Change Detection and Adaptation Strategies for Long-Term Estimation of Pedestrian Flows

The sharp growth of the global urban population entails a series of challenges for city councils towards enhancing the quality and safety of pedestrian infrastructures. Among them, an accurate identification of streets that are prone to experiencing congestion might improve the design and management of urban spaces and assets. However, the behavior of pedestrian flows is not exempt from environmental and social aspects that impact on its evolution over time. As a result, new circumstantial factors may alter the pedestrian profiles and ultimately, degrade the quality of flow estimations. This work presents a novel long-term pedestrian flow estimation model capable of adapting its knowledge to alleviate the aforementioned degradation of its estimations due to circumstantial changes. For this purpose, changes are detected by our approach based on heuristic rules that depend on the performance of the estimation model over time. Adaptation is then triggered reactively once a change is declared. We assess the performance of the proposed model over a real-world pedestrian flow dataset. Our experiments reveal that the proposed change detection and adaptation framework resiliently guarantees a stable quality of the estimations over time, paving the way towards its utilization in other mobility scenarios.