Information Entropy-Guided Multi-Scale Feature Fusion for Crowd Density Estimation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Crowd density estimation faces challenges due to spatial heterogeneity, leading to occlusion and feature ambiguity in dense areas.
• Sparse regions and backgrounds have lower informational complexity, complicating accurate density assessment.
• Existing methods struggle to adapt computational focus to varying crowd densities.

Purpose of the Study:

• To develop an entropy-inspired crowd density estimation framework that adaptively allocates computational attention.
• To introduce a novel network, DGCC-Net, leveraging a Density-Guided Map (DGMap) for improved density differentiation.
• To enhance the accuracy of crowd counting in scenarios with significant spatial heterogeneity.

Main Methods:

• Proposed an entropy-inspired framework allocating attention proportional to local information complexity.
• Developed a Density-Guided Map (DGMap) using nearest-neighbor statistics as an entropy proxy.
• Introduced DGCC-Net with a Twins-Transformer backbone, Local Attention Module (LAM), Multi-Level Feature Fusion Module (MLFM), and Density Guidance Module (DGM).

Main Results:

• DGCC-Net achieved competitive or state-of-the-art performance on four benchmark datasets (ShanghaiTech PartA, UCF-QNRF, UCF_CC_50, JHU-Crowd++).
• The entropy-inspired attention allocation effectively addressed challenges in heterogeneous crowd distributions.
• DGMap successfully differentiated between dense, sparse, and isolated pedestrian regions.

Conclusions:

• The proposed entropy-inspired attention allocation is effective for crowd density estimation in heterogeneous scenarios.
• DGCC-Net demonstrates superior performance, validating the benefits of density-adaptive feature refinement.
• This approach offers a promising direction for improving crowd counting accuracy and robustness.