Multi-scale pyramid fusion with overlap density attention module for crowd counting

Crowd counting is an important research area in computer vision that plays a crucial role in ensuring public safety in crowded environments. Despite significant research progress, accurate crowd counting remains challenging due to severe occlusion, large-scale variations, and overlapping individuals in dense scenes. This paper proposes a novel Multi-Scale Pyramid Fusion with Overlap Density Attention network (MSPF) to address these challenges. The proposed framework consists of three specialized modules: a multi-scale pyramid fusion module, an overlap density attention module, and a features enrichment module. These components collaboratively enhance the model's ability to handle overlapping individuals and scale variations in high-density crowd scenes. Initially, the encoder extracts feature representations at multiple receptive field sizes. The overlap density attention module then emphasizes overlapping and non-overlapping regions by selectively focusing on informative features. Subsequently, the pyramid fusion module adaptively integrates multi-scale features, while the features enrichment module further refines the most relevant spatial information. Finally, the network generates an accurate density map for crowd estimation. Extensive experiments conducted on the proposed Highly-Packed-Crowd dataset and four challenging benchmark datasets demonstrate the superior performance of MSPF in terms of accuracy, robustness, and efficiency compared to state-of-the-art methods.