The prototype of automatic head counting, based on deep learning, was developed in the H2020 project FAIR Stations for crow monitoring in stations. However, the approach can be applied wherever there is an interest in assessing how many people are present within the field of view of the used device (e.g. vending machines, shops, bank desks, etc.).
The system can be connected to an already installed surveillance camera, or can be enclosed in a single low-cost module that includes a micro-camera and a PC card for image processing. In this version the device is a "smart sensor" that directly provides the occupancy data.
To better manage crowding and occlusion situations, the system has been trained to count only the heads, regardless of posture and somatic characteristics. In this respect, it differs from similar systems based on the whole person (which require wider visibility) or on the faces (which require them to be oriented towards the camera).
Being completely based on a large number of examples collected in different situations, the system does not need to be programmed in advance, and is immediately operational as soon as it is connected. In some cases, it may be useful to tune the system with new images to ensure optimal performance.