J. Behmann, K. Hendriksen, U. Müller, S. Walzog, H. Sauerwein, W

Automated recognition of activity states in dairy cows
by combining multiple sensors
J. Behmann, K. Hendriksen, U. Müller, S. Walzog,
H. Sauerwein, W. Büscher, L. Plümer
Objectives
Animals and Housing
• Continious and automated monitoring of
activities: lying, standing, standing up, lying
down, walking, feeding and drinking
• Frankenforst research station at the University of Bonn
• Interpretation of signals from spatial and nonspatial sensors
• Herd of 65 German Holstein Friesian cows
• Loose-housed in two row open free-stall barn (Fig.1)
• Cubicles and concrete floor

• High level information from low level data
• Detection of deviations from “normal behaviour“
• Aggregated daily routine stored in a “Digital
Dairy Diary“
• Provides data for root cause analysis
Sensors
Heart rate sensor (Polar)
Local Positioning Meaurement System (Abatec)
• x,y-Position at 0.5Hz
• Accuracy depends
strongly on position
within the barn
• Provides Heart Rate (HR)
and Heart Rate Variability
(HRV) at 1Hz

• Chest belt with sensor and
clock with reciever and
data logger
Data analysis methods
• Synchronizing the sensors and extraction of 13 spatial and
non-spatial features for each time step at 1Hz [1]

• Derivation of class probabilities by Support Vector Machine (SVM) [2]
• Linking the class probabilities to a chronological sequence by a
Condition Random Field (CRF) [3] and include contextual information
• Spatial constraints prevent some class transition (Fig.3)
• Distribution of activity durations differ significantly between the classes
 a cow remains longer in state “lying“ than in state “standing up“
Results


Resting Pulse Rate [bpm]
• 43 time series of 12 different cows
(each 4h)
• Continious recognition of activities (Fig. 4)
• Particularly good accuracy for „lying“
 Automated determination of the
Resting-Pulse-Rate (RPR, Fig. 5)
Day of pregnancy
Conclusions
References
• Automated monitoring system based on position and heart signals
[1] Mohr E, Langbein J, Nürnberg G (2002) Heart rate
variability: a noninvasive approach to measure stress in calves
and cows. Physiol Behav 75(1): 251–259
• Expendable to futher sensors, e.g. like rumination sensors,
pedometers and NFC chips
• Data analysis uses contectual knowledge about animals and barn
• Resting Pulse Rate was significantly correlated to day of
pregnancy
Institut für Geodäsie und Geoinformation
Professur für Geoinformation
http://www.igg.uni-bonn.de
Meckenheimer Allee 172
53115 Bonn
Tel.: +49 228 731756
[2] Cortes C, Vapnik V (1995) Support-vector networks. Mach
Learn 20(3): 273–297
[3] Lafferty JD, McCallum A, Pereira FCN (2001) Conditional
Random Fields: Probabilistic Models for Segmenting and
Labeling Sequence Data. In: Proceedings of the 18th
International Conference on Machine Learning, pp 282–289
Ansprechpartner:
M.Sc. Jan Behmann
[email protected]