Regional crowd flow estimation from aerial view.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Data Science

Background:

• Crowd analysis is crucial for real-world applications.
• Existing aerial crowd flow methods lack flexibility, simplicity, and accuracy, especially with top-down views.
• Clear crowd appearance features are often missing in aerial perspectives.

Purpose of the Study:

• To propose a novel challenge for monitoring large-scale crowd scenes from an aerial view.
• To develop a crowd flow estimation model for accurately measuring flow into and out of specific regions over time.
• To enhance crowd flow estimation by addressing limitations of existing top-down view methods.

Main Methods:

• A two-stream network jointly regresses crowd density and individual velocities.
• A locality-confined attention module enhances flow estimation by considering local relationships.
• Inverse-temporal loss is incorporated for spatio-temporal regularization in top-down views.

Main Results:

• The proposed model accurately estimates instantaneous crowd flow at each location.
• The locality-confined attention module improves flow estimation accuracy.
• The inverse-temporal loss enhances spatio-temporal regularization for aerial views.

Conclusions:

• The developed crowd flow estimation model surpasses prior methods in performance.
• The approach is effective for drone-based overhead crowd videos.
• The model has potential applications in various crowd analysis tasks for understanding social dynamics.