

Feature

Extraction

...

less frequent

transmissions

Figure 1: Activity recognition chain with on-sensor calculations up to and including the feature extraction stage.

Two aspects are decisive for the resulting energy

savings: the workload of the microcontroller and the

rate with which data is communicated over the wire-

less interface. We introduce an estimation method,

which is based on a combination of two energy mod-

els of the target architecture: a model which captures

the energy consumption of the communication, and

a model which captures the additional computational

load on the sensor device. As we are interested in the

change of energy consumption when varying the data

rate or the computational effort, both models are not

required to capture the absolute energy consumption,

but rather the relative average energy consumption to

both parameters. Thus, the resulting model is referred

to as energy trade-off model. It can be used to calcu-

late the difference in energy consumption of two con-

ﬁgurations.

As this model depends on two parameters, it be-

comes a surface in a three dimensional space. The

structure of the surface depends on the relation be-

tween the energy consumption and the two model pa-

rameters. Note, that the model parameters need to be

distinguished from the design parameters of the fea-

ture extraction. The model parameter capturing the

amount of communicated data can be calculated from

the design parameters of the feature extraction.

Since data is acquired from sensors, the amount

of data can be naturally captured by the output fre-

quency, of either raw samples or feature values.

Therefore, it is intuitive to describe the ﬁrst model pa-

rameter in terms of a data transmission frequency f

Thus, the energy dependency on the amount of com-

municated data is described by function e

( f

). Note,

that the size of the data types transmitted with f

must

equal, when comparing two conﬁgurations.

The computational effort on the other hand can be

described in terms of processing load l

in %, deﬁned

by the processing time t

in a ﬁxed time interval T :

(1)

Since the calculation of feature extraction usually is

done in a sample-based or sliding-window based fash-

ion, which both is triggered with the sampling fre-

quency, it is obvious to use the sampling period as the

time interval T . This simpliﬁes the processing load

to the average processing time per sample t

. The

resulting energy dependency on the processing load

is described as function e

). Note, that the sam-

pling frequency describes at which rate sensor sam-

ples are acquired. This needs to be distinguished from

the data transmission frequency, which captures the

rate at which either raw samples or feature values are

communicated from the wireless sensor node. As we

will see in Section 4, both model parameters allow a

fast calibration, if e

( f

) and e

) are unknown for

a particular sensor node.

The resulting equation for calculating the differ-

ence of the energy consumption ∆e is:

∆e = e

( f

) − e

( f

) + e

) − e

), (2)

with f

and f

being the output frequencies and

and t

being the average computation times per

sample of the reference and the test conﬁguration re-

spectively. Note, that this model assumes that com-

putational load and the amount of communicated data

inﬂuence the energy consumption in a additive way,

and both components do not inﬂuence each other

As shown in Section 4, the energy consumption of

our wireless sensor node on both model parameters

shows a linear behavior.

In that case, the resulting model becomes a plane

in a 3D space with the slopes m

for e

( f

) and m

for e

). Thus, the equation for ∆e resolves to:

∆e = ( f

− f

) ∗ m

+ (t

−t

) ∗ m

. (3)

In the following we will explain, how both model pa-

rameters can be acquired from the design parameters

of a certain set of features.

The output data rate f

of the feature values

mainly depends on the number of features N

, and

the sliding window rate f

. The latter is calculated

from the sensor sampling frequency f

in Hz and the

sliding window parameters window size S

in Num-

ber of Samples and window overlap O

as a fraction

on the interval (0,1), by:

(1 − O

) ∗ S

. (4)

The parameter f

can be calculated by:

= f

∗ N

, (5)

for a conﬁguration calculating the feature extraction

on the wireless sensor node or directly by using f

in a conﬁguration of transmitting raw sensor sam-

ples. As we can see from Equations (3), (4), and (5),

Energy Efﬁcient On-Sensor Processing for Online Activity Recognition

Wireless Sensor Node

BHI160 Sensorhub

DA14583

µC

I2C

INT

I2C

FUSER

96K ROM

48K RAM

ACC+GYR

BMI160

BTLE

Figure 2: Architecture of our custom wireless sensor node.

the resulting absolute data rate reduction depends on

the sampling frequency. Thus, the impact of energy

savings by calculating feature extraction on the wire-

less sensor node increases with the sampling rate of

the sensor.

For determining the workload dependency e

() of

the target conﬁguration, different approaches are pos-

sible. A simple but fast way is the use of energy mod-

els which depend on the execution time of the feature

extraction code. Such models can be acquired by ex-

periments as we have done for our sensor node like

shown in Section 4.1. The workload can be calculated

by Equation (1), which only requires an estimation of

the average processing time per sample t

of the fea-

ture extraction code.

The processing time t

can be acquired by do-

ing worst-case execution time analysis on the assem-

bler code compiled for the target architecture. Fea-

ture extractions are usually performed using data ﬂow

oriented algorithms with few branches. Furthermore,

different execution paths usually have neglectable dif-

ferences in their path lengths. Thus, the instruction

count of the assembler code together with the micro-

controller frequency and the cycles per instruction ac-

quired from their data sheets can be a fast way of es-

timating the execution time. This should be possible

for most wireless sensor nodes, as mainly microcon-

troller architectures are chosen, because of their low

power consumption. Typically microcontroller archi-

tectures neither have deep pipelines nor operating sys-

tems performing dynamic scheduling. Thus, the exe-

cution time of the assembler code is not subject to

essential variations.

However, there also exist more ﬁne grained ap-

proaches, which directly model the energy consump-

tion of assembler instructions for a piece of source

code (Ruberg et al., 2015; Bazzaz et al., 2013).

4 EXPERIMENTS

In our experiments we used a custom wireless sen-

sor node which was designed for low power applica-

tions. It is equipped with an ultra low power BHI160

sensor hub from Bosch Sensortec (Bosch Sensortec,

2017), a Dialog Semiconductor DA14583 microcon-

troller with Bluetooth Low Energy (BTLE) interface,

acting as a host controller which is connected with

the BHI160 via I

C and a CR1225 coin cell battery

as power supply. For an overview see Fig. 2. In our

setup, the sensor node either sends raw samples or cal-

culated features to a Samsung Galaxy S5 smartphone,

acting as a data aggregating device. All on-sensor fea-

ture extractions are performed on the BHI160’s inte-

grated FuserCore in our experiments.

When evaluating the energy consumption of the

system we followed a similar approach introduced by

(Russell and Jacome, 1998) since it meets the neces-

sary requirements for comparing average energy con-

sumptions and it is easy to deploy. For current mea-

surement we are using a 10 Ω shunt resistor R

in se-

ries with the power source. In our experiments we

substituted the coin cell by a constant voltage source

providing 3.0 V supply voltage. The voltage drop U

over R

is captured and averaged by a DSO-X 3034A

oscilloscope from Agilent Technologies over a time

of at least 100 s with a resolution of 10 MSamples/s.

Due to its proportional nature, the average energy

consumption can be deduced from the measured av-

erage voltage drop with sufﬁcient accuracy. The tem-

perature dependency was considered to be neglectable

in our setup, which could be substantiated with our

results for the model accuracy in Section 5.1. When

comparing the average voltage drop over time of dif-

ferent scenarios of the sensor, the systematic error of

the measuring apparatus and the tolerance of the shunt

resistor is canceled out and thus, propagation of un-

certainty can be neglected for our evaluations.

4.1 Model Calibration

In order to show how much energy could be saved

by reducing data communication, we measured the

energy consumption of our wireless sensor node pro-

viding accelerometer and gyroscope data at different

output data rates transmitted over BTLE. While vary-

ing the output data rates, all other factors like sensor

sampling rate, work load and energy mode are kept

constant for all measurements. The output data rate is

reduced by just skipping the corresponding number of

samples to be propagated to the sensor output. Fig. 3

shows the overall average energy consumption of the

sensor node at different output data rates as well as

the energy consumption without BTLE transmissions,

to show the fraction of energy consumed by wireless

transmissions.

It can be seen that the amount of data send via

BTLE can have a great impact on the nodes energy

consumption, especially at data rates like 100 Hz or

more, which are broadly used in the literature con-

PEC 2018 - International Conference on Pervasive and Embedded Computing

cerned with activity recognition.

In order to show the energy increase of the added

workload when calculating additional algorithms on

a sensor subsystem, we implemented synthetic pro-

cessing load by performing Multiply and Accumulate

(MAC) operations in a loop and measured their en-

ergy consumption as well as the average computation

time. We used MAC operations, since our feature im-

plementation including online versions of statistical

feature calculations as well as an FFT implementa-

tion, mainly consist of MAC operations. The pro-

cessing load is deﬁned as the fraction of time the pro-

cessing unit is actively performing the additional cal-

culations, as described by Equation (1). The exper-

iments were done in the operating point at transmit-

ting accelerometer and gyroscope values with 2.5 Hz

as shown in Fig. 3. The resulting dependency can

be seen in Fig. 4. The trade-off between reduced

data rate and additional workload gets obvious, as re-

ductions in energy consumption caused by decreas-

ing the output data rate from 60 Hz to 2.5 Hz can al-

ready be neutralized when the additional workload

for doing that is more than approximately 57%. This

shows, that the trade-off between reduced communi-

cation and additional processing load needs to be ex-

plicitly evaluated for different conﬁgurations.

To calibrate our model we used the experimen-

tal results of the aforementioned measurings. As the

energy dependency of both model parameters is lin-

ear, the model is described by Equation (3). Both

slopes, m

and m

can be determined by calculat-

ing a regression of measuring values acquired for the

overall energy consumption. These lines are col-

ored green in Figures 3 and 4. The slopes of the

regression lines (m

= 0.0654172 mV /Hz and m

0.067873mV /average workload %) are used as the

calibrated model parameters.

0

5

10

15

20

0 20 40 60 80 100

[mV]overShuntResistor

[Hz]

RawDatawithBTLETransmissions

FitofRawDatawithBTLETransmissions

RawDatawithoutBTLETransmissions

FitofRawDatawithoutBTLETransmissions

Figure 3: The sensor nodes average energy consumption at

different output data rates of both accelerometer and gyro-

scope with and without BTLE communication.

4.2 Feature Sets

We implemented a feature extraction conﬁguration on

the BHI160 sensor subsystem which showed good re-

sults for activity and intention recognition for meal

preparation and related activities in a kitchen sce-

nario (Kr

uger et al., 2014). This feature set consists

of mean, variance, skewness, kurtosis, peak (domi-

nant frequency), and energy (magnitude of the domi-

nant frequency) on accelerometer and gyroscope data.

Both signals are sampled at 100 Hz and segmented

by a sliding window of 128 samples with an overlap

of 75%. This leads to a 12-dimensional feature vec-

tor with a sliding window rate of 3.125 Hz and is re-

ferred to as MVSKPE. This setup includes frequency

domain features by calculating an 128-point FFT. Due

to its high computational effort, we also implemented

a more lightweight feature set, by substituting peak

and energy by a zero-crossing rate (ZCR), since it

can also give an idea about with which average fre-

quency an sensor is shaken. This setup is referred to

as MVSKZ. In some literature, only statistical time-

domain features are used with mean and variance at

least (see (Dama

sevi

cius et al., 2016)). Therefore,

we also tested a feature set of mean, variance, skew-

ness and kurtosis referred to as MVSK, and a set of

mean and variance only, referred to as MV. All feature

sets have been implemented on the BHI160’s Fuser-

Core on our wireless node and its BTLE connection

interval was chosen as the maximal possible interval

which could still provide the necessary corresponding

data rate. In Fig. 5, our setup is shown.

In order to keep the computational effort for cal-

culating features on the sensor as low as possible,

we implemented the online version of mean, vari-

ance, skewness, and kurtosis from (John D. Cook,

2018), which is based on the one pass algorithms

0

5

10

15

20

0 10 20 30 40 50 60 70

[mV]overShuntResistor

AdditionalProcessingLoadl

[%]

RawDatawithBTLETransmissions

FitofRawDatawithBTLETransmissions

RawDatawithoutBTLETransmissions

FitofRawDatawithoutBTLETransmissions

Figure 4: Sensor nodes average energy consumption at 2.5

Hz output data rate and different additional workloads with

and without BTLE communication.

Energy Efﬁcient On-Sensor Processing for Online Activity Recognition

Wireless Sensor Node

BHI160 Sensor Hub

Fuser

Core

µC

less frequent

wake-ups

less frequent

transmissions

ACC

Figure 5: Shifting the feature extraction to the sensor hub reduces the wake-up frequency of the host and wireless transmis-

sions.

for arbitrary-order statistical moments from (Pebay,

2008). This enables those features to be updated for

each new sample with a computational complexity

O(1). The implemented ZCR is a comparison of a

new sample with its preceding sample to detect a zero

crossing accumulated over the time of a sliding win-

dow. The implemented FFT is a 128-Point Radix-

2 implementation with 7 butterﬂy stages. All algo-

rithms use single precision ﬂoating point arithmetics.

5 RESULTS

The energy consumptions of MV, MVSK, MVSKZ,

and MVSKPE as well as the reference scenario of

sending raw accelerometer and gyroscope data at

100 Hz to a smartphone were measured. The mea-

sured improvement in energy consumption of each

feature set compared to the reference scenario is

shown in column ”Meas.” in Table 1. It can be seen

that by implementing our chosen feature sets on the

sensor node, the energy consumption is reduced by

20.53% to 26.8% compared to our reference scenario,

depending on the feature set.

Note, that the relative energy savings depend on

the overall energy consumption, including the en-

ergy consumption independent of the sensors work-

load and the BTLE transmission. Reducing this part,

e.g. by utilizing sleep modes of the host controller,

has a potential for considerable further reductions of

the overall energy consumption and thus a higher im-

pact of the studied approach.

From the model parameters of the implemented

feature sets in Table 2, it can bee seen that the data

reduction of our implementations range from 81%

to 93%, depending on the chosen feature conﬁgura-

tion. This results can cope with state of the art data

compression algorithms for wireless sensor nodes,

without neither introducing additional processing ef-

Table 1: Predicted and Measured Average Energy Savings.

Features Pred. [%] Meas. [%] Pred. Error [%]

MVSKPE 20.60 20.53 0.5%

MVSKZ 23.68 23.70 0.1%

MVSK 24.63 24.61 0.1%

MV 27.02 26.87 0.6%

fort, nor loosing information. Note, that by shifting

the feature extraction to the sensor subsystem of our

node, less computations on the smartphone have to

be performed as well, which should reduce its en-

ergy consumption as well. However, this is not further

evaluated in this work.

5.1 Model Accuracy

In order to evaluate the accuracy of our energy trade-

off model, we acquired the average computation time

of the reference scenario and the feature sets calcu-

lated on the BHI160’s FuserCore. We measured the

average computation time instead of using one of the

prediction methods from Section 3.1, in order to only

evaluate the model accuracy without additional pre-

diction uncertainties, as they may vary among differ-

ent methods. The results are shown in row l

of Ta-

ble 2. It can be seen that by calculating the ZCR in-

stead of the FFT features, the average workload can

be reduced from 10% to 4%.

With the differences of the average workload l

(Equation (1)) and the output frequency f

(Equation

(5)) between the feature set to be tested and the ref-

erence scenario, the savings in energy consumption

were estimated by our model (Equation (3)). The

model parameters used for the evaluation are shown

in Table 2. We compared the estimated savings with

the measured results for the selected feature sets and

calculated the error, which can be seen in Table 1. The

prediction error of our model is less than 1%, which is

small enough to make feasible estimations of energy

savings when considering feature calculations on the

sensor node at design time. Note, that the uncertainty

of the methods for predicting the additional average

workload from an existing implementation (Section

3.1) will further affect the overall prediction error cor-

respondingly.

Table 2: Model parameters for the selected features sets.

Features avg. workl. l

[%] output rate f

[Hz]

Raw 0.3 100

MVSKPE 10.4 18.75

MVSKZ 3.3 15.625

MVSK 3.2 12.5

MV 1.3 6.25

PEC 2018 - International Conference on Pervasive and Embedded Computing

Table 3: Actual and achieved processing delays.

Features proc. delay [ms] delay in our setup [ms]

MVSKPE 22.8 120

MVSKZ 0.34 110

MVSK 0.35 100

MV 0.14 80

5.2 Delay Analysis

As already stated in (Rault et al., 2017), the sensor-

side computation of features can inﬂuence the latency

of the activity recognition system. Therefore, we ana-

lyzed the processing delay for the features calculated

on the sensor. With timestamp functions we measured

the additional processing time of the feature extrac-

tion implementations.

We calculated the average processing time of at

least 200 measurements per feature set. These results

can be seen in Table 3 as ”proc. delays”. With calcu-

lating the feature extraction on the sensor itself, multi-

ple values (feature vector entries) per sliding window

are propagated to the sensor output in one batch, in a

very short time. As in our setup a microcontroller is

acquiring the sensor data for transmitting via BTLE,

which has a limited input data rate due to it’s ﬁrmware

implementation, we had to delay the output of fea-

tures on the sensor subsystem artiﬁcially. When the

features are calculated for a sliding window, the out-

put of each feature vector entry is delayed until the

next sensor value is sampled, which is 10 ms in our

case. Since the feature vector for scenario MVSKPE

has 12 entries which is the greatest feature vector

among our feature sets, the maximum delay in our ex-

periments is 120 ms. This and the delay for the other

feature sets can be seen in Table 3 as ”delay in our

setup”. Note, that the artiﬁcial delay in our setup was

necessary because of our ﬁrmware implementation on

the sensor node. With a BHI160 evaluation board, we

veriﬁed that this is not a general issue, but caused by

our particular setup. Only the feature entries for the

peak frequency and its magnitude have to be delayed

by approx. 30 ms, since the FFT calculation for all 6

axis takes longer than the 10 ms sampling period and

has to be distributed among 3 sampling periods. How-

ever, even 120 ms is a fairly good result for most ac-

tivity recognition tasks. Furthermore, the calculation

of features on multiple wireless sensor nodes is inher-

ently done in parallel, which might not be the case

when calculating the feature extraction for all sensors

on a data aggregating device, depending on its parallel

processing capabilities.

6 CONCLUSION

Our work investigates the idea of calculating the

feature extraction stage of activity recognition algo-

rithms on wireless sensor nodes. We have shown the

outcomes of this approach by porting four broadly

used feature sets to a BHI160 sensor subsystem resid-

ing on a wireless sensor node and compared the en-

ergy savings to reference scenarios transmitting raw

sensor samples instead. In our experiments, this ap-

proach reduced the energy consumption of the wire-

less sensor node by up to 27%.

To show the applicability of on-sensor feature ex-

tractions, we evaluated the processing delay of our

implementations, which was 120 ms at most.

More importantly, in our work we examined the

decisive design parameters for the resulting energy

trade-off when calculating data reducing stages like

feature extraction on wireless sensor nodes. From

these design parameters an energy trade-off model is

built which allows to estimate the energy savings of

sensor side feature calculations at design time. This

improves the development process of activity recogni-

tion systems as design decisions can be substantiated

early in the design process. By calculating the energy

savings from the model ﬁtted to our wireless sensor

node and comparing it to four implemented feature

sets in our experiments, we achieved a model accu-

racy of more than 99%.

ACKNOWLEDGMENTS

This work is partially supported by the German Fed-

eral Ministry of Education and Research (BMBF),

grant number 03ZZ0519D.

REFERENCES

Anguita, D., Ghio, A., Oneto, L., Parra, X., and Reyes-

Ortiz, J. L. (2013). Energy efﬁcient smartphone-based

activity recognition using ﬁxed-point arithmetic. J.

UCS, 19(9):1295–1314.

Atallah, L., Lo, B., King, R., and Yang, G.-Z. (2011). Sen-

sor positioning for activity recognition using wearable

accelerometers. IEEE transactions on biomedical cir-

cuits and systems, 5(4):320–329.

Bazzaz, M., Salehi, M., and Ejlali, A. (2013). An accurate

instruction-level energy estimation model and tool for

embedded systems. IEEE Transactions on Instrumen-

tation and Measurement, 62(7):1927–1934.

Berlin, E. and Van Laerhoven, K. (2010). An on-line piece-

wise linear approximation technique for wireless sen-

sor networks. In Proc. of LCN’10, pages 905–912.

IEEE.

Energy Efﬁcient On-Sensor Processing for Online Activity Recognition

Berlin, E. and Van Laerhoven, K. (2012). Detecting leisure

activities with dense motif discovery. In Proc. of Ubi-

Comp’12, pages 250–259. ACM.

Bosch Sensortec (2017). BHI160 / BHI160B - Ultra

low-power sensor hub incl. integrated imu. BST-

BHI160(B)-DS000-01.

Bulling, A., Blanke, U., and Schiele, B. (2014). A tutorial

on human activity recognition using body-worn iner-

tial sensors. ACM Comput. Surv., 46(3):33:1–33:33.

Capela, N. A., Lemaire, E. D., and Baddour, N. (2015).

Feature selection for wearable smartphone-based hu-

man activity recognition with able bodied, elderly, and

stroke patients. PloS one, 10(4):e0124414.

Dama

sevi

cius, R., Vasiljevas, M.,

Salkevi

cius, J., and

zniak, M. (2016). Human activity recognition in

aal environments using random projections. Compu-

tational and mathematical methods in medicine, 2016.

Gordon, D., Czerny, J., Miyaki, T., and Beigl, M. (2012).

Energy-efﬁcient activity recognition using prediction.

In Proc. of ISWC ’12, pages 29–36.

utzmacher, F., Wolff, J.-P., Hein, A., Lepidis, P., Dorsch,

R., Kirste, T., and Haubelt, C. (2017). Towards energy

efﬁcient sensor nodes for online activity recognition.

In Proc. of IECON’17, pages 8291–8296. IEEE.

Huynh, T. and Schiele, B. (2005). Analyzing features for

activity recognition. In Proc. of sOc-EUSAI ’05, pages

159–163, New York, NY, USA. ACM.

John D. Cook (visited 03/2018). Comput-

ing skewness and kurtosis in one pass.

https://www.johndcook.com/blog/skewness kurtosis/.

uger, F., Nyolt, M., Yordanova, K., Hein, A., and Kirste,

T. (2014). Computational state space models for activ-

ity and intention recognition. a feasibility study. PloS

one, 9(11):e109381.

Laerhoven, K. V., Gellersen, H.-W., and Malliaris, Y. G.

(2006). Long term activity monitoring with a wearable

sensor node. In Proc. of BSN’06., pages 4–pp. IEEE.

Liang, Y., Zhou, X., Yu, Z., and Guo, B. (2014). Energy-

efﬁcient motion related activity recognition on mobile

devices for pervasive healthcare. Mobile Networks

and Applications, 19(3):303–317.

Lorincz, K., Chen, B.-r., Challen, G. W., Chowdhury, A. R.,

Patel, S., Bonato, P., Welsh, M., et al. (2009). Mer-

cury: a wearable sensor network platform for high-

ﬁdelity motion analysis. In SenSys, volume 9, pages

183–196.

Mamaghanian, H., Khaled, N., Atienza, D., and Van-

dergheynst, P. (2011). Compressed sensing for real-

time energy-efﬁcient ecg compression on wireless

body sensor nodes. IEEE Transactions on Biomedi-

cal Engineering, 58(9):2456–2466.

Marcelloni, F. and Vecchio, M. (2008). A simple algo-

rithm for data compression in wireless sensor net-

works. IEEE communications letters, 12(6).

Pebay, P. P. (2008). Formulas for robust, one-pass parallel

computation of covariances and arbitrary-order statis-

tical moments. Technical report, Sandia National Lab-

oratories.

Rault, T., Bouabdallah, A., Challal, Y., and Marin, F.

(2017). A survey of energy-efﬁcient context recogni-

tion systems using wearable sensors for healthcare ap-

plications. Pervasive and Mobile Computing, 37:23–

44.

Ruberg, P., Lass, K., and Ellervee, P. (2015). Microcon-

troller energy consumption estimation based on soft-

ware analysis for embedded systems. In Proc. of

NORCAS’15, pages 1–4. IEEE.

Russell, J. T. and Jacome, M. F. (1998). Software power

estimation and optimization for high performance, 32-

bit embedded processors. In Proc. of ICCD’98, pages

328–333. IEEE.

Van Laerhoven, K. and Aronsen, A. K. (2007). Memoriz-

ing what you did last week: Towards detailed actigra-

phy with a wearable sensor. In Proc. of ICDCSW’07.,

pages 47–47. IEEE.

Wang, Y., Lin, J., Annavaram, M., Jacobson, Q. A., Hong,

J., Krishnamachari, B., and Sadeh, N. (2009). A

framework of energy efﬁcient mobile sensing for au-

tomatic user state recognition. In Proc. of Mobisys

’09, pages 179–192. ACM.

Yan, Z., Subbaraju, V., Chakraborty, D., Misra, A., and

Aberer, K. (2012). Energy-efﬁcient continuous ac-

tivity recognition on mobile phones: An activity-

adaptive approach. In Proc. of ISWC ’12, pages 17–

24. IEEE.

Ye, W., Heidemann, J., and Estrin, D. (2002). An energy-

efﬁcient mac protocol for wireless sensor networks. In

Proc. of INFOCOM ’02, volume 3, pages 1567–1576.

IEEE.

Younis, O. and Fahmy, S. (2004). Heed: a hybrid, energy-

efﬁcient, distributed clustering approach for ad hoc

sensor networks. IEEE Transactions on mobile com-

puting, 3(4):366–379.

Zappi, P., Lombriser, C., Stiefmeier, T., Farella, E., Roggen,

D., Benini, L., and Tr

oster, G. (2008). Activity recog-

nition from on-body sensors: accuracy-power trade-

off by dynamic sensor selection. In Wireless sensor

networks, pages 17–33. Springer.

PEC 2018 - International Conference on Pervasive and Embedded Computing