Development and Validation of an Automated Video Tracking Model for Stabled Horses

Abstract

Changes in behaviour are often caused by painful conditions. Therefore, the assessment of behaviour is important for the recognition of pain, but also for the assessment of quality of life. Automated detection of movement and the behaviour of a horse in the box stall should represent a significant advancement. In this study, videos of horses in an animal hospital were recorded using an action camera and a time-lapse mode. These videos were processed using the convolutional neural network Loopy for automated prediction of body parts. Development of the model was carried out in several steps, including annotation of the key points, training of the network to generate the model and checking the model for its accuracy. The key points nose, withers and tail are detected with a sensitivity of more than 80% and an error rate between 2 and 7%, depending on the key point. By means of a case study, the possibility of further analysis with the acquired data was investigated. The results will significantly improve the pain recognition of horses and will help to develop algorithms for the automated recognition of behaviour using machine learning.