Classifying sows’ activity types from acceleration ... · Classifying sows’ activity types from...
Transcript of Classifying sows’ activity types from acceleration ... · Classifying sows’ activity types from...
Classifying sows’ activity types from
acceleration patterns
An application of the Multi-Process
Kalman Filter
Cecile Cornou a,*, Søren Lundbye-Christensen b
a Department of Large Animal Sciences, Faculty of Life Sciences, University of Copenhagen,
Groennegaardsvej 2, 1870 Frederiksberg C. Copenhagen, Denmarkb Institute of Mathematical Sciences, Center for SundhedStatistik, Aalborg University,
Fredrik Bajers Vej 7G, 9220 Aalborg SØ, Denmark
Accepted 25 June 2007
Available online 13 August 2007
Abstract
An automated method of classifying sow activity using acceleration measurements would allow the
individual sow’s behavior to be monitored throughout the reproductive cycle; applications for detecting
behaviors characteristic of estrus and farrowing or to monitor illness and welfare can be foreseen. This
article suggests a method of classifying five types of activity exhibited by group-housed sows. The method
involves the measurement of acceleration in three dimensions. The five activities are: feeding, walking,
rooting, lying laterally and lying sternally. Four time series of acceleration (the three-dimensional axes, plus
the length of the acceleration vector) are selected for each activity. Each time series is modeled using a
Dynamic Linear Model with cyclic components. The classification method, based on a Multi-Process
Kalman Filter (MPKF), is applied to a total of 15 times series of 120 observations, which involves 30 min
for each activity. The results show that feeding and lateral/sternal lying activities are best recognized;
walking and rooting activities are mostly recognized by a specific axis corresponding to the direction of the
sow’s movement while performing the activity (horizontal sidewise and vertical). Various possible
improvements of the suggested approach are discussed.
# 2007 Elsevier B.V. All rights reserved.
Keywords: Group-housed sows; Body activity; Dynamic Linear Models; Multi-Process Kalman Filter
www.elsevier.com/locate/applanim
Applied Animal Behaviour Science 111 (2008) 262–273
* Corresponding author. Tel.: +45 35333364; fax: +45 35333055.
E-mail addresses: [email protected] (C. Cornou), [email protected] (S. Lundbye-Christensen).
0168-1591/$ – see front matter # 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.applanim.2007.06.021
1. Introduction
When sows are group-housed it can be difficult to gain access to individual animals. Often this
leads to serious management problems. The development of sensor technology (Eradus and
Jansen, 1999) opens up new possibilities for monitoring single animals within a group, and
current automation systems aim to facilitate ’management by exception’ by drawing the farmer’s
attention to particular individuals.
A large range of automation systems for animal husbandry are based on Dynamic Linear
Models and the Kalman Filter (Kalman, 1960). Thus in pig production it is possible to monitor
the condition of young pigs via their drinking behavior (Madsen et al., 2005); and in group-
housed sows it is possible to monitor estrus via individual body activity (Cornou and Heiskanen,
submitted for publication). A similar approach has been described for monitoring milk quality in
dairy cattle (Thysen, 1993), and de Mol et al. (1999) suggest a method of this kind for detecting
estrus and diseases. Finally, an application for use in poultry production is presented in Roush
et al. (1992).
The behavior of the individual sow can be affected both by its physiological state and by
illness: body activity tends to increase at the onset of estrus (Cornou and Heiskanen, submitted
for publication; Freson et al., 1998; Geers et al., 1995; Serlet, 2004); nest-building behavior is
performed at the approach of farrowing (Jensen, 1993) lameness mainly influences the sow’s
walking activity, while other diseases may affect specific behaviors such as feeding (Forbes,
1995). Automated monitoring of the activities of an individual group-housed sow would
therefore help the farmer to detect deviations from normal behavior and provide information
about the specific state of the animal.
The objective of this study is to develop a method for automatically classifying particular
activities that group-housed sows perform. The method tracks acceleration measurements. An
accelerometer, fixed on individual sows, allows activity data to be recorded at any time. The
modeling of activity patterns could allow the individual animal to be monitored for the full
duration of its reproductive cycle, i.e. from the mating section to the farrowing house. Other
applications, such as monitoring animal welfare, can also be foreseen.
The following section describes the collection of acceleration measurements and the five types
of activity selected. Section 3 sets out the methods used to model and classify the activity types.
Section 4 presents and evaluates the results. Section 5 further discusses the results; it explores
perspectives for improvement and suggests new applications of the classification method
presented.
2. Time series of accelerations and activity types
2.1. Collection of acceleration measurements
The time series of acceleration measurements referred to in this article are extracts of data
collected from five group-housed sows in a production herd in Denmark over a period of 20 days
during March 2005. The sows were fed ad libitum; they had access to two electronic sow feeders
(ESF) and three nipple drinkers. Resting areas were straw-bedded and activity areas had plain or
slatted floors. Acceleration data were measured in three dimensions using a digital accelerometer
(LIS3L02DS from STMicroelectronics) four times per second, 24 h a day. A box containing the
accelerometer and the battery package was fitted on a neck collar which was put on the
experimental sow. The neck collars tended to loosen after few days. However, the weight of the
C. Cornou, S. Lundbye-Christensen / Applied Animal Behaviour Science 111 (2008) 262–273 263
box ensured that it generally stayed in position under the neck. Data were transferred to two PCs
via an external Bluetooth dongle which hung from the ceiling in the middle of the pen. Video
recordings covering a large proportion of the pen were also performed over 20 days, 24 h a day
(four pictures being recorded per second). The experimental protocol is described in detail in
Cornou and Heiskanen (submitted for publication).
Acceleration is a vector quantity defining the rate at which the sow changes its velocity. The
sow is treated as accelerating if its velocity is changing. The initial series included values for the
three axes (x, y and z), measured in volts. Of these axes, x corresponded to the vertical dimension;
y corresponded to the horizontal dimension, with the acceleration being measured sidewise,
while z corresponded to the horizontal dimension, with the acceleration being measured
forwards. Before further processing, the data were converted into the acceleration unit (g) and the
length of the acceleration vector was calculated as
acc ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiacc2
x þ acc2y þ acc2
z
q(1)
where accx, accy and accz are the acceleration values for the axes x, y and z. Acceleration values
for the three axes ranged from 2 to 2. The values of the length of the acceleration vector (acc)
ranged from 0 to 3.46, where 1 g corresponds to no acceleration. When the accelerometer was
placed immobile on a plane surface, the acceleration values for the axes x, y and z, were
respectively 1, 0 and 0 g (the first value being due to the effect of gravity).
2.2. Selection of the activity types and time series associated
With the help of the video recordings, extracts of time series were associated with five activity
types. The activity types were: feeding (FE), rooting (RO), walking (WA), lying sternally (LS)
and lying laterally (LL). Fig. 1 shows the raw values of the length of the acceleration vector (acc)
for the five types of activity (four measurements per second). It can be seen that each activity type
displays a distinct acceleration pattern.
Some of these activities are not exclusive. For example, sows can root and walk at the same
time. Since a walking sow may also pause and stand immobile more or less frequently for a
limited period of time, walking activity, when measured over a sufficient time, will generally be
associated with short periods of temporary immobility.
The data sets selected for this study satisfied two criteria: (i) the selected activities fill the
entire period; (ii) the overlapping of activities are reduced to a minimum (e.g. in extracts
corresponding to the rooting activity the sows only root over a limited area in order to limit the
effects of the walking activity).
For the purpose of modeling and classifying the activities two data sets satisfying the above
criteria were used:
� Learning data set: 10 min of each activity type (presented in Fig. 1); this data set is used to
estimate the model parameters for each activity type.
� Test data set: 10� 2 min of each activity type; this data set is used in the implementation of the
classification method, after each activity type has been modeled.
Each data set included four time series of acceleration measurements: axes x, y, z and the
length of the acceleration vector acc (referred to as the axis acc in the following sections).
The total duration of the data sets was 10 h (30 min of each activity type � 5 activity types �4 axes). These data sets correspond to acceleration measurements collected for three
C. Cornou, S. Lundbye-Christensen / Applied Animal Behaviour Science 111 (2008) 262–273264
different sows: sows 1, 3 and 5 for feeding, walking and lying laterally; sow 5 for rooting;
sows 1 and 3 for lying sternally. The activities of rooting and lying sternally were generally
performed in the straw-bedded resting areas. The videos only covered parts of these
areas, which made it difficult to associate these types of series with videos recordings. Hence,
here, the number of sows for both activities is limited, compared with the other types of
activity.
The selected data sets are relatively short in length. This is, first, to satisfy the criteria
presented above (regarding the filling of an entire period and exclusivity). Secondly, the
experimental conditions hindered efforts to associate particular activity types for a longer period
of time: more often than not the timestamps of the video recordings and the accelerometers were
unsynchronized. Therefore, series were selected around the time at which the sows visited the
ESF; at this time correct timestamps were available. Furthermore, the shifting of the battery
inside of the accelerometer boxes may have resulted in axes inversions when the neck collar was
repositioned. The limited length of the series was also designed to provide an assurance, during
the development of this classification method, that the position of each axis would remain
identical.
C. Cornou, S. Lundbye-Christensen / Applied Animal Behaviour Science 111 (2008) 262–273 265
Fig. 1. Ten minutes selected extracts of the length of the acceleration vector (acc) for the five activity types (four
measurements per second); vertical axes indicate the value of acceleration (g); horizontal axes indicate the duration of the
extracts (min).
3. Modeling of the acceleration patterns and classification of the activities
Modeling of the activity patterns was performed using time series from the learning data set,
previously averaged per second. As Fig. 2 shows, the analysis of correlation, for the x axis, of the
feeding series of the learning data set showed periodic movement, with a period of 22 s.
In the learning data set, periodicity was only observed for this above series (1 out of 20). In the
test data set, periodicity in the range of 15–25 was observed in 12 of the 200 series (10 series� 5
activities � 4 axes). The patterns are pseudo-cyclic, with a smoothly changing wavelength;
therefore, the suggested model includes a gradually changing sinoid movement.
3.1. Model design
The general DLM is represented as a set of two equations (West and Harrison, 1997). The
observation equation (2) defines the sampling distribution for the observation Yt conditional on
an unobservable state vector ut. The system equation (3) defines the time evolution of the state
vector ut.
Yt ¼ F>t ut þ nt; nt �Nð0;VÞ (2)
ut ¼ Gtut�1 þ vt; vt�Nð0;WtÞ (3)
The error sequences nt and vt are assumed to be internally and mutually independent. The DLM
combined with a Kalman Filter (KF) (Kalman, 1960) estimates the underlying state vector ut by its
conditional mean vector mt and its variance-covariance matrix Ct (the model variance) given all
previous observations Dt ¼ fY1; . . . Ytg of the acceleration measurements. Thus, the conditional
C. Cornou, S. Lundbye-Christensen / Applied Animal Behaviour Science 111 (2008) 262–273266
Fig. 2. Autocorrelation function for the x axis of the feeding series of the learning data set. A periodicity of 22 s is
observed.
distribution of ut is
ðutjDtÞ�N ðmt;CtÞ: (4)
The updating equations of the KF used for stepwise calculation of mt and Ct can be found in West
and Harrison (1997).
The suggested DLM includes a sine-cosine movement that follows the sinoid movement of the
observation data: the state vector ut consists of a set of parameters describing the model level (mt)
and the sine-cosine components (st, ct) at time t, i.e.
ut ¼mt
st
ct
0@
1A (5)
The systems matrices are defined in (6): Ft is labeled the design matrix; Gt is labeled the system
matrix, which is defined here as identity matrix.
F>t ¼�
1; sin
�2p
Tt
�; cos
�2p
Tt
��Gt ¼ I (6)
The period T is defined in the design matrix (6). However, the random variation over time of the
sine-cosine components of the state vector ut (5) allows the model to adapt to periodic
movements with periods varying near T.
The system variance Wt ¼ W (3) is defined as:
W ¼Wm 0 0
0 W sc 0
0 0 W sc
0@
1A (7)
The observation variance V (a scalar) and the parameters Wm and W sc of the system variance W,
characteristic of each axis of the respective activities, were estimated using the EM algorithm
(Dempster et al., 1977; Jørgensen et al., 1996; Dethlefsen, 2001). The EM algorithm is an
iterative algorithm used to estimate unknown parameters by maximum likelihood estimation; it
uses the conditional mean vector mt and the model variance Ct from the Kalman filtering and
their respective smoothed components mt and Ct obtained after Kalman smoothing (West and
Harrison, 1997).
The estimated values for the parameters bV , bWmand bW sc
, for each axis of each activity,
converged after 200 iterations. These values are available on request.
3.2. Multi-Process Kalman Filter
Following the procedure set out in Section 3.1, 20 DLMs (5 activities� 4 axes) were defined.
Each DLM was described by the quadruple at each time t, denoted by:
Mt : fF;G;V;Wgt (8)
In the Multi-Process model of class I, a single DLM (out of a range of possible DLMs) is
appropriate for describing the entire time series. However, there is uncertainty about the ‘true’
value of the defining parameter vector a ¼ aði; jÞ, where aði; jÞ is the set of parameters for the 20
possible DLMs, i.e. the five activity types indexed by i (FE, WA, RO, LL, LS) and their respective
C. Cornou, S. Lundbye-Christensen / Applied Animal Behaviour Science 111 (2008) 262–273 267
axes j (x, y, z, acc).
Mt ¼ MtðaÞ ðt ¼ 1; 2; . . .Þ (9)
Each DLM MtðaÞ was analyzed using the updating equations of the Kalman Filter. At each
observation time t, the model one-step forecast mean f t and its respective variance Qt were
calculated. The posterior probabilities ( pt) were estimated for each DLM, as
ptðiÞ/ftðiÞ � pt�1ðiÞ; (10)
where ftðiÞ is the predictive distribution of the observation given both the past Dt�1, and that
model i is appropriate.
In practice, each DLM was analyzed using the parameters estimated from the learning data
set; the value of the period T was set to 22. Initial values of the posterior probabilities were set to
0.2, corresponding to a uniform initial distribution for the five activity types. At each time t, the
probabilities were updated for the 20 DLMs according to (10).
4. Results
Fig. 3 shows the evolution of posterior probabilities for two time series of 120 observations
(2 min of observation) corresponding to the walking activity. The upper frame (a) corresponds to
C. Cornou, S. Lundbye-Christensen / Applied Animal Behaviour Science 111 (2008) 262–273268
Fig. 3. Evolution of the posterior probabilities for the five activity types, for two time series of 120 observations
corresponding to the walking activity (a) y axis and (b) z axis.
the y axis. The MPKF tends to indicate that the sow is rooting. This indicates that walking and
rooting activities have similar movements in a horizontal sidewise direction. Where the axis z
(lower frame) is concerned, the MPKF recognizes the ‘true’ activity type; the posterior
probabilities reach the value 1 after 40 observations.
Fig. 4 shows the mean posterior probabilities (wide line) for each axis for the five activity
types. The parameters estimated from the learning data set are applied to the respective activity
type/axis of the test data set.
The feeding activity is best recognized, while the walking activity is recognized only on one
axis, i.e. the horizontal z axis measuring forwards acceleration; the posterior probabilities exceed
the value 0.5 after approximately 30 s. Lying laterally is correctly recognized by the y axis
(horizontal sidewise direction). Lying sternally is also rapidly recognized. However, the results
here should be regarded with precaution, since both the learning and test data set are extracts of
time series from two sows only. The speed of recognition of the rooting activity is relatively slow,
and the posterior probabilities tend to stay around 0.5 for axes y and z; rooting is best recognized
by the x axis (vertical direction). Closer observation of the posterior probabilities of walking and
rooting activities indicates that these activities tend to be confused with each other: indeed for the
walking activity, in 8 out of the 10 times series the posterior probabilities of the y axis tend to
indicate that the sow is rooting. Similarly, for the rooting activity, the posterior probabilities of
the z axis indicate that the sow is walking in 8 out of 10 series. Particular axes perform better
C. Cornou, S. Lundbye-Christensen / Applied Animal Behaviour Science 111 (2008) 262–273 269
Fig. 3. (Continued ).
when used to classify specific activity types. These results are in keeping with the types of
movement that a sow performs for each activity type: sidewise movements are very limited when
a sow is lying laterally, so the activity is better recognized by the horizontal sidewise y axis.
Upwards and downwards movements are performed while the sows is rooting, which may
explain the better recognition on the vertical x axis. The fact that a sow tends to walk while
rooting may explain why neither of these activities are clearly recognized.
In order to corroborate the results of the classification method, new parameters were estimated
from the test data set and applied on the learning data set, which had previously been divided into
five series of 2 min. Table 1 indicates the percentage of posterior probabilities (pp) above 0.5
during the 2-min time series. The left panel shows the results from the test data set, where the
DLMs were analyzed using the parameters estimated from the learning data set. The right panel
C. Cornou, S. Lundbye-Christensen / Applied Animal Behaviour Science 111 (2008) 262–273270
Fig. 4. Posterior probabilities for the axes x, y and z, and acc for the five types of activity. Implementation of the
classification method on the test data set. Mean of the 10 time series (wide plain line) and 95% confidence intervals. The
horizontal axes indicate the observation time (s) (120 observations, i.e. 2 min).
shows the results from the learning data set, where the DLMs were analyzed using the
parameters estimated from the test data set.
The activities of feeding, rooting and lying sternally are well recognized on most axes of both
data sets. Walking activity is best recognized on the forward horizontal z axis; when the MPKF is
applied to the test data set, the very low percentage of recognition for the y axis confirms the
observation derived from Fig. 3(a): the classification method tends to indicate that the sow is
rooting. Where the acc axis is concerned, results indicate that the sow is either rooting or lying
sternally. The activity of lying laterally is also very poorly recognized on the acc axis: the
percentages of posterior probabilities > 0:5 are 12.7% and 2.5%, respectively, for the test and
learning data set, and the classification method tends to indicate that the sow is feeding; results
from the other three axes are in the range 56.2–97.9%. Differences between the two panels
observed here may be explained by the value chosen for the threshold, i.e. 0.5. As shown in
Fig. 3(a), some activities may be partly recognized, but the values of the posteriors probabilities
generally stay below, and rarely exceed, the 0.5 threshold.
5. Discussion and conclusion
The results of the classification method explored in this article show that all activity types can
be recognized, either by using all axes or by focusing on a specific axis. Both the three axes (x, y,
z) and the length of the acceleration vector (acc) were included. Reference to each of the three
dimensions appears to be desirable; this was especially seen in connection with the walking
activity, where the z axis (horizontal forward direction) showed the best results.
The results presented in Table 1 may be used as an indicator in an initial classification of the
activity types. The choice of a 2-min ‘window’ is in accordance with the average speed of
recognition by the axis, which can take up to 1 min (as happens, for instance, with the x axis of
rooting). Specific activities, such as feeding, may be of short duration: in our experimental
conditions the duration of feeding activity was approximately 10 min/day, which corresponds to a
sow consuming her entire ration at once. This also supports the argument for classifying activities
over a short time. In this study, a threshold of 0.5 was arbitrarily set in order to evaluate the results of
the classification method. As seen in Table 1, this resulted in differences in the recognition of the
same activity according to the data sets. Other threshold values should be tested and optimized by
mean of larger data sets. Other methods, based on time-moving windows (Shasha and Zhu, 2004),
may be used to detect whether or not the sow is active, using the length of the acceleration vector.
Class II Multi-Process models (West and Harrison, 1997, pp. 443–456) offer another kind of
C. Cornou, S. Lundbye-Christensen / Applied Animal Behaviour Science 111 (2008) 262–273 271
Table 1
Percentage of posterior probabilities where pp> 0:5 during the 2-min windows for the respective axes of each activity
type
Test data set Learning data set
x y z acc x y z acc
FE 79.3 91.9 87.3 90.8 90.2 39.0 95.5 92.2
WA 22.8 05.0 81.9 00.0 49.2 42.3 66.2 49.3
RO 57.6 54.1 36.5 78.1 46.5 64.3 74.7 86.7
LL 77.9 97.9 65.6 12.7 85.3 97.3 56.2 02.5
LS 81.7 87.7 82.3 82.1 70.5 87.0 71.7 42.7
Left panel: application of the MPKF on the test data set (1200 observations per result); right panel: application of the
MPKF on the learning data set (600 observations per result).
classification, over a shorter period of time. The interest of this method appears, however, to be
limited for two reasons. First, it is more complex: prior distributions for each activity type are
defined according to the average daily length of the activity, and this may vary from one sow to
another. Second, it is more demanding from a computational point of view, and this may limit its
interest so far as practical application is concerned.
Further developments of the method explored in this article may include combining the three
axes into a single multivariate model. For longer-term perspectives it may be desirable to fit
accelerometers to an ear tag. The use of an ear tag, rather than fitting accelerometers to a neck
collar, will generate more noise in the time series. If that happens, the direction of each axis may
become less clear, which will limit their effectiveness as separate entities.
The five activity types were selected by associating acceleration patterns with specific
activities observed in video recordings. The number and choice of activity types depends on the
practical application of the method being envisaged. In group-housed sows fed by ESF the
detection of feeding activity is less important, since the ESF already registers the sow entering
the feeding station. However, for group-housed sows not being fed by ESF, or sows at any other
stages (mating or farrowing section), the monitoring of feeding activity may be used to detect,
for example, illness: reduced feed intake is considered one of the first signs that an animal is ill
(Forbes, 1995). Some pairs of activities, such as standing and lying sternally, and walking and
rooting, may also present similar acceleration patterns, and may be difficult to distinguish one
from another. The interest of detecting walking and rooting separately is likely to be very
limited, except in the detection of lameness; and in that particular case a more detailed analysis
of the walking pattern should be carried out. It can be argued that a larger number of activities
will make acceleration patterns more difficult to recognize and may, therefore, affect the
performance and reliability of the classification method. Types of activity may also be grouped
and reduced to two general categories: ‘active behaviors’ (e.g. eating, walking, rooting) and
‘passive behaviors’ (e.g. lateral/sternal lying). These general categories could be used, as
indicators of general activity level, to detect estrus (Cornou and Heiskanen, submitted for
publication). Finally, the frequency of change in activity type, or posture, may be used as an
indicator of restlessness, in order to monitor, for example, parturition (Harris and Gonyou,
1998; Hartsock and Barczewski, 1997).
In further experimentation, inter-pig and inter-pen variations need to be explored. Particular
behaviors, such as walking activity, may be influenced by the size of the pen. However, the
method presented in this article does not detect the daily distribution of the duration of an activity
(which is needed in an MPKF of class II). Differences may be observed more readily within the
variance parameters, as a result of the speed of walking, for instance; this may nevertheless be
due to inter-pig variability. The data sets used in this study were obtained under ideal conditions:
a single activity type filled the whole period, and the activities were either not overlapping or such
that overlapping was reduced to a minimum. This may have resulted in overestimation of the
performance of the method. More extensive data sets including several types of activity should
now be tested. The use of a more complete data set, covering a larger number of individuals and
with optimal synchronization between video recordings and acceleration measurements, may
help researchers to carry out these further analyzes.
The main interest of classifying activity types automatically is to supplement visual
observation with automatic registration. This makes it possible to monitor a larger number of
individuals at the same time, instantaneously, if the method performs well. Development of a
method of automatically classifying activity types is the first step towards further automated
methods designed to detect estrus, farrowing, illness or welfare status.
C. Cornou, S. Lundbye-Christensen / Applied Animal Behaviour Science 111 (2008) 262–273272
In conclusion, then, the classification method presented in this article opens up new
possibilities for automatic monitoring of the types of activity an individual sow performs. Further
developments, involving the modeling each activity type, will require larger data sets to be used.
It will also be necessary to incorporate a method of assessing inter-pig and inter-pen variation.
Acknowledgements
The authors gratefully acknowledge the Department of Computer Science, Copenhagen
University, for assistance with data collection. Funding was provided by the Danish Research
Agency.
References
Cornou, C., Heiskanen, T. Pilot experiment: oestrus detection for group housed sows using sow’s acceleration
measurements, submitted for publication.
de Mol, R.M., Keen, A., Kroeze, G.H., Achten, J.M.F.H., April 1999. Description of a detection model for oestrus and
diseases in dairy cattle based on time series analysis combined with a Kalman Filter. Comput. Electron. Agric. 22 (2–
3), 171–185.
Dempster, A., Laird, N., Rubin, D., 1977. Maximum likelihood from incomplete data via the em algorithm (with
discussion). JRSS(B) 39, 1–38.
Dethlefsen, C., 2001. Space time problems and applications. Ph.D. thesis, Aalborg University.
Eradus, W.J., Jansen, M.B., 1999. Animal identification and monitoring. Comput. Electron. Agric. 24 (1–2), 91–98.
Forbes, J.M., 1995. Voluntary Food Intake and Diet Selection in Farms Animals. CAB International, Wallingford, Oxon
OX10DE, UK.
Freson, L., Godrie, S., Bos, N., Jourquin, J., Geers, R., 1998. Validation of an infra-red sensor for oestrus detection of
individually housed sows. Comput. Electron. Agric. 20 (1), 21–29.
Geers, R., Janssens, S., Spoorenberg, S., Goedseels, V., Noordhuizen, J., Ville, H., Jourquin, J., 1995. Automated oestrus
detection of sows with sensors for body temperature and physical activity. In: Proceedings of ARBIP95, Kobe, Japan.
Harris, M.J., Gonyou, H.W., 1998. Increasing available space in a farrowing crate does not facilitate postural changes or
maternal responses in gilts. Appl. Anim. Behav. Sci. 59 (4), 285–296.
Hartsock, T., Barczewski, R., 1997. Prepartum behavior in swine: effects of pen size. J. Anim. Sci. 75 (11), 2899–2904.
Jensen, P., 1993. Nest building in domestic sows: the role of external stimuli. Anim. Behav. 45, 351–358.
Jørgensen, B., Lundbye-Christensen, S., Song, P., Sun, L., 1996. State-space models for multivariate longitudinal data of
mixed types. Can. J. Stat. 24 (3), 385–402.
Kalman, R., 1960. A new approach to linear filtering and prediction problems. J. Basic Eng. 82.
Madsen, T., Andersen, S., Kristensen, A., 2005. Modeling the drinking pattern of young pigs using a state space model.
Comput. Electron. Agric. 48, 39–62.
Roush, W., Tomiyama, K., Garaoui, K., Alfonso, T., Cravener, T., 1992. Kalman Filter and an example of its use to detect
changes in poultry production responses. Comput. Electron. Agric. 6, 347–356.
Serlet, S., 2004. Optimalisatie van bronstdetectie bij zeugen. Thesis/dissertation.
Shasha, D., Zhu, Y., 2004. High Performance Discovery in Time Series. Springer.
Thysen, I., 1993. Monitoring bulk tank somatic cell counts by a Multi-Process Kalman Filter. Acta Agric. Scand. Sect. A
Animal Sci. (43), 58–64.
West, M., Harrison, J., 1997. Bayesian Forecasting and Dynamic Models, second ed. Springer.
C. Cornou, S. Lundbye-Christensen / Applied Animal Behaviour Science 111 (2008) 262–273 273