A study on deep learning model based on global-local structure for crowd flow prediction.

Area of Science:

• Data Science
• Artificial Intelligence
• Urban Planning

Background:

• Crowd flow prediction is vital for urban planning, business, and public health, especially during pandemics like COVID-19.
• Existing models often overlook the hierarchical structure within crowd data (e.g., by age, gender).

Purpose of the Study:

• To develop a deep learning model that effectively utilizes the global and local structures of crowd flow data.
• To improve the prediction accuracy of crowd flow for various subgroups by considering their hierarchical characteristics.

Main Methods:

• Proposed a deep learning model integrating global crowd flow data with local site-specific data.
• Simultaneously analyzed overall crowd movement and site-specific crowd dynamics.
• Further refined the model by optimizing global data composition based on subgroup correlations.

Main Results:

• The model demonstrated significant improvements in subgroup crowd flow prediction accuracy, ranging from 5.2% to 45.95% compared to related works.
• Further accuracy enhancements of 5.6% to 48.65% were achieved by refining global data composition and excluding low-correlated subgroups.

Conclusions:

• The proposed deep learning approach effectively leverages hierarchical data structures for more accurate crowd flow prediction.
• The findings offer valuable insights for urban planning, resource allocation, and public health strategies, particularly in managing crowd dynamics.