Dairy cows have been selected to produce high milk volumes but production output is frequently limited by lameness, mastitis and other diseases. Disease in its subclinical or clinical stage compromises herd health and welfare and hampers dairy sector sustainability, representing an increase in carbon footprint. In addition, high animal welfare and health practices are more important than ever to satisfy societal demands for food. Spotting the early, subclinical stages of disease in individual cows is hence essential to successful treatment and disease transmission control. Our recent research has found an early sickness-related decline in social exploration and interactivity in cows with subclinical mastitis . A network of 25 CCTV cameras has been set up within our new dairy farm research platform, the John Oldacre Centre at Bristol Vet School  with the goal of providing exquisitely-detailed tracking of individual animal activities, behaviours and interactions. We have developed novel deep learning methods to individually identify cows , laying the groundwork to develop automated video monitoring of social behaviours and from this the prediction of the early stages of disease.
Aims and objectives
Hypothesis: We hypothesise that (a) social behaviour changes are early predictors of subclinical disease; (b) deep activity classifiers operating on video footage can robustly detect this social behaviour, and (c) from this prediction models can be developed to automatically detect early disease from video.
Technical objective: Creation of a multi-camera platform with machine vision classifiers for tracking the activities (loafing, lying, feeding) and social behaviours (face/body collisions, direction and nature of interaction) of the cattle herd for predicting individual-animal early stage disease through time-series prediction models basic on social network analysis.
The student will develop skills in AI and statistical modelling while learning livestock sustainability and welfare - a combination to underpin a future leading role in environmental sustainability in academia, industry or policy work. Our established visual cattle identification work  will be adopted to continually locate and re-identify every cow in real time across our camera network at the John Oldacre Centre. The student will then develop deep activity classifiers using video clips labelled with expert annotations of various behaviours; following human action recognition trends, we utilise spatiotemporal networks and associated self-attention components [4,5]. For prediction modelling, a key facet is that community members in social networks influence each other, causing non-independence that must be accounted for statistically . We will then evaluate if the techniques translate to single-camera setups for translational impact across the dairy sector, while working with farmers, vets and other stakeholders to design a dashboard for communicating herd health to them. A multidisciplinary supervisory team will support the student, which is led by data scientist Prof Andrew Dowsey and supported by machine vision expert Dr Tilo Burghardt, animal bioscientist Dr Suzzane Held, daily scientist Dr Enriquez-Hidalgo, and applied welfare expert Prof Siobhan Mullan.
Apply for this project
This project will be based in Bristol Veterinary School.
Please contact [Email Address Removed] for further details on how to apply.