Towards a Holistic Assessment of Animal Welfare using Emotion and Deep Learning

Understanding animal emotions is an important part of animal welfare. Humans can recognise emotions by synthesising information from facial expressions, body posture, and movements. This ‘holistic approach’ has also been applied to the observation of animals, e.g. combining various body movements and -parts, or applying Qualitative Behavioural Assessment (QBA). However, which observable features in an animal's appearance and movements are used and which combinations of these are most relevant in perceiving differing states of valence and arousal remains unknown. Furthermore, until now the mentioned behavioural observations are very time consuming and therefore application in a commercial context is limited. Therefore, our project aims to explore a novel spatiotemporal form of supervised machine learning for holistic assessment of animal wellbeing by interpretation of body language, which is informed by work in human activity and emotion recognition using AI. Computer vision-based machine learning techniques will be applied, in which models will be trained using many examples of (individual) pig body language when experiencing known differing emotional states (as ground truth), e.g. positive emotions during feeding or negative emotions when in an unfamiliar environment. Over time, it should be possible to correlate the same or a similar animal's stance and/or type of movement with a particular valence and level of arousal. As a part of the model development, eye tracking during experimental observations will allow to explore aspects of human perception of animal body language. The model will be applied and tested under various other situations on-farm, in groups of animals, various husbandry and age situations. Automated detection of body postures and movements offers the possibility for extending welfare monitoring beyond human time limitations to provide monitoring of large numbers of animals. Such a system could both complement welfare monitoring (e.g. on-farm, on abattoirs, experimental situations).