Development of automatic body condition scoring using a low-cost 3-dimensional Kinect camera

The objective of this study was to assess objectively the ability of observers to assess body condition of dairy cows. Four observers independently assigned a body condition score (five-point scale, .25 increments) and described the appearance of seven body regions of 225 Holstein cows. Areas described were the thurl region, ischial and ileal tuberosities, ilio-sacral and ischio-coccygeal ligaments, transverse processes of the lumbar vertebrae, and spinous processes of the lumbar vertebrae. An absolute body condition score was designated for each cow based on the modal body condition score for all observers. If there was no modal body condition score, the mean score was used for the absolute body condition score. Statistical analysis of principal components was used to examine the relationship between body region description and absolute body condition score. Descriptions of body regions were highly correlated across all absolute body condition scores. Four principal component vectors explained 83.6% of the variation of the body region correlation matrix. The first principal latent vector accounted for 55% of the variation and was uniformly correlated with all body regions. Analysis of variance of first principal latent vector as the dependent variable and absolute body condition score as the class variable indicated that body condition could be separated into .25 units between 2.5 and 4.0, inclusively. Below 2.5 and > 4.0, body condition could only be separated by .5 units. Distinct changes in specific body regions were associated with change in absolute body condition score. Observers agreed with the absolute score 58.1% of the time, deviating by .25 units 32.6% of the time. A body condition score can be given to a cow based on principal descriptors of specific body regions between 2.5 and 4.0 by .25 units.