Comparison of a three-dimensional and two-dimensional camera system for automated measurement of back posture in dairy cows

In this study, two different computer vision techniques to automatically measure the back posture in dairy cows were tested and evaluated. A two-dimensional and a three-dimensional camera system were used to extract the back posture from walking cows, which is one measurement used by experts to discriminate between lame and not lame cows. So far, two-dimensional cameras positioned in side view are used to measure back posture. This method, however, is not always applicable in farm conditions since it can be difficult to be installed. Shadows and continuous changes in the background also render image segmentation difficult and often erroneous. In order to overcome these problems, a new method to extract the back posture by using a three-dimensional camera from top view perspective is presented in this paper. The experiment was conducted in a commercial Israeli dairy farm and a dataset of 273 cows was recorded by both the three-dimensional and two-dimensional cameras. The classifications of both the two-dimensional and the three-dimensional algorithms were evaluated against the visual locomotion scores given by an expert veterinary. The two-dimensional algorithm had an accuracy of 91%, while the three-dimensional algorithm had an accuracy of 90% on the evaluation dataset. These results show that the application of a three-dimensional camera leads to an accuracy comparable to the side view approach and that the top view approach can overcome limitations in terms of automation and processing time.