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

The spatial heterogeneity of crowd distributions poses significant challenges for density estimation. Dense regions exhibit high local information entropy due to severe occlusion and feature ambiguity, while sparse regions and backgrounds carry progressively lower informational complexity. To address this, we propose an entropy-inspired crowd density estimation framework that allocates computational attention in proportion to the local information complexity of crowd regions. A Density-Guided Map (DGMap), constructed from nearest-neighbor distance statistics of head annotations, serves as a proxy for local information entropy, enabling the model to differentiate among dense, sparse, and isolated pedestrian regions. The proposed network, termed DGCC-Net, comprises four components: a Twins-Transformer backbone for hierarchical feature extraction, a Local Attention Module (LAM) that enhances high-resolution features through multi-scale receptive fields and rotational attention, a Multi-Level Feature Fusion Module (MLFM) with cross-scale dense connectivity and learnable branch weights for integrating semantic and spatial information, and a Density Guidance Module (DGM) supervised by the entropy-inspired DGMap to achieve density-adaptive feature refinement. Extensive experiments on four benchmark datasets (ShanghaiTech PartA, UCF-QNRF, UCF_CC_50, and JHU-Crowd++) demonstrate that DGCC-Net achieves competitive or state-of-the-art performance, validating the effectiveness of entropy-inspired attention allocation in heterogeneous crowd scenarios.