*Result*: Efficiently recording of sow’s activity patterns through embedded computer vision
*Further Information*
*International audience ; The assessment of a sow’s maternal behavior relies on manual annotations which are time consuming and are susceptible to an observer’s bias. This maternal behavior and particularly the level of activity of a sow is linked to piglet growth and to preweaning mortality. The aim of this study was to provide an overview of the data collected with the MoSBReal device, a user-friendly tool that automatically records changes in sow postures during their stay in the farrowing unit. Thirty devices (60 cameras) were installed in 10 selection farms to collect images from Landrace and Large-White sows. In total, 168,589 images coming from 275 sows were collected over 8 postures: kneeling, standing, sitting, lying on their left or right side with or without showing their udder and ventrally. A classification model, based on YOLO v11, was trained on 154,780 images. The algorithm was tested on 13,809 new images and had an overall precision of 91.2%, from 83% to 98% depending on the posture. Using this model, the activity of 13 sows was recorded between the third day before farrowing till the tenth day following it. The activity profile of those sows was studied. The number of daily posture changes was peaking for nine sows on the day of farrowing and the day before for the other four. When looking at the time spent in each posture, we found that on the day of farrowing sows spent 53.8% of the time with the udder showing. After farrowing, sows spent significantly more time in those postures (81.45% on the next day). Through this sample some diversity of activity profile has been observed. By studying the activity profile of our 275 sows we could find a new way to improve maternal abilities and weaning performances with it.*