On farm implementation of a fully automatic computer vision system for monitoring gait related measures in dairy cows

The objective of this study was to implement a computer vision system for automatic monitoring of animal based measures relevant for lameness detection in a commercial dairy farm. The implementation procedure comprised the following steps: (1) start and stop of the video recordings, (2) identification of the cow in the video, and (3) video processing including the filtering of good quality images and the calculation of the back posture parameters used for classifying cows as lame or not lame. After implementation, the performance of the system was evaluated. All data were gathered from a Belgian commercial dairy farm. Between 20 September 2013 and 30 March 2014, 323 recording sessions were performed, together with 33 locomotion scoring events spread over time. The first step after recording the videos was identifying the cows in the video, which was successful for 79.2% ± 6.2% of the milked cows. In the second step of the video processing where the lameness related feature variables are extracted from the images, obtained an average analysis rate of 49.9% ± 11.3%. On average 80%) of the individual cows were at least 5 times per week automatically scored. Based on 3130 complete cow observations spread over time, a group level analysis was performed in the form of a receiver operating characteristics curve. The back posture measure (BPM) and ⊖₂ were the two feature variables that reached the level of a fair measure for lameness detection.