Exploring density rectification and domain adaption method for crowd counting.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Crowd counting is crucial for security, business, and public health.
• Existing methods struggle with nonuniform density and cross-domain challenges.

Purpose of the Study:

• To develop a robust crowd counting model addressing density variations and domain shifts.
• To improve the accuracy and generalizability of crowd counting systems.

Main Methods:

• Proposed a Density Rectifying Network (DRNet) with Dual-layer Pyramid Fusion Modules (DPFM) for multi-scale feature integration.
• Introduced a Density Rectification Map (DRmap) auxiliary module for adaptive density map correction.
• Developed a domain adaptation technique using randomly cut mixed dual-domain images to learn domain-invariant features.

Main Results:

• DRNet achieved state-of-the-art performance with the lowest Mean Absolute Error (MAE) on four benchmark datasets.
• The domain adaptation method effectively reduced the domain gap, significantly decreasing MAE in cross-domain experiments.
• Demonstrated superior performance compared to existing methods in both standard and cross-domain crowd counting tasks.

Conclusions:

• The proposed DRNet and domain adaptation method offer significant improvements in crowd counting accuracy and robustness.
• These advancements are vital for reliable crowd analysis in diverse real-world applications.
• The study highlights the effectiveness of integrating multi-scale features and adaptive learning for complex crowd counting challenges.