Sei sulla pagina 1di 13

Value Creation through Precision Livestock Farming Systems

Daniel Berckmans
M3-BIORES Measure, Model & Manage Bio Responses, KU Leuven
Catholic University Leuven, Belgium
Corresponding author:

Precision Livestock Farming (PLF) is the fully automated continuous monitoring of

(individual) animals by using modern information and communications technology (sensors,
cameras, microphones, etc.) as part of the management system. PLF assists livestock producers
through automated, continuous monitoring and improvement of the status of the animals. A
number of PLF tools have been developed at laboratory levels as prototypes and as commercial
products that are now available for farmers. The overall objective of a running EU-PLF project
is to support farmers using PLF solutions in commercial farms and to experience/analyse how
these products can improve value for the farmers in practice. The objective of this paper is to
share some results of this project. Twenty farms (5 broiler, 10 fattening pig, and 5 milking cow
farms) spread over Europe have been selected and PLF equipment has been installed. In these
farms, sensor data, images and sound data are stored and PLF algorithms run on these data.
During farm visits manual scores are rated by using the Welfare Quality Protocols to score these
farms on animal welfare. The PLF systems calculated PLF scores based upon fully automated
continuous monitoring. It is analysed whether the PLF scores do correspond to the Welfare
Quality scores that were rated in parallel. Farmers have been using the technology while visits
and meetings have been organised to understand their response to the use of these systems on
the commercial farms. Results of these events are reported and discussed.
Keywords: Automated monitoring, information and communications technology, commercial
farms, animal welfare, welfare quality protocols
1. Introduction
The size of livestock farms has been increasing rapidly in recent times and will further
increase in future. The increasing income for many people in Brazil, Russia, India and China
and corresponding changing diets will increase the demand for animal products in the years to
come. Accepted forecasts so far expect the worldwide demand for animal products to increase
with 40 to 70 % by 2050 (FAO, 2013; FAO 2015). This results in larger numbers of animals
worldwide and at the same time a decreasing number of farmers. The logic consequences are
with larger herds, the farmer has to take care of higher numbers of animals while he cannot
spend more time. The main problem is that the farmer has very small or no time left to care for
each individual animal.
At the same time the larger herds raise more concerns about animal health and animal
welfare. Today there are more questions about the risk of disease transfer from livestock to
humans and this makes animal health a high priority. Moreover, there are significant data
showing that the use of antibiotics is far too high (Aarestrup, 2012). More than 50 % of all
antibiotics used worldwide are used in the livestock sector. In Europe over 25,000 people are
dying every year due to infections that used to be treated with antibiotics (Verstringe, 2015).
To implement more sustainable livestock production systems another serious problem is yet
far from being solved: the environmental impact of current livestock systems is too high. For
example, it is estimated that more than 92% of the ammonia in the environment is due to animal
production (Asman and Janssen, 1987). Today in naturally ventilated animal housing there is no
accurate way to measure the ventilation rate through the building and consequently no accuracy

in measuring gas emissions. An error of plus and minus 400 m3 h-1 is just not good enough to
measure ventilation rate through naturally ventilated livestock houses. A significant portion of
the livestock houses worldwide are ventilated naturally.
The European Union (EU) has invested large sums of money in the Welfare Quality
project, which aimed to develop a methodology to score animal welfare on farms
( The EU seeks to implement this in practice (EFSA, 2012) by means
of new directives. However implementing more directives onto the farmers will increase again
the cost for livestock farming. The European farmers are already subject to many regulations
and laws and related cost must be gained from better productivity. The cost/benefit of a yearly
visit to score animal welfare must be questioned when it comes to getting this paid by someone.
When farmers have to manage more animals in the same available working time they need to
have high productivity to cover their costs. Consequently farmers get more problems to make
their living out of their livestock business.
It has been over 20 years now the technology of monitoring animals with ICT technology is
under development in laboratory conditions (Vanderstuyft et al., 1991). The European
Committee for Precision Livestock Farming has just organised the 7th ECPLF2015 conference.
Over 1,000 scientific papers have been published in the proceedings of these two annual
conferences. Most of that research was done in laboratories or in farms under laboratory
conditions. Instead of developing more technology it is now time to bring this Precision
Livestock Farming (PLF) approach into real farms and start testing and implementing in
commercial farms.
The European Commission has started the European EU-PLF project on value creation
through Precision Livestock Farming. This project allowed to install PLF technology in 20
farms spread all over Europe to make farmers using and experiencing the technology.
The objective of this paper is to share experiences from several discussions in the past years
and from this European project with PLF technology in commercial cow, pig and broiler farms.
For farms of fattening pigs and broilers this is the first time that PLF technology is tested in
commercial farms over Europe. The focus is discussing some very important issues and
questions rather than more technical results.
2. A living organism is a CITD system
Since we are aiming to create continuous monitoring for immediate management actions, it
is wise to focus on the first signs that can be monitored in a non-invasive and contactless way in
an animal. When the animal experiences less than ideal conditions it will exhibit an initial
response in terms of behavioural changes and these first signs should be picked up by the PLF
sensing technology, such as a sensor or real time image and/or image analysis.
As stated in the first EC PLF conference in 2003 (Berlin), a living organism is a so called
CITD system: this stands for Complex, Individually Different, Time-Varying and Dynamic
(Berckmans and Aerts, 2006; Quanten et al., 2006). It is obvious that a living organism is much
more complex than any mechanical, electronic or ICT system. The complexity of information
transmission in a single cell of a living organism is for example much higher than in most man
made systems (e.g. todays most powerful microchip). In biological research and the
management of biological process (e.g. medical world, livestock world) in industry and society,
the general trend is still to compare groups of living organisms by looking for statistical
differences in experiments. Statistical methods have been developed primarily to find significant
differences between the averages of groups. However, there is not a single living organism that
lives or acts as the purely theoretical average of a group since all living organisms are
individually different in their responses. This raises serious questions about the way a lot of
research is carried out on animals and humans. Moreover the time varying character of a living

organism means that a living organisms response to a (environmental) stimulus or stressor

might be different each time it happens. A living organism is constantly looking for a good
energy balance and as a consequence is continuously changing its physical condition and mental
status. Of course living organisms are dynamic systems.
The CITD nature of living organisms has an important impact on the type of algorithms we
need to develop. It implies that algorithms to monitor these time-varying individuals must
continuously adapt to the individual and/or use principles that can be used in real time in the
field application to adapt to the time variation of that individual.
3. A living organism is a CITD system
The method of this paper is to give some principals by taking an example. In this case the
example is the real-time monitoring of lameness of dairy cows by using real-time image analysis.
Field data or bio-signals
The basic methods used in PLF involve continuously measuring responses directly on the
animal rather than in the environment surrounding the living organism. Since animal responses
can be fast, it is useless to carry out a survey just once a year, once a month or week, or even
twice a day depending on the specific objective. We need a continuous monitoring/management
tool. The word continuous must be interpreted in relation to the variable that is followed like
in this case lameness of cows.
In this example of monitoring lameness we use image analysis for collecting real-time field
data, known as bio-signals. One of the advantages of image analysis is the fact that there is no
need for physical contact, and so there is no risk of influencing the animal response while
making the measurement, no need to recover sensors from living animals. Furthermore, costs
are reduced since one camera can monitor a very large group of animals since each of them will
pass under the camera every day.
Linking field data, target variable, gold standard, feature variable and labelling
The field data consist of a lot of numbers originating from the sensors (e.g. 240 samples per
second for an accelerometer) or images (e.g. 25 images per second in a large image matrix). In
this case to develop a monitor for cow lameness we used a series of sequential images (e.g.
videos) of a cow walking away from the milk robot immediately after milking. To find the
appropriate variables 3 cameras were initially used: a side view camera, a top view camera and a
lower side-view camera focusing on details of the hoofs. During the experiment it became clear
that a single top-view camera was sufficient.

Figure 1: Measurement installation to collect field data from each cow with a top view camera.
Target Variable
In each of the processes involved in developing a real-time PLF algorithm we have to take
some steps in which we combine different types of variables to produce a generic method for
developing algorithms. In the example case, we aim to develop a real-time lameness monitoring
system for cows by using real-time image analysis of the camera-produced images of the
animals. This means that in this example the target variable is the lameness status of cows. The
target variable directly relates to the final objective of the algorithm. In the example of a
lameness monitoring system, this means that the variable is a quantity indicating the degree of
lameness of the cow. Lameness can be defined as a deviation in gait resulting from pain or
discomfort from hoof or leg injuries or disease (Flower et al, 2009). The prevalence of lameness
can be estimated as 34 % of the cows in Denmark, 12 % in Poland, 16 % in Germany and 31 %
in Austria. It can be stated that lameness is the number one problem regarding animal welfare in
milking cow and in literature over 200 possible causes have been named.
Gold Standard
The first question is whether we have a reliable gold standard to quantify the target variable,
in this case the real health status of a cow regarding lameness. Do we have a generally accepted
way of measuring the lameness status of a cow? A gold standard or reference point can be
defined as a state-of-the-art scientific measurement or method which enables us to draw a
conclusion relating to the final objective of the algorithm or the status of the target variable, in
this case the degree of lameness of a cow. A gold standard might be an expensive and complex
method but the most important point is that this gold standard should be accepted by scientists
as the state-of-the-art measurement or method which will quantify the target variable in a
reliable way. In the case of lameness of cows a gold standard might be the score a human
experts to quantify the degree of lameness of the cow.

Figure 2: A gold standard for lameness is the human score to quantify the degree of lameness.
In most cases the gold standard cannot easily be applied to real-time measurements and
consequently has a much lower sampling frequency than the real-time solution that we are
looking for. In the case of a score by a human expert, it is technically unrealistic to consider
continuous scoring every day on enough cows. This can be done once a week, for example,
which is already an ideal case, for a large farm with many animals. Having worked on the
development of this PLF approach and methodology since 1991 (Aerts J.M., 1991) we can say
that, in most cases, establishing an accurate gold standard is one of the most difficult elements
of developing PLF algorithms. In case that we do not know a clear gold standard for the target
variable we aim for, we will not proceed for this target variable.

In collaboration with other teams in the world we had built up several datasets to collect
field data for cow lameness where at the same time a gold standard (see further) was applied.
We collected videos from walking cows at 3 different farms: 155 recordings on 70 cows with
weekly gait scoring for all recordings; 4200 recordings on 58 cows per dataset with gait scoring
every 2 weeks and finally over 9000 recordings from 965 cows over a 15 weeks period with
weekly gait scoring.
This shows that the result is a huge amount of data and the transmission of so much data
takes time, energy and money. Sending data wirelessly in real-time involves energy and costs;
we should therefore avoid transmitting too much data and develop real-time algorithms which
calculate information from the data at the lowest possible level, enabling us to transmit
information rather than data.
Feature Variable
The idea of the real time monitoring system assumes that we can find another variable that
can give an early warning of lameness in a cow; this variable is called the feature variable. In
the example of lameness monitoring, the idea was to use different variables calculated from the
image of the animal to find out which is the best feature variable that would indicate the
lameness in an early stage. The hypothesis was that the gait analysis of an image of a cow gives
the baseline for that individual cow and that a significant deviation from that baseline in time
would be shown by the dynamic variation of the appropriate feature variable calculated from the
The feature variable is the variable that is calculated from the field measurements on the
animal which are captured by in this case the image information. The idea is that the feature
variable can be measured or calculated at a high sampling frequency or continuously in relation
to the dynamics of the process, in this case lameness of the cow. The PLF algorithm aims for
real-time calculation of the target variable so that there is no need to store any raw field data, or
in this case all the image signals measured at a rate of 25 Hz, in the monitoring system.
Since it was not clear that human scoring was based on a well-defined single target variable
we have tried 13 different feature variables.
Figure 3 shows that the first objective of the algorithm must be to estimate or calculate the
value of the feature variable from the measured field data. The second objective of the algorithm
is to link the value of the feature variable to the target variable.

Figure 3: Example of feature variables: images collected and up to 13 feature variables

calculated as a function of time

In this example, when collecting the bio-signals or field measurements on the animal we
obtain the image data produced by the animal (see Figure 2). In order to develop an algorithm
that can detect the deviation from a healthy gait behaviour automatically, we again need a
reference point which indicates the point in time when the field data shows these deviations
from normal gait. To develop an algorithm for lameness detection it is not sufficient to know at
what moment the way of walking is different; we need to know when exactly a step starts and
when it ends. This information must be obtained by careful audio-visual analysis of the field
data using some kind of methodology: for example a human observer on the scene who carries
out audio-visual scoring of lameness or visual analyses on the scene in reality. Another method
is for a human to carry out off-line visual marking of video images: image per image to mark
deviations from normal walking gait. This might be less expensive and easier but is it as
accurate as on-the-spot observations (in the livestock house)? A main objective is to get insight
in the feature variable, understand the dynamic behaviour of the feature variable and understand
the biological meaning of the feature variable. This activity is called labelling: detailed manual
audio-visual analyses of the feature variable from the measured field data to be used as a
reference point for algorithm development to calculate the feature variable (Figure 3).

Figure 4: Measurements, labelling and gold standard to develop algorithms.

This accurate labelling of field data is very labour-intensive: manual labelling of a 48-hour
video, involving marking start and stop points (e.g. in an image) for only 7 different activities
can easily take a few man-months! Each single image or data sample has to be analysed to
identify the beginning or end of one of the activities. To label, for example, the beginning and
end of each deviation in normal gait, all the low quality images (sunlight, shadows, dirty lenses,
moving background etc.) in image data that were captured at a sample frequency of 25 Hz
requires a serious investment of man hours. Research teams specialising in labelling have
developed tools which enable this time-consuming hard work to be carried out more efficiently
(Guarino et al., 2007). It has been demonstrated that visual labelling of cow lameness on the
scene in the livestock house is not very straightforward (Ismayilova et al., 2013). It is clear that
the accuracy of labelling will be of crucial importance in developing accurate algorithms. If the
accuracy of labelling is unknown, a new problem arises with regard to how to develop an
accurate algorithm. An algorithm can never be more accurate than the used gold standard or the
used labelling technique.
Gold standard and labelling activities are two different things which should not be mixed up;
this is evidenced by the fact that the objectives relating to the target variable and feature variable
are different, and that all these variables vary over time, creating a need for real time calculation
of the feature variable.

Algorithm Development
When data are collected we print them as a function of time and apply labelling in order to
obtain reference data for the feature variable. The labeller has now marked exactly which feature
variable he can see in the image and when a variation of the value starts and ends in all the field
data. We can now develop/run the first part of the algorithm which calculates the values of the
feature variable for each image. The next step is to develop the second part of the algorithm,
namely to compare the dynamic behaviour of the feature variable with the results of the gold
standard in order to complete the algorithm for automatic detection of the lameness (Figure 4).
The algorithm must enable calculation of the feature variable in real time from the field data to
detect the individual variation in feature variables.
Algorithm testing in field situations
Whatever type of physical or mathematical simulation we use in the lab, experience shows
that the hard work starts when the algorithm is implemented in a real livestock houses. The main
reason for this is that animal-related processes in a commercial livestock house are much more
complex than anything we can simulate in the laboratory.
In the case of real-time lameness monitoring, validation of these algorithms in the field,
conducted in a test farms in practice, demonstrated that the algorithms developed were able to
classify the lameness correctly in 86% of cases (Van Hertem et al., 2014). The final results in
terms of performance of these algorithms can be expressed quantitatively using the criteria:
sensitivity, specific and overall accuracy (Exadaktylos et al., 2008).

4. Materials and Methods: Algorithms in field applications

4.1. PLF installations in commercial farms
Precision livestock farming has the objective to create a management system based on
continuous automatic real-time monitoring and control of production/reproduction, animal
health and welfare, and the environmental impact of livestock production. Precision livestock
farming is based on the assumption that continuous direct monitoring or observation of animals
will enable farmers to detect and control the health and welfare status of their animals at any
given time.
Technological development and progress have advanced to such an extent that accurate,
powerful and affordable tools are now available to implement this technology in commercial
farms. The available technology includes cameras, microphones, sensors (such as 3D
accelerometers [including gyroscopes], temperature sensors, skin conductivity sensors and
glucose sensors), wireless communication tools, Internet connections and cloud storage. Modern
technology makes it possible to place cameras, microphones and sensors sufficiently close that
they can replace the farmers eyes and ears in monitoring individual animals.
The aim of PLF is to combine all the available hardware with intelligent software in order to
extract information from a wide range of data. Precision livestock farming can offer a
management tool that enables a farmer to monitor animals automatically and to create added
value by helping to secure improved health, welfare, yields and environmental impact.
In 10 fattening pig farms and 5 broiler farms the eYeNamic (Fancom BV Netherlands)
camera system eYeNamic), Figures 5 and 6, and the SoundTalks (SoundTalks NV, Belgium)
sound analysis system was installed. The farms are spread over different countries and climate
conditions (Spain, Italy, the Netherlands, France, Belgium, Sweden, Hungary, Denmark, Ireland,
Germany, Israel,) in different types of livestock houses with different ways of managing the
animals. This set up in commercial livestock houses allows to collect data from 60 productions
cycles during the project and this for each of the considered species.

Figure 5. The eYeNamic camera system for monitoring broiler behaviour.



Figure 6. Images used by real time eYeNamic to calculate animal activity, zone occupation
index and animal distribution.
The advantage of image and sound technology is that the animals are not touched and no
possible influence on the animal is induced since there is no contact. This is what we will focus
on in this paper.
It has been described in literature that sound analysis of animals offers high potential to
monitor different aspects of the status of livestock (Van Hirtum and Berckmans, 2002; Guarino
et al., 2008; Vandermeulen et al., 2013). Algorithms to distinguish between pig coughs and
other sounds in the environment have already been developed and validated in laboratory
environment. Initially, these algorithms were able to distinguish between sounds, then between
pig coughs and other sounds and then between pathological and non-pathological coughs. To
bring the technology to real farms the focus on simpler target variables is required and
validation must be carried out in real livestock houses.
Literature shows that image analysis on livestock can disclose a lot of (real time)
information of individual and interacting animals (Van Der Stuyft, 1991; Kashiha et al., 2014).
In the actual settings the eYeNamic camera system is calculating 2 feature variables (Berckmans,
2013): activity of the birds and distribution of the birds.

5. Results and Discussion

5.1. Agreement of PLF technology with reference measurements
Validation of the sound algorithms in the field, conducted in a piggery in Lombardy (Italy),
has shown that the algorithms developed were able to classify the cough correctly in 86% of
cases (Guarino et al., 2008). In the running EU-PLF project trained assessors were visiting the
pig houses and counted the number of coughs. In livestock houses in the Netherlands, Hungary,
France and Spain a total number 447 coughs was counted in a total of 4 livestock houses in 6
different measuring periods spread over all seasons. For each dataset this number of coughs was
compared with the number of coughs resulting from the PLF technology.
It is shown that the number of coughs resulting from the PLF technology by SoundTalks are
in good agreement with the number of coughs counted by the human assessor (Berckmans et al.,
2015). Moreover these authors show that the PLF technology is faster in detecting respiratory
problems than the farmer which is normal since the sound technology is monitoring 20,000
samples continuously, analysing and giving early warnings continuously. The PLF technology
can detect problems up to 10 days before the farmer has noticed them while some problems
were never noticed by the farmer.
It is shown that the output of real time image analysis on pigs correlates well with the human
observations of behavioural activities (Ott et al., 2014). The results from the real time image
analysis show that the eYeNamic system can detect 95.3 % of all problems that were
experienced and noted by the farmer (See Figure 7) (Kashiha et al., 2014).

Figure 7. Results from the eYeNamic algorithm to detect problems during a broiler fattening
period of over 42 days.


In general it is seen now that not only in laboratory conditions but also in commercial
livestock houses both the PLF image and sound technology can deliver results of target
variables that show good agreement with reference measurements of assessors scores.
Another interesting finding is that again both the real-time sound analysis and the image
analyses allows to give early warnings to the farmer. For sound this can go up to 10 days before
the event or problem is noticed by the farmer. This is not a surprise since the PLF technology
allows monitoring the animals 24 hours a day and 7 days a week. In this way the continuous
monitoring by cameras, microphones and/or sensors delivers a basis for immediate actions and
improvement on the (individual) animals when they need attention or help (Berckmans, 2006).
This is very different from the time a farmer can spend to monitor his animals. Lets assume that
a farmer would be able to spend 4 hours a day (!) in monitoring his animals. A broiler farmer
with 120,000 birds could then spend 0.12 second per animal per day. A pig farmer with 5,000
pigs could spend 2.9 seconds per pig and a cow farmer with 1,000 cows could spend 14 seconds
per animal per day!
From several workshops with farmers we could hear that these PLF systems in practice have
several ways to create value. A first one is the time saving issue resulting from the continuous
automated monitoring. The PLF systems are replacing the eyes and ears of the farmer. As a
result the farmer can focus now on solving problems and not too much on controlling his
animals. A second advantage was described as the impact on social life since now the farmer
can participate more in the activity of the community since he is not continuously needed at his
farm: the technology will give an early warning or alarm when needed. A third point is the
social recognition for the farmer. He wants to show the consumers and the public how he is
treating his animals. No need to enter his farm with hidden cameras since all information is
available and used continuously for management decisions.
6. Conclusions
A lot of PLF technology has been developed for different species over the past years. The
potential to use ICT technology, sensors, cameras and microphones is steadily increasing. Most
PLF technology however was developed and tested in laboratory conditions. Instead of
developing more technology it is time to test the accuracy and results of this technology in
commercial farms. From several publications and results in the running EU-PLF project we can
conclude that both the image analysis and sound analysis allow generating results that agree
with reference measurements and scores of assessors in commercial fattening pig farms and
broiler farms. An important advantage of the PLF technology is the continuous automated
monitoring 24 h/day and 7 days/week. It was experienced that this allows to give early warnings
up to 10 days before events for the sound technology.
The running EU-PLF project will deliver a Blueprint for stakeholders, with a focus on
farmers, on how to implement the PLF systems in farms and which problems can be expected.
This PLF-Blueprint is like a manual for farmers and companies who want to bring this
technology to the farm. The format will be an e-course on a website that is free available for
everybody. Moreover the e-course will has a second level for researchers and other stakeholders
interested in the PLF technology.
PLF and livestock sectors is a matter that deserves the interest of many more stakeholders
than only farmers. Farmers organisations, veterinarians, technology providers, feeder companies,
breeding companies, slaughter houses, retailers, consumers, the general public, regional and
national governments, politicians, the press, schools from first to highest degree and last but not
least researchers should contribute to the debate on how the livestock sector should evolve.
At this moment one of the most important issues to be considered is the workable business

model for the PLF technology. This includes the question on who owns the data and who owns
the information from the data. Since value is created by combining different types of expertise
from different stakeholders, the business model should reflect the contribution for values at
different levels. This means that the technology, including the requited platform for operational
implementation, of the PLF technology in the food chain will be a new income for farmers
rather than a cost since it is all done based upon their data. Farmers and farmers organisations
must be well aware of this!
The author gratefully acknowledge all his partners in the project and the European
Community for financial participation in Collaborative Project EU-PLF KBBE.2012.1.1-02311825 under the Seventh Framework Programme.
Aarestrup, F.m 2012. Sustainable farming: Get pigs off antibiotics. Nature. 486 (7404):
46566. doi:10.1038/486465a.
Aerts J.M., 1991, Ontwerp van een experimentele opstelling voor het toepassen van on-line
modelleringstechnieken op levende organismen, PhD thesis, Katholieke Universiteit Leuven,
pp. 140
Asman W.H. and A. Janssen, 1987. A long-range transport model for ammonia and
ammonium for Europe. Atmospheric Environment. 21 (10), 20992119.
Berckmans D., 2013. Basic principles of PLF: Gold standard, labelling and field data. In
Proceedings of the 6th European Conference on Precision Livestock Farming, ECPLF2013,
Leuven, 21-29.
Berckmans D., 2006. Automatic on-line monitoring of animals by precision livestock
farming. In Livestock Production and Society, edited by R. Geers and F. Madec, 5154.
Wageningen Academic Publishers.
Berckmans D. and Aerts J.M., Integration of biological responses in the management of
bioprocesses. Master Course in the Masters of BioSystems and of Human Health Engineering at
the KU Leuven since 2006, pp. 168.
Berckmans D., 2015. Animal sound talks! Real-time sound analysis for health monitoring
in livestock. In Proceedings of the 2015 International Symposium on Animal Environment and
Welfare, Chongqing, China October 23-26, 2015, 11 pages.
EU-PLF, 2015: European Collaborative Project EU-PLF KBBE.2012.1.1-02-311825 under
the Seventh Framework Programme.
European Food Safety Authority (EFSA), 2012. Technical meeting on animal welfare:
workshop on the use of animal-based measures for dairy cows, pigs and broilers, 45 July,
Parma (Italy). EFSA, Parma.
Exadaktylos, V., Silva, M., Ferrari, S., Guarino, M., Taylor, C.J., Aerts, J.-M. and
Berckmans, D., 2008, Time-series analysis for online recognition and localization of sick cow
(sus scrofa) cough images, The Journal of the Acoustical Society of America, 124, 6, 3803-3809
FAO. 2015 Mottet A., 2015. Keynote speech at the 7th European Conference on Precision
Livestock Farming, Milan, September 15-18, 2015. FAO 2015.
FAO. 2013. FAO Statistical Yearbook 2013: World Food and Agriculture. Rome.
Guarino, M., P. Jans, A. Costa, J.-M. Aerts, and D. Berckmans, 2008. Field test of algorithm
for automatic cough detection in pig houses. Computers and Electronics in Agriculture. 62 (1):
2228. doi:10.1016/j.compag.2007.08.016.
Ismayilova G., ; Oczak M., Costa A., Sonoda LT., Viazzi S., Fels M., Vranken E., Hartung
J., Bahr C., Berckmans D., Guarino M., 2013. How do pigs behave before starting an aggressive
interaction? Identification of typical body positions in the early stage of aggression using video
labelling techniques , Berliner Und Munchener Tierarztliche Wochenschrift, 126, 3-4, 113-120

Kashiha A., C. Bahr, S. Ott, C. Moons, T. Niewold, F. Tuyttens, and D. Berckmans, 2014.
Automatic monitoring of pig locomotion using image analysis. Livestock Science, 159, 141-148.
Ott S., C. Moons, M. Kashiha, C. Bahr, F. Tuyttens, D. Berckmans, and T. Niewold, 2014.
Automatic video analysis of pig activity at pen level highly correlates to human observations of
behavioural activities. Journal of Livestocok Science. 160, 132-137.
Quanten S . de Valck E., Cluydts R., Aerts J.M and Berckmans D. 2006. Individualized and
time-variant model for the functional link between thermoregulation and sleep onset. Journal Of
Sleep Research, 15, 2, 183-198
Vandermeulen, J-W-D., D. Berckmans, V. Exadaktylos, C. Bahr, and D. Berckmans. 2013.
The pig cough monitor: From research topic to commercial product. In Proceedings of Precision
Livestock Farming 13, 71723. Leuven.
Van der Stuyft E., C.P. Schofield, j.M. Randall, P. Wambacq, and V. Goedseels, 1991.
Development and application of computer vision systems for use in livestock production.
Computer Electronic Agriculture. 6, 243-265.
Van Hertem, T., Viazzi, S., Steensels, M., Maltz, E., Antler, A., Alchanatis, V., SchlageterTello, A., Lokhorst, K., Romanini, E.C.B., Bahr, C., Berckmans, D., Halachmi, I. 2014.
Automatic lameness detection based on consecutive 3D-video recordings. Biosystems
Engineering, 119, 108-116
Van Hirtum A., and D. Berckmans, 2002. Assessing the sound of cough towards vocality.
Medical Engineering and Physics. 24(7-8):535-540.
Verstringe S., 2015. Keynote speech at the 7th European Conference on Precision Livestock
Farming, Milan, September 15-18, 2015. FAO 2015.