LSC-CNN

LSC-CNN implements a detection-based convolutional neural network that locates individual heads, estimates head sizes with bounding boxes, and counts people in dense crowds for accurate localization and counting.

Key Features:

• Detection Over Regression: Shifts from density regression to a detection framework that localizes individuals and produces bounding boxes.
• Multi-Column Architecture: Employs a multi-column architecture with top-down feature modulation to enhance resolution and deliver refined multi-resolution predictions.
• Head Localization and Sizing: Detects heads across sparse to extremely dense crowds and estimates approximate head size via bounding boxes.
• Training with Minimal Annotations: Trains using only point head annotations while inferring size information for bounding boxes.
• Superior Performance: Empirical evaluations report improved localization accuracy and counting precision compared to density regressors.

Scientific Applications:

• Public Safety: Provides precise localizations and counts to inform crowd control and public safety measures.
• Urban Planning: Informs urban planning and public-space design through accurate crowd quantification.
• Event Management: Supports monitoring and capacity planning at large events to accommodate gatherings safely.
• Surveillance and Crowd Control: Enhances surveillance systems and crowd-control strategies by supplying localization and count information rather than only density estimates.

Methodology:

Uses a detection framework implemented as a multi-column CNN with top-down feature modulation to predict head locations and bounding-box sizes, and is trained using point head annotations to infer size information instead of relying on density regression.