Deep learning allows the automation of video recorded pig behaviour assessment.

Animal behaviour studies require the use of a precise ethogram. Cameras can be used to collect data continuously, allowing ad libitum sampling. However, watching and measuring all the recordings afterwards is time-consuming. This work aims to propose an automated recognition pipeline for a wide range of pig behaviours including both social and regular ones. A pig detection dataset was first built by extracting 1000 frames from a 3-month video recording and labelling pigs. Next, a YOLOv8-based object detector was trained to detect pigs and reached a mean average precision of 0.993. A pig behaviour recognition dataset was then built to cover a total of 11 social and non-social regular behaviours (i.e., fighting, mounting, nosing, social playing, exploring, standing, laying, eating, drinking, playing, and walking). Annotations were made on the basis of YOLOv8 predictions on consecutive frames extracted from ten 5-min video clips at an interval of 1 s. Targeted short clips were complemented to deal with class imbalance. Finally, a SlowFast-based spatiotemporal convolutional network was trained to reach a detection probability of 0.9. Deep learning-based recognition technology can thus be used to automate behavioural measurements on pigs.