Locate, Size, and Count: Accurately Resolving People in Dense Crowds via Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current crowd counting models primarily use density regression, which often lacks precise person localization.
• This limitation hinders their applicability in scenarios requiring accurate individual identification beyond mere counts.

Purpose of the Study:

• To introduce a novel detection framework for dense crowd counting that overcomes the limitations of density regression.
• To develop a model capable of accurately detecting and localizing individuals in diverse crowd densities.

Main Methods:

• Proposed a detection-based framework, LSC-CNN, that locates each person, sizes their head with bounding boxes, and then counts them.
• Employed a multi-column architecture with top-down feature modulation for enhanced resolution and person distinction.
• Utilized a training regime requiring only point head annotation, enabling estimation of head size.

Main Results:

• LSC-CNN demonstrates superior localization accuracy compared to traditional density regression methods.
• The model achieves state-of-the-art performance in crowd counting tasks across various crowd densities.
• Successfully addresses challenges unique to dense crowd detection, such as high person diversity and contiguous box prediction.

Conclusions:

• The LSC-CNN framework provides a robust alternative to density regression for crowd counting.
• The detection-based approach offers improved localization and counting accuracy, making it suitable for a wider range of applications.
• The model's ability to work with minimal annotation (point head) further enhances its practical utility.