Behavioural information relevant to calving is extracted and exploited successfully for automatic calving prediction from videos. Calving prediction is key for preventing fatal accidents such as stillbirth and dystocia. Such a prediction has been performed using contact sensors to capture a cow's typical movements before calving. However, directly attaching sensors to a cow's body is not desirable from the viewpoint of animal welfare, economic load, and safety for livestock farmers. This paper presents a camera-based, noncontact calving prediction system that captures typical precalving movements such as rotations, turns and step-backs. The information on the frequency of such behaviours and that on the frequency of changes in the behaviours are extracted every few minutes using deep neural networks and used as inputs to a calving predictor based on support vector machines. Experimental comparisons conducted using the videos of four Japanese black beef cows of normal and precalving statuses demonstrated that the system developed with five cows' videos achieved a precision rate of 97% and a recall rate of 82% for a cow that was active before calving.