Crowd Gathering Detection Method Based on Multi-Scale Feature Fusion and Convolutional Attention.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Public Safety

Background:

• Rapid urbanization leads to more frequent and dense crowd gatherings, challenging public safety management.
• Existing crowd detection methods struggle with occlusions, complex backgrounds, and varying crowd densities due to reliance on local features and fixed scales.

Purpose of the Study:

• To develop an advanced crowd counting framework that overcomes limitations of existing methods.
• To enhance the accuracy and robustness of crowd detection in challenging urban environments.

Main Methods:

• Proposed a point-supervised framework named Multi-Scale Convolutional Attention Network (MSCANet).
• Integrated a context-aware architecture with multi-scale feature extraction and convolutional attention mechanisms.
• Enabled dynamic adaptation to varying crowd densities and focus on critical regions.

Main Results:

• MSCANet demonstrated high counting accuracy and robustness, especially in dense and occluded scenarios.
• The network effectively handles complex scenes and varying crowd scales.
• Achieved superior performance compared to existing methods on public datasets.

Conclusions:

• MSCANet offers a robust solution for crowd counting in complex urban environments.
• The proposed framework shows strong potential for real-world public safety applications.
• Advanced feature representation and attention mechanisms are key to improved crowd detection.