Prediction of daily activity of gestating sows from manual video analysis periods Measuring sows' physical activity is important for predicting their nutritional requirements and can be performed using accelerometers or manual video analysis. However, accelerometers are expensive, invasive, and do not last long, while video analysis is time-consuming and cannot be done continuously. Therefore, the aims of this project were 1) to study the correlation between the activity measured continuously by accelerometers and the one estimated by video analysis over short periods, and 2) to determine whether 24 h of activity could be predicted from these video data. The data set contained individual activity measurements, from accelerometers (recorded continuously) and video analysis (two periods of 5 h per day), taken over 24 days from 38 group-housed gestating sows. Statistical analyses were performed using a generalized linear model considering the observer of the videos (5), the sow's group (4), and the hierarchical rank. Prediction of accelerometer activity using data from the video analysis explained 23-64% of the variability for the two 5 h periods per day, and 39-46% for a combined period of 10 h (P < 0.001). For the prediction of 24 h of activity, the explained variability ranged from 26-44%, depending on the data set used. These results indicate that the prediction of 24 h of activity is possible but needs to be improved by reducing the observer effect and perhaps by extending the duration of the video analysis.