The Dual-Attention Mechanism-Based Subway Station Crowded Crowds Counting Method

Abstract The subway station is an important place for passenger flow distribution in subway networks, and real-time monitoring of passenger flow within stations helps promote the safe and efficient operation of the entire subway network. However, large-scale crowded passenger flow still exist in subway stations, making it challenging to accurately assess the massive passenger flow inside the stations. In this regard, this paper proposes a crowd counting method based on a deep learning framework with a dual attention mechanism. It aims to tackle the problem of counting large crowds within a subway station. This method provides strong support for ensuring passenger safety in subsequent operations. The key components of our proposed model are the multi-scale attention module and the deformable attention module. The multi-scale attention module can effectively extract multi-scale features of crowds and extract informative features from heavily occluded areas while ensuring that the channel count and resolution of the input feature map remain unchanged. In the Transformer framework, the deformable attention module dynamically assigns attention weights to each feature position, enabling a more suitable crowd counting model for congested conditions in subway station. Three commonly used benchmark datasets for crowd counting were used in a wide range of experiments. The experimental results demonstrate that our model achieves relatively good performance compared to existing popular algorithms. Additionally, existing crowd counting datasets do not adequately capture the variations in multi-scale and crowd occlusion scenarios specific to subway station environments. Therefore, this paper constructed a custom dense crowd dataset for subway station platforms. Our method performs better on this self-built dataset, which focuses on capturing the challenges of multi-scale variations and crowd occlusions in subway station scenarios, as demonstrated by experimental results.