Comparative analysis of methods for automated detection of pig behaviour from visual imaging

The development of AI-based systems for monitoring animal behaviour is constrained by the scarcity of labelled data, particularly in real farm environments where behavioural patterns are complex and uncertain. This study systematically compares alternative learning frameworks for detecting feeding and drinking behaviours in pigs from video data collected under commercial farm conditions. Salient changes in these behaviours have been shown to relate to health and welfare challenges in these species. We first evaluate a YOLO-based object detector trained directly on video frames to identify behavioural events. We then assess feature-based video analytics approaches that employ features derived from tracklets and bounding boxes as behavioural evidence, comparing conventional machine learning classifiers with a novel Dempster-Shafer evidential reasoning framework. The latter integrates uncertain evidence to mitigate the impact of limited annotated data. Results show that, under data scarcity and uncertainty, the evidential reasoning approach achieves superior performance in behaviour detection compared with classical machine learning models, while the object detector exhibits limited generalisation. On our realistic farm dataset, Dempster-Shafer evidential reasoning achieved a combined behaviour detection F1-score of 0.87, surpassing the Support Vector Machine F1-score of 0.83 and a YOLO-based object detector’s F1-score of 0.70. These findings suggest that Dempster-Shafer combination-based methods offer a robust alternative for AI-driven analysis of animal behaviour in variable farm environments, thereby enhancing the scalability and reliability of precision livestock monitoring systems.