Pedestrian Flow Prediction Method Based on LSTM

Abstract

With the acceleration of urbanization in China, the urban population density has risen sharply, leading to increasingly severe social security and traffic congestion problems. To address these issues, this study proposes a dynamic and accurate regional pedestrian flow prediction model based on the LSTM neural network, which fuses temporal, statistical and semantic features. Meanwhile, the prediction results are applied to assist relevant departments in pedestrian flow management. Experimental results show that the model achieves a mean squared error (MSE) as low as 29.2530 and a mean absolute error (MAE) of 3.5059, outperforming traditional prediction methods significantly. Additionally, the integration of the SmoothL1Loss function and early stopping mechanism has notably improved the robustness of the model. Error analysis indicates that 83.5% of the prediction errors are controlled within ±20 people, which fully meets the requirements of public safety early warning and traffic congestion prediction. This model provides data support for relevant departments to manage urban pedestrian flow efficiently and accurately, effectively helping the government reduce management and control costs and improve the efficiency of emergency response.