Changes in animal’s behaviour may be good indicators of health and welfare variations. However, human observation is time-consuming and labour-intensive. Development of video technology and image processing is a non-invasive method which may offer the opportunity for a better prevention by detecting behavioural and welfare issues continuously and automatically and therefore at an early stage. We are developing deep learning algorithms to analyse routinely the behaviour of young bulls. In the present work, performances of algorithms developed to automatically detect the different activities of bulls on images are evaluated. Bulls originating from 2 different breeds (Limousine, 6 bulls/pen; Charolais, 13±1 bulls/pen) were housed accordingly to the standard management conditions of their respective stations (Pôle de Lanaud, Ferme des Etablières). Two cameras 2 D colour were installed above each pen with different angular views. Nine postures (standing, lying) and behaviours (eating, drinking, moving, autogrooming, fighting, standing up and lying down) were labelled on 1,108 images extracted from the videos. Annotations are evenly distributed with an average of 123 sequences per type of posture or activity and a standard deviation of 37.0. This annotated set of images was used to train the algorithm, an object detection model that uses convolutional neural networks to detect and classify objects in an image. Preliminary training of the algorithm with 419 standing bulls and 373 lying bulls’ pictures is promising with 88% sensitivity and 79% precision. Complementary results of algorithm’s performances will be presented by valuing the full dataset. This project BeBoP will contribute to the current need for on-farm, operational behavioural welfare indicators that can be easily used to assess not only the individual welfare but also the welfare of the whole group.