From movement to meaning: Bridging computer vision and behaviour research to investigate activity patterns and social interactions in pigs

This thesis investigated the potential of computer vision algorithms for monitoring social interactions and activity patterns. Proximity networks derived from a tacking-by-detection algorithm were built to examine how methodological choices affect network metrics. The results demonstrated that network metrics were sensitive to both the definition of proximity and sampling rate. Dyadic models were used to characterise variation in proximity and damaging behaviours, revealing substantial variation between individuals and between pairs. Two approaches were explored for identifying early signs of damaging behaviours.  First, higher activity levels were associated with higher tail-biting frequency on following days, suggesting that activity may help identify potential future tail biters. However, activity was also influenced by environmental conditions, which should be considered when interpreting activity data.
Second, a video-based behaviour recognition model was developed and achieved 72% accuracy in detecting ear and tail biting. Overall, this thesis provides novel approaches for quantifying and analysing pig behaviour using computer vision, while highlighting key methodological considerations for their application.