Multi-scale Transformer-Based Crowd Counting Network

In dense crowd scenes, the accuracy of crowd counting is often compromised by complex backgrounds and significant variations in target scales. To address this challenge, this paper proposes a Multi-scale Transformer-based Crowd Counting Network (MSTNet). At the feature extraction stage, we introduce a backbone network of multi-scale Transformer with spatial reduction attention, which effectively captures global contextual information from the image, to address the low counting accuracy caused by complex backgrounds and Multiscale change. A novel method is designed to tackle these challenges by reconstructing features from different levels and fusing local detail with deep semantic of the crowd image. Experimental results on three public datasets demonstrate that our approach achieves higher accuracy in crowd counting compared to current mainstream methods, as well as better robustness in complex backgrounds and Multiscale changes.