Automatic Lameness detechtion by computer vision and behavior and performance sensing

T. van Hertem, M. Steensels, S. Viazzi, C.E.B. Romanini, A.A. Schlageter Tello, C. Lokhorst, E. Maltz, I. Halachmi, C. Bahr, D. Berckmans, S. Hong

Research output: Contribution to conferenceAbstractAcademic

Abstract

The objective was to compare automatic lameness detection methods based on daily automatic measurements of cow’s back posture and behavioral and performance variables. The experimental setup was located in a commercial Israeli dairy farm of 1,100 Israeli Holstein cows. All cows were housed in open, roofed cowsheds with dried manure bedding and no stalls. All cows were equipped with a commercial neck activity and ruminating time data logger. Milk yield was measured with a milk flow sensor. Cow gait recordings were made during 4 consecutive nighttime milking sessions with a 3D image camera. From the videos, the “inverse radius” of the back posture contour and the “back posture measurement” were extracted. The reference in this study was a daily live locomotion score of the animals. A dataset of 186 cows with 4 video-based lameness scores and 4 live locomotion scores was built. In total, 38 model input variables were constructed from the sensor data comprising absolute values, relative values, daily standard deviations, slope coefficients, daytime and nighttime periods and milk session related variables. Data of lame cows – cows recognized and treated for lameness – was compared with data of non-lame cows. A logistic regression model was built with the highest correlated behavioral and performance variables. Model validation was done with 10-fold cross-validation. The analysis of the video-based scores as independent observations leads to a correct classification rate of 53.0% on a 5-point level scale. A multinomial logistic regression model based on 4 consecutive “back posture measurement”-scores and “inverse radius”-scores obtained a correct classification rate of 60.8%. Strict binary classification to lame vs. not-lame categories reached 80.7% correct classification rate. In addition, the logistic regression model included 7 model input variables (the daily milk yield; the slope coefficient of the daily milk yield; the nighttime/daytime neck activity ratio; the milk yield week difference ratio; the milk yield week difference; the neck activity level during the daytime; the ruminating time during nighttime). After 10-fold cross-validation, the model obtained a sensitivity of 0.89 and a specificity of 0.85, with a correct classification rate of 0.86 when based on the averaged 10-fold model coefficients. The combination of image processing and behavioral monitoring is believed to further improve the lameness detection accuracy.
Original languageEnglish
Publication statusPublished - 2014
EventAgEng 2014 - Zurich, Switzerland
Duration: 6 Jul 201410 Jul 2014

Conference

ConferenceAgEng 2014
CountrySwitzerland
CityZurich
Period6/07/1410/07/14

Fingerprint

computer vision
lameness
cows
milk yield
posture
neck
model validation
sensors (equipment)
locomotion
milk
Daily Values
back (body region)
gait
milking
cameras
dairy farming
animal manures
Holstein
image analysis
taxonomy

Cite this

van Hertem, T., Steensels, M., Viazzi, S., Romanini, C. E. B., Schlageter Tello, A. A., Lokhorst, C., ... Hong, S. (2014). Automatic Lameness detechtion by computer vision and behavior and performance sensing. Abstract from AgEng 2014, Zurich, Switzerland.
van Hertem, T. ; Steensels, M. ; Viazzi, S. ; Romanini, C.E.B. ; Schlageter Tello, A.A. ; Lokhorst, C. ; Maltz, E. ; Halachmi, I. ; Bahr, C. ; Berckmans, D. ; Hong, S. / Automatic Lameness detechtion by computer vision and behavior and performance sensing. Abstract from AgEng 2014, Zurich, Switzerland.
@conference{6e51fa8991f942fab345fe32aa9ae9c6,
title = "Automatic Lameness detechtion by computer vision and behavior and performance sensing",
abstract = "The objective was to compare automatic lameness detection methods based on daily automatic measurements of cow’s back posture and behavioral and performance variables. The experimental setup was located in a commercial Israeli dairy farm of 1,100 Israeli Holstein cows. All cows were housed in open, roofed cowsheds with dried manure bedding and no stalls. All cows were equipped with a commercial neck activity and ruminating time data logger. Milk yield was measured with a milk flow sensor. Cow gait recordings were made during 4 consecutive nighttime milking sessions with a 3D image camera. From the videos, the “inverse radius” of the back posture contour and the “back posture measurement” were extracted. The reference in this study was a daily live locomotion score of the animals. A dataset of 186 cows with 4 video-based lameness scores and 4 live locomotion scores was built. In total, 38 model input variables were constructed from the sensor data comprising absolute values, relative values, daily standard deviations, slope coefficients, daytime and nighttime periods and milk session related variables. Data of lame cows – cows recognized and treated for lameness – was compared with data of non-lame cows. A logistic regression model was built with the highest correlated behavioral and performance variables. Model validation was done with 10-fold cross-validation. The analysis of the video-based scores as independent observations leads to a correct classification rate of 53.0{\%} on a 5-point level scale. A multinomial logistic regression model based on 4 consecutive “back posture measurement”-scores and “inverse radius”-scores obtained a correct classification rate of 60.8{\%}. Strict binary classification to lame vs. not-lame categories reached 80.7{\%} correct classification rate. In addition, the logistic regression model included 7 model input variables (the daily milk yield; the slope coefficient of the daily milk yield; the nighttime/daytime neck activity ratio; the milk yield week difference ratio; the milk yield week difference; the neck activity level during the daytime; the ruminating time during nighttime). After 10-fold cross-validation, the model obtained a sensitivity of 0.89 and a specificity of 0.85, with a correct classification rate of 0.86 when based on the averaged 10-fold model coefficients. The combination of image processing and behavioral monitoring is believed to further improve the lameness detection accuracy.",
author = "{van Hertem}, T. and M. Steensels and S. Viazzi and C.E.B. Romanini and {Schlageter Tello}, A.A. and C. Lokhorst and E. Maltz and I. Halachmi and C. Bahr and D. Berckmans and S. Hong",
year = "2014",
language = "English",
note = "AgEng 2014 ; Conference date: 06-07-2014 Through 10-07-2014",

}

van Hertem, T, Steensels, M, Viazzi, S, Romanini, CEB, Schlageter Tello, AA, Lokhorst, C, Maltz, E, Halachmi, I, Bahr, C, Berckmans, D & Hong, S 2014, 'Automatic Lameness detechtion by computer vision and behavior and performance sensing' AgEng 2014, Zurich, Switzerland, 6/07/14 - 10/07/14, .

Automatic Lameness detechtion by computer vision and behavior and performance sensing. / van Hertem, T.; Steensels, M.; Viazzi, S.; Romanini, C.E.B.; Schlageter Tello, A.A.; Lokhorst, C.; Maltz, E.; Halachmi, I.; Bahr, C.; Berckmans, D.; Hong, S.

2014. Abstract from AgEng 2014, Zurich, Switzerland.

Research output: Contribution to conferenceAbstractAcademic

TY - CONF

T1 - Automatic Lameness detechtion by computer vision and behavior and performance sensing

AU - van Hertem, T.

AU - Steensels, M.

AU - Viazzi, S.

AU - Romanini, C.E.B.

AU - Schlageter Tello, A.A.

AU - Lokhorst, C.

AU - Maltz, E.

AU - Halachmi, I.

AU - Bahr, C.

AU - Berckmans, D.

AU - Hong, S.

PY - 2014

Y1 - 2014

N2 - The objective was to compare automatic lameness detection methods based on daily automatic measurements of cow’s back posture and behavioral and performance variables. The experimental setup was located in a commercial Israeli dairy farm of 1,100 Israeli Holstein cows. All cows were housed in open, roofed cowsheds with dried manure bedding and no stalls. All cows were equipped with a commercial neck activity and ruminating time data logger. Milk yield was measured with a milk flow sensor. Cow gait recordings were made during 4 consecutive nighttime milking sessions with a 3D image camera. From the videos, the “inverse radius” of the back posture contour and the “back posture measurement” were extracted. The reference in this study was a daily live locomotion score of the animals. A dataset of 186 cows with 4 video-based lameness scores and 4 live locomotion scores was built. In total, 38 model input variables were constructed from the sensor data comprising absolute values, relative values, daily standard deviations, slope coefficients, daytime and nighttime periods and milk session related variables. Data of lame cows – cows recognized and treated for lameness – was compared with data of non-lame cows. A logistic regression model was built with the highest correlated behavioral and performance variables. Model validation was done with 10-fold cross-validation. The analysis of the video-based scores as independent observations leads to a correct classification rate of 53.0% on a 5-point level scale. A multinomial logistic regression model based on 4 consecutive “back posture measurement”-scores and “inverse radius”-scores obtained a correct classification rate of 60.8%. Strict binary classification to lame vs. not-lame categories reached 80.7% correct classification rate. In addition, the logistic regression model included 7 model input variables (the daily milk yield; the slope coefficient of the daily milk yield; the nighttime/daytime neck activity ratio; the milk yield week difference ratio; the milk yield week difference; the neck activity level during the daytime; the ruminating time during nighttime). After 10-fold cross-validation, the model obtained a sensitivity of 0.89 and a specificity of 0.85, with a correct classification rate of 0.86 when based on the averaged 10-fold model coefficients. The combination of image processing and behavioral monitoring is believed to further improve the lameness detection accuracy.

AB - The objective was to compare automatic lameness detection methods based on daily automatic measurements of cow’s back posture and behavioral and performance variables. The experimental setup was located in a commercial Israeli dairy farm of 1,100 Israeli Holstein cows. All cows were housed in open, roofed cowsheds with dried manure bedding and no stalls. All cows were equipped with a commercial neck activity and ruminating time data logger. Milk yield was measured with a milk flow sensor. Cow gait recordings were made during 4 consecutive nighttime milking sessions with a 3D image camera. From the videos, the “inverse radius” of the back posture contour and the “back posture measurement” were extracted. The reference in this study was a daily live locomotion score of the animals. A dataset of 186 cows with 4 video-based lameness scores and 4 live locomotion scores was built. In total, 38 model input variables were constructed from the sensor data comprising absolute values, relative values, daily standard deviations, slope coefficients, daytime and nighttime periods and milk session related variables. Data of lame cows – cows recognized and treated for lameness – was compared with data of non-lame cows. A logistic regression model was built with the highest correlated behavioral and performance variables. Model validation was done with 10-fold cross-validation. The analysis of the video-based scores as independent observations leads to a correct classification rate of 53.0% on a 5-point level scale. A multinomial logistic regression model based on 4 consecutive “back posture measurement”-scores and “inverse radius”-scores obtained a correct classification rate of 60.8%. Strict binary classification to lame vs. not-lame categories reached 80.7% correct classification rate. In addition, the logistic regression model included 7 model input variables (the daily milk yield; the slope coefficient of the daily milk yield; the nighttime/daytime neck activity ratio; the milk yield week difference ratio; the milk yield week difference; the neck activity level during the daytime; the ruminating time during nighttime). After 10-fold cross-validation, the model obtained a sensitivity of 0.89 and a specificity of 0.85, with a correct classification rate of 0.86 when based on the averaged 10-fold model coefficients. The combination of image processing and behavioral monitoring is believed to further improve the lameness detection accuracy.

M3 - Abstract

ER -

van Hertem T, Steensels M, Viazzi S, Romanini CEB, Schlageter Tello AA, Lokhorst C et al. Automatic Lameness detechtion by computer vision and behavior and performance sensing. 2014. Abstract from AgEng 2014, Zurich, Switzerland.