Energy-efﬁcient Operation of GSM-connected Infrared Rodent Sensor

abor Paller and G

abor

echenyi Istv

an University, Information Society Research & Education Group, Egyetem t

er 1., Gy

or, Hungary

Keywords:

Agriculture, Infrared Camera, Power Efﬁciency.

Abstract:

Camera sensors have been deployed in the agriculture for various use cases. Most of the applications tried to

infer the health and development of the plants based on image data in different wavelength domains. In this

paper, we present our research of rodent population estimation with infrared camera sensors.

The usual camera sensor applications in the agricultural domain are quite simple from the sensor architecture

point of view as the environment rarely changes. Image capture/transmission at preconﬁgured moments is

usually enough. Rodents move quickly so the sensor must be able to capture images with low capture time

interval. As the data link to the server backend is relatively slow, this fast capture rate may require image

processing capability in the sensor.

The paper analyzes the effects of such an image processing capability, in particular the power consumption

trade-offs. Inadequate power management support of the selected embedded Linux platforms is identiﬁed as

a problem and proposals are made for improvement.

1 INTRODUCTION

Data captured by agricultural sensors may be scalar

(Adamchuk et al., 2004), (L

opez et al., 2009) (Arm-

strong et al., 1993) or of more complex data types, e.g.

spectrogram. Among the latter, 2D imaging sensors

(”cameras”) have been found to be efﬁcient in detect-

ing effects of drought (Grant et al., 2006), (Alderfasi

and Nielsen, 2001), plant phenotype (Li et al., 2014)

or diseases (Moshou et al., 2004).

AgroDat.hu is an ongoing research project funded

by the Government of Hungary (Paller et al., 2015).

In the ﬁrst phase the project concentrated on sen-

sors providing scalar value like soil temperature, soil

moisture, concentration of salts in groundwater and

concentration in the ground, air temperature, hu-

midity, rainfall, wind speed and direction, solar radi-

ation intensity and leaf wetness. The project decided

to use GSM/GPRS network as its data carrier.

Due to the fact that one type of our sensor

poles has only underground parts, solar cells as en-

ergy source could not be employed hence energy-

efﬁciency was an important concern from the begin-

ning. These decisions made in the ﬁrst phase when the

project deployed only sensors providing scalar time-

series data inﬂuenced the second phase when cameras

are added. In particular, we wanted to reuse the phys-

ical sensor pole structure and the GSM/GPRS data

carrier layer. This led us into researching energy ef-

ﬁciency in case of the data capture and transmission

patterns of 2D image structures.

2 COMMON VOLE DETECTION

Our intention was to create a ﬂexible framework for

camera sensors therefore we looked for a more de-

manding use case. Simple use cases of agricultural

camera sensors normally involve taking a picture 1-4

times a day about a speciﬁc feature of the environ-

ment, e.g. the selected part of the foliage. From the

communication pattern and sensor framework points

of view, these use cases are just slightly different from

the scalar value use case. The sole difference is the

larger data size but as these larger data packets are in-

frequently sent, the conclusions are not signiﬁcantly

different from the ones presented in (Paller et al.,

2015).

One of the more challenging use cases we iden-

tiﬁed is animal monitoring, speciﬁcally rodent track-

ing. Population outbreaks of certain rodent species

can cause signiﬁcant damage in crop production.

Grain Producers Association-Hungary (GPAH) esti-

mated that common vole (Microtus arvalis, Figure 1)

caused 500000 tons of damage in winter wheat alone

in 2014. More agressive rodenticides are applied ac-

cording to population estimation hence this estima-

Paller, G. and Élõ, G.

Energy-efﬁcient Operation of GSM-connected Infrared Rodent Sensor.

DOI: 10.5220/0005628500370044

In Proceedings of the 5th International Confererence on Sensor Networks (SENSORNETS 2016), pages 37-44

ISBN: 978-989-758-169-4

Figure 1: Common vole (Microtus arvalis) (source: Na-

tional Forestry Association, Hungary).

tion is an economically important task.

Common vole population estimation was reported

using radio collars (Jacob and Hempel, 2003) but ob-

viously this method is not suitable to detect animals

that have not been trapped previously. Detection of

itinerant animals which just try to settle on a certain

ﬁeld is best accomplished with cameras but the noc-

turnal nature of the animal and the excellent camou-

ﬂage of their fur make them difﬁcult to spot in visible

light. (Ou-Yanga et al., 2011) reports about the ob-

servation of laboratory animals with infrared camera.

The monitored area was rather small - about 50x100

cm - and they employed Kinect sensor. Kinect op-

erates in the near-infrared range (830nm) therefore it

requires external illumination of the target which lim-

its its range.

Figure 2 shows a small rodent (Phodopus sun-

gorus) similar in size to the common vole in near-

infrared image. The picture was taken in complete

darkness, the target was illuminated by near-infared

light. Note that even though the animal is recogniz-

able, it would be hard to create a computer algorithm

that spots it. In this image the animal is close to the

camera but as the distance grows, illuminating the tar-

get becomes more and more complicated.

Long-wavelength infrared (LWIR) cameras detect

the infrared radiation emitted by the object in the pic-

ture hence they do not require infrared illumination

of the target. LWIR cameras has existed for a long

time but their price and other restrictions (e.g. export

control) conﬁned them to specialist use cases. Rela-

tively low cost LWIR cameras appeared just recently.

We experimented with FLIR Lepton camera module

whether small rodents can be detected reliably. This

camera operates in the 8000-14000 nm wavelength

range and has a resolution of 80x60 pixels. An animal

similar to the common vole (Phodopus sungorus) was

http://www.ﬂir.com/cores/content/?id=66257

Figure 2: Small rodent similar to a common vole (Phodopus

sungorus) in near-infrared image. The animal is marked by

the red ellipse.

placed in a cage and images were captured with dif-

ferent distances between the camera and the animal.

The background was lawn and other common foliage.

The images were made in the night.

Figure 3 shows the result. Minimum and maxi-

mum intensity values were calculated from raw image

(automatic gain control (AGC) switched off) and the

0-255 pixel intensity in the resulting greyscale image

was mapped into this minimum-maximum range so

that

greyscale

= 255

raw

− p

raw,min

raw,max

− p

raw,min

(1)

where p

raw

is the intensity value of the raw image

pixel produced by the Lepton camera, p

raw,max

and

raw,min

are the maximum and minimum intensities

detected in the raw image, p

greyscale

is the pixel in-

tensity value produced for the greyscale image.

When the animal is farther from the camera, its

observed infrared radiation decreases as the size of the

animal gets closer to the size of a pixel in the image.

The advantage of the dynamic calculation of the pixel

intensity range is that the animal remains bright, even

if it is farther from the camera. Disadvantage is that

this method makes features in the background with

higher infared radiation (trees, bushes, etc.) to be-

come more pronounced (”brighter”) in case of larger

distances to the target which makes the identiﬁcation

of the animal in the image more challenging. The an-

imal is clearly visible up to 4 meters of distance, in

fortunate situation the distance can be as large as 5

meters.

The experiments made with the FLIR Lepton

LWIR camera convinced us that small animals can be

reliably recognized with infrared camera, even with

such a limited resolution as the FLIR Lepton has. Due

to the non-trivial infared signature of the background,

however, we had to develop an image processing al-

gorithm to identify the pictures which are suitable for

population estimation.

SENSORNETS 2016 - 5th International Conference on Sensor Networks

Figure 3: Small rodent similar to a common vole (Phodopus

sungorus) in long-wavelength infrared image.

3 IMAGE PROCESSING

ALGORITHM IN THE SENSOR

According to the architectural goal of the AgroDat.hu

project, data processing happens on the server side.

The sensor unit, however, is responsible of sending

data that can be meaningfully used to extract relevant

information, in our case the population estimate of the

small rodents in the area. These animals move rela-

tively quickly so for meaningful observation, the bait

area needs to be monitored with an interval of some

seconds. Sending an image with this frequency is in-

feasible for mobile bandwidth and power consump-

tion reasons. The sensor itself must make the ﬁrst ﬁl-

tering and upload only pictures which has high prob-

ability of containing relevant information.

The requirements of the image processing algo-

rithm are the following.

• Remove large non-moving elements from the im-

age. These are supposed to be background ele-

ments.

• Look for relatively small objects that are moving.

These are potentially the rodents we are looking

for.

Our implementation is based on OpenCV image

processing framework

and comprises the following

steps:

• A greyscale image is produced from the raw Lep-

ton output image according to the transformation

described in section 2.

• The greyscale image is transformed into a binary

image with a ﬁxed threshold of 204 (0.8*255). As

the greyscale intensity values are calculated dy-

namically, taking into account the minimum and

maximum intensity values in the raw image, this

step is really dynamic thresholding with respect to

the original raw infrared image. The high thresh-

olding limit is due to the assumption that our ro-

dents are among the warmest things in the night

scene in the agricultural ﬁeld.

• Contour tracing algorithm from the OpenCV li-

brary is applied, then the resulting contours’ con-

vex hull is ﬁlled. This step gets rid of spurious

noise in the image resulting from the thresholding

step.

• Elements in the image are dilated by a kernel of

6x6, then again contour traced. This step merges

features that are separated by just a gap of up to 6

pixels.

• The encloding circle of each resulting contour is

calculated.

• The enclosing circles of this iteration and the pre-

vious iteration are compared. In order for two en-

closing circles to be considered the same, their

overlapping area must be at least 70% of the

smaller circle’s area. Circles present in both the

current and the previous picture are removed from

the set of circles in the image. This ﬁltering step

ensures that the feature we are looking for needs

to move.

• If there are circles remaining that have not been

considered the same as any circle in the previous

image and the any of the remaining circle’s radius

is smaller than 5 pixels then we have a candidate

image for uploading to the server.

Figure 4 demonstrates the steps of the image pro-

cessing algorithm. The image labelled as ”eq” is the

input greyscale image. ”th” is the result of the thresh-

olding, note the large amount of unconnected dots.

”c1” is the result of the ﬁrst contour tracing-convex

hull ﬁlling step. ”c2” is the output of the dilation-

second contour tracing. At the end, ”circle” shows

the resulting object circles identiﬁed in the picture.

Of the 5 circles identiﬁed, only one corresponds to

http://opencv.org/

Energy-efﬁcient Operation of GSM-connected Infrared Rodent Sensor

Figure 4: Demonstration of image processing steps in the

sensor. Image labels are detailed in section 3.

the test rodent. The rodent is identiﬁed when it starts

to move. Then its small circle will not overlap with

the circle on the previous picture, triggering an image

upload.

The algorithm above is presented as a demonstra-

tion that the image processing expected from the sen-

sor is not trivial and efﬁcient implementation is heav-

ily based on a publicly available library (OpenCV).

As we will see, these conclusions will have impor-

tant consequences when the power consumption of

the sensor is analyzed.

4 THE AGRODAT.HU SENSOR

NETWORK INFRASTRUCTURE

As the ﬁrst version of the AgroDat.hu sensor network

will target corn, typical cornﬁeld locations were con-

sidered when designing the communication architec-

ture. Due to large ﬁeld sizes and the production area

often located far from existing infrastructure, only a

radio technology with large coverage area was accept-

able. There are a number of alternative radio tech-

nologies with this characteristic (e.g. WiMax or cus-

tom VHF/UHF system) but due to its wide availabil-

ity, low cost and well-established regulatory frame-

work, we decided to use the GSM mobile network.

The ﬁrst version of the sensor network collects

data that change slowly (e.g. soil temperature, soil

moisture) and the data representation requires only

short data packets (with our coding format it means

200-400 bytes of data). This means that the sensor

communicates on the mobile network relatively rarely

(1-3 times a day) and even then only low amount

Figure 5: Conceptual communication architecture of the

AgroDat.hu sensor network.

of data is sent. Part of the sensor stations is only

equipped with underground sensors and minimally

protrude above ground level therefore solar cell-based

power supply was not possible. Energy efﬁciency

was a key requirement when designing the sensor sta-

tion. Due to the low amount of data transfer in the

ﬁrst version and the energy efﬁciency requirement,

the prorotype was designed using the low-bandwidth

services of the GSM network. This may mean GPRS

or SMS-based data transfer.

The second iteration added image and video trans-

fer to the requirement set. This new requirement

changes the amount of data to transmit. One infrared

image is typically in 3-4 kByte of size (80x60 pixel

resolution, 16 bit greyscale, PNG format). The sensor

may also send short videos (20-30 sec) created from

consecutive infared images. In MP4 format, these

videos have the typical size of 30-50 kBytes. Im-

ages captured by ordinary web cameras are up to 60

kBytes in PNG or JPEG format. This higher payload

size may require the usage of more recent telecommu-

nication technology, e.g. 3G or 4G. Our experience,

however, is that in the deployment areas of our sen-

sors, even basic GSM coverage may be problematic,

particularly with an antenna mounted so close to the

ground that some of our sensors have. For the exper-

iments reported in this paper, we did not change the

GSM/GPRS communication technology but analyz-

ing the possibility of 3G/4G is deﬁnitely a work to be

done.

Figure 5 shows the architecture used in the sec-

ond iteration of the AgroDat.hu sensor network. The

GPRS/HTTP is used to send bulk data to the server.

(Paller et al., 2015) demonstrated that the use of pro-

tocols more optimized than HTTP like CoAP does not

yield a signiﬁcant power consumption saving over the

GPRS bearer. The SMS infrastructure is employed

as the push bearer for server-initiated operations like

SENSORNETS 2016 - 5th International Conference on Sensor Networks

management operations. (Paller et al., 2015) ana-

lyzed the energy-efﬁciency of two ”push” solutions

and found that SMS is signiﬁcantly more energy-

efﬁcient than long-lived TCP connections. The man-

agement server initiates management operation by

sending SMs to the sensor. Some of these commands

like status query, changing a single conﬁguration vari-

able, reset request ﬁt into one SM and the entire op-

eration is carried out over the SM service. In case

of other commands like sensor log download or soft-

ware update, the SM sent by the management server

is used to trigger a GPRS connection to the manage-

ment server and the data transfer is carried out over

the GPRS connection. The main change compared to

the soil sensor is the sensor control component. In the

ﬁrst iteration, this component controlled a set of low-

bandwidth sensors with scalar value output therefore

an Atmel ATxmega128A4 microcontroller performed

the sensor control task. As argued in section 3, for

the camera sensor case we have a set of requirements

that demands more powerful sensor control. These

requirements are the following.

• In case of the cameras operating in visible light

frequency domain, we want a platform that fa-

cilitates the interfacing of popular cameras, e.g.

USB-connected webcams.

• We need a platform that is powerful enough to

execute the relatively complex image processing

tasks detailed in section 3.

• In order to implement the image processing logic

efﬁciently, we need a platform that is able to de-

ploy popular image processing frameworks like

OpenCV.

Considering these requirements led us to the con-

clusion that the sensor control component in our cam-

era sensor should be implemented based on an em-

bedded ARM-based computer (we used BeagleBone

Black, based on the TI Sitara AM335x ARM Cortex-

A8 system-on-chip (SoC)) and an embedded Linux

operating system. BeagleBone Black is supported by

a number of Linux distributions. The measurements

were made with 2 of them: Ubuntu Snappy

which

is a variant of the Ubuntu distribution particularly

targeted to Internet of Things (IoT) applications and

Texas Instrument’s own version of Linux speciﬁcally

targeted to the Sitara family of SoCs, called Linux

EZ Software Development Kit

. Snappy is attrac-

tive due to the large software base that Canonical, the

developer of the Ubuntu distribution constantly up-

dates. For example OpenCV and the libraries it de-

pends on are available as software packages for this

https://developer.ubuntu.com/en/snappy/

http://www.ti.com/tool/linuxezsdk-sitara

Figure 6: Image processing-image sending cycle in case of

communication logic running on the Sitara CPU.

platform. TI EZSDK on the other hand provides the

best available integration with the Sitara SoC’s hard-

ware features of which the power management was

particularly important for us.

5 IMAGE PROCESSING AND

POWER CONSUMPTION

It has been already pointed out in the literature

(Margi et al., 2006) that the power consump-

tion of a camera sensor node can be chatego-

rized as idle, processing-intensive, storage-intensive,

communication-intensive and visual sensing. In this

section we try to optimize the power consumption by

allocating idle consumption to processing units and

by ﬁnding a balance between processing-intensive

and communication-intesive tasks.

In the ﬁrst version of the AgroDat.hu sensors the

main control logic of the sensor was executed by the

Telit GL865 module which combines the GSM mo-

dem with a low-power microcontroller. As Snappy

Ubuntu is positioned as a platform with secure up-

date option, the ﬁrst sensor architecture we evaluated

was that both the image processing and the commu-

nication logic was executed by the main Sitara CPU

running Snappy Ubuntu Linux, the Telit GL865 was

used only as a modem. Figure 6 shows the power con-

sumption of the entire system (BeagleBone Black and

the GL865 GSM modem) when running the image

processing-image sending cycle. The image process-

ing cycle happens between the timestamps of 435-

440 seconds while the image sending is between the

timestamps of 440-540 seconds. The image size to be

sent to the server with a HTTP POST request was 4

Kbytes, over GPRS bearer.

The resulting power consumption for the image

sending cycle is 10.755 mAh which is much higher

than the expected consumption of about 1 mAh in

(Paller et al., 2015). The source of this signiﬁcant dif-

ference is the Sitara CPU’s base non-idle consump-

tion of 250 mA. This alone results in 250 mA*100

Energy-efﬁcient Operation of GSM-connected Infrared Rodent Sensor

Figure 7: Image sending cycle in case of communication

logic running on the GL865 module.

seconds=25000 mAs (about 7 mAh) consumption.

This example shows that the high-performance Sitara

performs very poorly from the power consumption

point of view if the computation pattern is of action-

and-wait type which is very common in case of

telecommunication state machines. In this case the

event the application code waits for happens too

quickly therefore the CPU cannot be put into an idle

state while the overall length of the execution is quite

long. This level of power consumption overhead was

unacceptable for us, hence we decided to rearchitect

the sensor so that the code is executed on a CPU that is

most suitable for the task with regards to power con-

sumption.

In the rearchitected model we returned to the orig-

inal sensor control model where the main control

logic is in the GL865 module. The microcontroller

in the GL865 is very efﬁcient in low-power, low-

performance execution and can sleep with very low

power consumption, using its internal real-time clock.

The experience also applies to different communica-

tion modules where the controller and the modem are

not integrated. In this case a separate low-power mi-

crocontroller can take care of the main control. Due

to its high active power consumption, the Sitara CPU

is kept in inactive state as long as possible.

Figure 7 shows the power consumption of the

GL865 sending the image of 4 Kbytes. The sending

cycle takes much longer (600 seconds vs. the 100 sec-

onds when the Sitara CPU executed the same logic),

this is due to the slowness of the GL865 Python en-

gine. The power consumption for the sending cycle is

much lower, however, about 3 mAh while the power

needed to send over the image from the Sitara CPU to

the GL865 over the serial line is negligible.

In case of the rearchitected sensor control, the

Sitara CPU waits for wakeup signal, acquires and pro-

cesses images and if suitable image is available, up-

loads the images to the GL865 over the serial connec-

tion between the two units. Then the Sitara CPU goes

to sleep. Currently we see the following use cases for

the combinations of timed picture acquisition vs. au-

tomatic identiﬁcation of relevant pictures.

Figure 8: Image processing power consumption, 5 itera-

tions, 500 msec inter-image delay.

• Images or video (infrared and/or visible-light) are

acquired at preconﬁgured moments of time. In

this case the advantage of the embedded Linux in

the sensor is its wide support of different cameras.

Image processing is not done.

• At preconﬁgured moments of time, automatic im-

age acquisition is triggered and if image process-

ing ﬁnds a relevant feature, the images/videos

(visible light/infrared) are uploaded to the server.

• Image acquisition/image processing is executed

continuously and images are uploaded to the

server if relevant feature is found by the image

processing pipeline. There are sub-requirements

about the acquisition time interval between the

images (from zero to up to 60 seconds) but the

images are not sent to the server at preconﬁgured

moments, only when the image processing algo-

rithm identiﬁes something important.

Depending on the use case, there are different

power consumption balance between the image pro-

cessing and image sending activities. For the ﬁrst

use case that sends images only at preconﬁgured mo-

ments, the use of high-performance SoC platform

with embedded Linux is not justiﬁed from the power

consumption point of view. There is some marginal

software engineering advantage as embedded Linux

comes with a large number of camera libraries and

image acquisition tools (e.g. V4L). Regarding the

third use case, image processing runs continuously

therefore this function consumes the most power,

there is no point in talking about power consumption

balance. Power consumption balance becomes rele-

vant in the second use case where images taken may

not be sent if there is no relevant feature on them. The

image processing performed to ﬁgure out if the image

is worth sending may be justiﬁed by lower amount

of data to transfer and lower power consumption as

GSM communication is an expensive operation from

the power consumption point of view. Prior literature

about camera sensor power consumption has already

pointed out that tasks with lower energy consumption

should trigger tasks with higher energy consumption

SENSORNETS 2016 - 5th International Conference on Sensor Networks

(Kulkarni et al., 2005) and in this case the image pro-

cessing corresponds to the task in with lower hierar-

chy level to trigger the GPRS transfer.

Figure 8 shows the power consumption of an im-

age acquisition/processing activity on Ubuntu Snappy

15.04. One iteration comprises image acquisition

from the infrared camera, running the algorithm de-

scribed in section 3 and waiting 500 msec if no

relevant feature was found. The image acquisi-

tion/processing was repeated 5 times, consuming

about 0.62 mAh. It can be observed that the power

consumption increased for the duration of the activ-

ity as clock scaling option was active (cpufreq with

”ondemand” governor). This automatically increased

the clock speed to the maximum of 720 MHz from

the base of 275 MHz when the CPU was active. Even

though the Sitara SoC contains a GPU, this time it was

not used because the small size (80x60) of the infrared

images would not make the GPU usage efﬁcient.

Considering strictly the image acquisition/image

processing step, the 0.62 mAh consumption of this

5-iteration activity compares favorably to the 3 mAh

consumption of sending the image with the GL865.

The Sitara CPU has a signiﬁcant idle consumption,

however. Our prototype was implemented on the

Ubuntu Snappy distribution which at the moment of

writing this paper, does not offer CPU idling support.

This means that an inactive CPU still consumes about

250 mA (same as active CPU with no load), consum-

ing 3 mAh (the cost of sending one image) in just 43

seconds. The TI EZSDK implements one sleep state,

the suspend-to-RAM (S3) state. TI EZSDK can enter

and exit this state in 3 seconds but the consumption

in this state is still 156 mA, which means 69 seconds

to reach 3 mAh. Ubuntu Snappy 15.04 consumes 120

mA even in shutdown state but TI EZSDK properly

shuts down. Unfortunately, a full shutdown-reboot

consumes 4.78 mAh with TI EZSDK which is more

than the 3 mAh required to send the image. Idling the

Sitara CPU with shutdown is therefore not an option.

Our conclusion is that saving battery power and

cellular data transfer by putting more intelligence into

the sensor and preﬁltering image data there is still an

attractive option. Unfortunately the current platforms

are inadequate from the power consumption point of

view, in particular the idle state management needs

more improvement. Until the embedded Linux sytem

can be placed into a state with near-zero consumption

in relatively short time, efﬁcient battery-powered op-

eration is not possible.

We found that a use case exists for applications

with image processing in the sensor. If the require-

ment is to monitor the environment continuously with

short image capture interval and the data link to the

server side is relatively slow, detection of the relevant

features must be done in the sensor. This was the case

for our rodent detection use case. According to our

experiments, if the image processing is implemented

on the BeagleBone Black using high-productivity,

popular software stacks (e.g. Linux/OpenCV), the

power consumption will be very high. It is certainly

possible to decrease this high power consumption

with dedicated hardware (e.g. microcontrollers) but

the software productivity will drop dramatically as

powerful image processing frameworks are not avail-

able for these devices. The outcome is that continuous

monitoring with high-productivity frameworks is an

expensive choice from the power consumption point

of view.

6 CONCLUSIONS

Camera sensors have been deployed in the agricul-

ture for various use cases. Most of the applications

tried to infer the health and development of the plants

based on image data in various wavelength domains.

These applications are simple from the sensor point

of view as capturing/sending images at predetermined

moments is usually enough. The larger data payload

that these sensors generate would justify the usage of

a more recent cellular standard (3G/4G) but coverage

is spotty in the areas of our interest. We intend to

analyze more the question of 3G coverage in areas

relevant for agricultural activity.

In our research, we looked for a use case that re-

quires more sophisticated processing in the sensor and

we found that rodent population estimation is an eco-

nomically relevant application and due to the quick

movement of the target animals, fast capture interval

is required. We also found that a bait area can be

efﬁciently monitored with a reasonably priced long-

wavelength infrared camera.

It was an attractive proposition that the power

consumption of the sensor system can be efﬁciently

decreased with image processing because the sen-

sor can ﬁlter out non-relevant images. Initial anal-

ysis of power consumption cost of a relatively com-

plex image processing operation was promising. Un-

fortunately the idle state support of the embed-

ded Linux platform of our choice prevented the ex-

ploitation of this possibility. We found that con-

tinuous image capture/monitoring is a use case that

still requires image processing capability in the sen-

sor but energy-efﬁcient implementation is not sup-

ported with the popular software stack we evalu-

ated (Linux/OpenCV). There is a trade-off here be-

tween implementation productivity and energy efﬁ-

Energy-efﬁcient Operation of GSM-connected Infrared Rodent Sensor

ciency with these platforms.

We propose the following directions to resolve

this trade-off. One direction is to improve the idle

state management of embedded Linux systems. This

is not trivial due to the complex interaction between

the peripherals (e.g. network cards) and the tasks run-

ning on the main processor. The expectation here

is that the Linux system may be placed into a state

where it consumes near to zero current. The other di-

rection is to port software libraries necessary for im-

age processing to microcontrollers without operating

systems. This would eliminate Linux and its power

management complexity entirely from the picture but

it would also make impossible the exploitation of se-

curity management and prepackaged software of em-

bedded Linux systems.

ACKNOWLEDGEMENTS

I would like to thank the Government of France

for the generosity in granting a scholarship to the

ESEO institute in Angers, France. I thank especially

ebastien Aubin at ESEO for facilitating my stay.

REFERENCES

Adamchuk, V. I., Hummel, J., Morgan, M., and Upadhyaya,

S. (2004). On-the-go soil sensors for precision agri-

culture. Computers and electronics in agriculture,

44(1):71–91.

Alderfasi, A. A. and Nielsen, D. C. (2001). Use of crop

water stress index for monitoring water status and

scheduling irrigation in wheat. Agricultural Water

Management, 47:69–75.

Armstrong, R., Barthakur, N., and Norris, E. (1993). A

comparative study of three leaf wetness sensors. In-

ternational journal of biometeorology, 37(1):7–10.

Grant, O. M., Chaves, M. M., and Jones, H. G. (2006).

Optimizing thermal imaging as a technique for de-

tecting stomatal closure induced by drought stress un-

der greenhouse conditions. Physiologia Plantarum,

127(3):507–518.

Jacob, J. and Hempel, N. (2003). Effects of farming prac-

tices on spatial behaviour of common voles. Journal

of Ethology, 21(1):45–50.

Kulkarni, P., Ganesan, D., Shenoy, P., and Lu, Q. (2005).

Senseye: a multi-tier camera sensor network. In Pro-

ceedings of the 13th annual ACM international con-

ference on Multimedia, pages 229–238. ACM.

Li, L., Zhang, Q., and Huang, D. (2014). A review of

imaging techniques for plant phenotyping. Sensors,

14(11):20078.

opez, J. A., Soto, F., S

anchez, P., Iborra, A., Suardiaz, J.,

and Vera, J. A. (2009). Development of a sensor node

for precision horticulture. Sensors, 9(5):3240–3255.

Margi, C. B., Petkov, V., Obraczka, K., and Manduchi, R.

(2006). Characterizing energy consumption in a vi-

sual sensor network testbed. In Testbeds and Research

Infrastructures for the Development of Networks and

Communities, 2006. TRIDENTCOM 2006. 2nd Inter-

national Conference on, pages 8–pp. IEEE.

Moshou, D., Bravo, C., West, J., Wahlen, S., McCartney,

A., and Ramon, H. (2004). Automatic detection of

yellow rustin wheat using reﬂectance measurements

and neural networks. Computers and electronics in

agriculture, 44(3):173–188.

Ou-Yanga, T.-H., Tsaib, M.-L., Yenc, C.-T., and Lina, T.-

T. (2011). An infrared range camera-based approach

for three-dimensional locomotion tracking and pose

reconstruction in a rodent. Journal of Neuroscience

Methods, 201:116–123.

Paller, G., Sz

armes, P., and

o, G. (2015). Efﬁcient

power consumption strategies for stationary sensors

connected to GSM network. In SENSORNETS 2015:

Proceedings of the 4th International Conference on

Sensor Networks., At Angers, France, Volume: pp. 63-

68., editors: C

esar Benavente-Peces, Patrick Plain-

chault, Octavian Postolache.

SENSORNETS 2016 - 5th International Conference on Sensor Networks