XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
Automatic dairy cow body condition scoring
using depth images and 3D surface fitting
Wen-Yong Li
Dept. Intelligent System
National Engineering Research Center
for Information Technology in
Agriculture
Beijing, China
liwy@nercita.org.cn
Zhan-Kui Yang
Dept. Computer Science
Beijing University of Technology
Beijing, China
a437112007@163.com
Yang Shen
Dept. Electronic Information and
Automation
Tianjin University of Science &
Technology
Tianjin, China
master0808@126.com
Xin-Ting Yang
Dept. Intelligent System
National Engineering Research Center
for Information Technology in
Agriculture
Beijing, China
yangxt@nercita.org.cn
Du-Jin Wang
Dept. Electronic Information and
Automation
Tianjin University of Science &
Technology
Tianjin, China
1458298431@qq.com
Abstract—Automatic and objective dairy cow body
condition scoring has received considerable attention as a tool to
aid in the management of nutritional programs in dairy herds.
This paper presents a 3-dimensional algorithm that provides a
topographical understanding of the cow’s body to estimate BCS.
The hypothesis tested was that the body shape of a fatter cow is
rounder than that of a thin cow and, therefore, may better fit a
paraboloid surface. Image processing and regression algorithms
were developed and included the following steps: (1) object
recognition and separation, identification and separation of the
cows; (2) image surface fitting; and (3) parameters
determination in BCS model. All steps were performed
automatically, including image acquisition and model training.
The novelty in this study compared to the previous ones was
completing the full-automation of the system. The model was
implemented and its outputs were validated against manual
body condition scoring (BCS). Pearson correlation between the
proposed BCS and the manual BCS was 0.84 for the test data
set.
Keywords—body condition scoring (BCS), computer vision,
dairy cows, precision livestock farming
I. I
NTRODUCTION
Body condition scoring (BCS) is known as a herd
technique for determining the energy balance of dairy cows
and to estimate their fatness or thinness. Evaluation of BCS on
dairy farms has implications for milk yield, herd health,
reproductive performance, animal well-being, and overall
farm profitability [1,2]. The most popular method for
evaluating BCS is the 5-point scale [3,4], with 1 representing
emaciated cows and 5 representing obese cows. Currently,
BCS is accomplished by visual and tactile assessment of a cow
by a trained evaluator. Consequently, the scoring is expensive,
prone to subjectivity and cannot be performed on a regular
basis due to the time commitment required. Therefore, there is
a need to develop methods to determine the BCS of dairy cows
in an automatic way, which would be more objective, cost
effective and easy to connect with data from a herd
management system.
Several attempts to automate BCS to achieve a more
objective and less time-consuming procedure are reported in
the literature [2, 4-9]. Reference [5] identified 23 anatomical
points manually on images captured automatically as cows
passed through a weigh station and two primary BCS systems
(UKBCS and USBCS) were used to measure the BCS for
lactating dairy cows. Reference [6] reported a model based on
thermal camera and image processing for evaluation of cows’
body reserve but it didn’t achieve full automation since the
these frames were manually examined to select the best one
from each cow. In order to address the problem of modeling
the shape of a cow’s body to build a robust descriptor for
automatic BCS estimation, [2] developed a technique to
model the body shape of a cow from which learned parameters
could be used in BCS estimation. However, in the anatomical
points labeling, users need to use a mouse to point to a region
and right clicks on the mouse to record the location. In pursuit
of full-automation BCS, some groups made an important step
forward.
Reference [7] extracted the contour of tailhead area from
2D image and used Fourier descriptors to describe the cow’s
signature which is a 1-dimensional vector of the Euclidean
distances from each point in the contour to the shape center.
The feature set seemed to be an efficient descriptor for
automatic BCS prediction. However, (1) the images were
selected manually in the image acquisition step, and (2) the
model developed in the research showed weaker results in
terms of R
2
(0.77). Reference [8] improved their previous
research [6] to reach the full-automation assessment of BCS
for dairy cattle by using thermal imaging and the best frame
selection method. However, the model [8] was established
using a thermal camera. Implementation of a thermal camera
is relatively expensive [3] and its output is easily influenced
by the temperature of its surrounding environment [10].
Recent advances in 3D imaging techniques have enabled
possibilities of accurately and efficiently obtaining 3D
structures of objects for measuring geometric traits. Using 3D
instead of 2D-image data has the advantage that not only the
contours but also changes of the cows' surfaces can be
examined in detail. Reference [11] estimated the backfat
thickness using extracted traits from an automatic 3D optical
system in lactating Holstein-Friesian cows. In that study, an
expensive TOF camera was used to acquire the animal’s rear
area images. In addition, several cuts through a cow’s surface