Video-based analysis of dairy cow behaviour: detection of lying down and standing up

Digital agriculture offers opportunities for improved monitoring and precision phenotyping of farm animals, crucial to achieving a more sustainable livestock production sector. Video-based analysis enables the quantification of animal behaviour in a non-invasive, automated way with few sensors. To unlock its full potential, appropriate computer vision techniques are needed. In this study, we propose an algorithm to detect lying-down and standing-up behaviour in dairy cows based on changes in bounding box properties detected via YOLOv5 and tracked with DeepSORT. We analysed 86 videos with a standing-up or lying-down event. With different criteria applied to the bounding box time series, we could detect up to respectively 92.3 and 80% for standing-up and lying-down events, respectively, with an accuracy of less than 2 seconds. Using bounding box properties as proxy for body shape and location, a general cow detection algorithm can serve multiple behavioural analyses simultaneously, whilst interpretability of the algorithms is maintained.