With the rapid development of the railway industry and the continuous increase of passenger transport tasks, railway passenger stations are facing increasingly severe passenger flow safety issues. To realize real-time monitoring of passenger flow dynamics and finely analyze the multi-granularity characteristics of passenger flow, a Multi-granularity Yardstick for Dynamic Crowds (MYDC) model for railway passenger stations based on video analysis technology is proposed. Firstly, a passenger flow dataset for railway passenger stations is constructed. Secondly, a fine-grained feature perception network for passenger flow is designed based on YOLO and Discriminative Correlation Filter (DCF) tracking algorithm, and the adaptive crowd localization Transformer (CLTR) model for railway passenger stations is improved to capture the coarse-grained features of the overall passenger flow distribution. Finally, based on the physical attributes of passenger flow as well as its micro and macro characteristics, a Multi-Attention Spatio-Temporal Graph Convolutional Network (MASTGCN) is constructed to mine the spatio-temporal dynamic trends of passenger flow and assess the safety risk level of passenger flow in the station. The results show that the cumulative error of fine-grained feature extraction is 6.9%, the recognition accuracy of coarse-grained features is 89.1%, and the recall rate of the passenger flow safety assessment model is 87.5%. The proposed model can provide accurate data support for passenger flow management and has strong engineering application value.