Enhancing Crowd Counting Efficiency Using Hybrid Attention-based Multi-column CNN

Abstract

Accurately estimating crowd sizes is crucial for numerous applications such as crowd management, urban planning, and public safety. However, accurately estimating crowd density remains a challenge due to factors like scene variations and occlusions. Exist- ing methods often struggle with generalization and sensitivity to these factors. This paper proposes a novel crowd-counting architecture that leverages hybrid attention mechanisms. By incorporating hybrid attention into a multi-scale CNN, our goal is to significantly improve the model’s adaptability to diverse crowd scenes and enhance counting accuracy. The primary contribution of our research is the incorporation of a spatial attention and channel module within the CNN architecture. This module allows the model to dynamically concentrate on essential areas within the image, effectively mitigating the impact of background clutter and occlusions. This targeted focus leads to more precise crowd density estimation and improved counting performance, especially in complex scenarios. Our experimental results show a substantial enhancement in counting accuracy compared to existing methods. Additionally, the proposed architecture achieves this improvement with a notable reduction in training time. This advancement represents a significant stride towards more robust and efficient crowd-counting methodologies, paving the way for further advancements in this field.