Smart gravimetric system based on Deep Learning for enhanced safety of accesses to public places

This paper describes the architecture and the performances of a smart robust gravimetric system for access monitoring in public places. The system is composed of a gravimetric footboard, equipped with load cells, an embedded electronics, and a VGG16 convolutional neural network based classification algorithm. The main objective is to spot irregularities, which correspond in this case to the contemporary presence on the footboard of more than one person, trying to enter in the place under observation using only one ticket. Moreover, other notable events can be recognized, such as the passage of a person with a stroller or a trolley, or a person with some injuries or disabilities on crutches or on a wheelchair. This additional information can be useful for enhancing the access safety and for the statistical characterization of the monitored place frequentation. The results are very promising, showing accuracies well over 90% for all the considered event categories. In particular, the overall 100% accuracy in the detection of irregular events (subdivided into queued people and side-by-side people categories) is truly remarkable and confirms that the proposed system can be reliably employed in real world applications.