We developed an automatic monitoring system, called MoSBReal, to record the postures of lactating sows kept in crates. The system consists of a Raspberry Pi, protected in a waterproof box, connected to two closed-circuit television (CCTV) cameras to monitor two sows at the same time. The system is easy to install inside a building on a pig farm and to use. The system can be controlled using a smartphone with a direct wireless connection, which does not require Wi-Fi. It can be used for example to set up the camera and check the angle of view. The system uses a convolutional neural network (CNN) to estimate sow postures from the CCTV images at a rate of one estimate every 30 s. A total of eight postures were considered: standing, sitting, kneeling, lying sternal, and lying on right or left side and with the udder exposed or not. The CNN was trained with 89,399 images of 164 different sows and evaluated with 40,206 images of 60 different sows not included in the training set. Using this evaluation procedure, we assessed the capacity of the method to estimate posture from sows never seen by the neural network as it is the case in practice. The overall precision of the prediction and recall were both equal to 92%. The system was installed on 10 farms and used for 10 months on each with no notable difficulties.