Crowd Counting Method Based on Dual-Path Multi-Feature Fusion with Transformer

In dense crowd scenes, the scale of human heads varies greatly, and current visual transformer models lack learning of different scales, resulting in the model being unable to perceive the target scale changes of human heads in crowd density images, leading to a decrease in crowd counting accuracy. To address this issue, we added a Multi Path Computing Network (MPCNet) in the feature extraction stage, applying attention mechanisms at different sampling rates and levels to capture hierarchical features of the image, and fitting proximity and multi-level dependencies to effectively perceive changes in the head scale. Adding a bidirectional feature fusion module (BIFFM) to the backend network captures more complex feature dependencies by simultaneously considering two different spatial positions, effectively enhancing the interaction between feature maps of different scales and enabling the model to perform counting efficiently. Finally, this chapter conducted extensive experiments on three mainstream population counting benchmark datasets to validate the effectiveness of the method.