Metro passenger alighting flow prediction for real-time crowding information: A deep learning approach based on sequential images

With the continued increase in metro ridership, real-time crowding information (RTCI) has become essential for improving travel experience and making better use of carriage space. By accurately predicting the number of passengers alighting from each carriage and estimating available space in real time, it is possible to guide waiting passengers to better boarding choices and reduce congestion. Most existing RTCI studies focus on total passenger flow prediction or instantaneous crowd level estimation, while paying little attention to how alighting behavior of passengers directly affects carriage space availability. To address this issue, this study presents three main contributions. (1) A feature vector is constructed to represent the dynamic state of a carriage, along with a mapping function that describes the evolution of available space, forming a theoretical basis for RTCI generation.(2) A deep learning model based on image sequences is developed by integrating EfficientNet-B0 with LSTM (Long Short-Term Memory), enabling accurate prediction of the number of passengers alighting at the carriage level. (3) Systematic experiments using a dataset from the Changsha metro demonstrate that the method achieves high prediction accuracy while maintaining computational efficiency. Comparative and ablation studies confirm the importance of each core component in improving model performance. The findings offer technical support for optimizing carriage selection and enhancing metro service operations.