Behavioural indicators of emotion are extremely useful for animals’ emotions’ assessment as animals under minimal or no stress should be able to display their natural behaviour patterns. Considering the challenges of behavioural observation, this study aimed to implement a skeleton-based Action Recognition model to detect and classify five behaviours in a group of 24 broiler chicks: Sitting Alert, Standing Alert, Preening, Exploration Pecks, and Feeding Pecks, which added up to 2687 total instances labelled across 144 videos.

A three-stage approach was used for action detection and classification, where object detection (1) and pose estimation (2) models were trained to identify individual chicks and their key body points and subsequently detect and classify the chick’s behaviours (3). A significant aspect of the study was evaluating how various approaches to data pre-processing affected the model's performance.