| Abstract |
Assessing how people typically move across public areas like city squares, shopping centers, airports, train stations, metro hubs, etc., is a relevant task required to both perform informed decisions for infrastructure design and to identify unusual behaviors to support surveillance activities. Thanks to the development of computer vision algorithms, methods for detecting and tracking pedestrians offer nowadays valid solutions to tackle this problem. However, most of the state-of-the-art approaches exploit color images that can pose threats to people’s privacy. Additionally, the classification of trajectories to estimate local origin-destination matrices and enable analysis of typical people flow is still poorly addressed. Therefore, in this paper, we propose a complete solution able to detect, track, and classify people flows in complex junctions by exploiting privacy-preserving thermal cameras. Thanks to a thorough examination of several machine learning methods, and to the identification of a new set of features for trajectory classification, also evaluated with eXplainable AI approaches, the proposed solution has been capable to obtain accurate trajectory classification – achieving an F1-score of 0.96 – that can be exploited to compute local origin-destination matrices. Moreover, early classification in real time using partial trajectories is possible. The solution has been implemented exploiting the Snap4City Smart City Digital Twin infrastructure and has been carried out in the context of CN MOST, the Italian National Center on Sustainable Mobility. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026. |
