基于多注意力特征融合的人群计数方法

A Crowd Counting Method Based on Multi-Attention Feature Fusion

摘要: 针对拥挤环境下人群分布的高度不均匀、复杂背景的干扰以及遮挡问题，本文提出一种基于MobileNet V3分类模型的特征融合网络。首先从MobileNet V3网络中提取四个不同尺寸的特征图，并对每个特征图进行HAM模块操作，该模块由通道、边缘、空间注意力以及动态卷积组成，特征图通过上采样对齐分辨率，并在通道维度上拼接成综合特征图，经过1 × 1卷积压缩通道，生成最终的融合特征图用于生成密度图，完成高精度人群计数任务。该方法在ShanghaiTech、NWPU和QNRF三个具有挑战的数据集上进行了实验验证，实验结果表明，所提出的方法在计数精度和鲁棒性方面显著优于现有主流方法。

Abstract: To address the challenges of highly non-uniform crowd distribution, complex background interference, and severe occlusions in crowded environments, this paper proposes a feature fusion network based on the MobileNet V3 classification model. The framework first extracts four multi-scale feature maps from the MobileNet V3 backbone. Each feature map undergoes processing through a Hybrid Attention Module (HAM), which integrates channel attention, edge attention, spatial attention, and dynamic convolution operations. The processed features are then upsampled to align their spatial resolutions, concatenated along the channel dimension, and compressed via a 1 × 1 convolutional layer to generate a unified fused feature map. This fused representation is subsequently used to regress high-precision density maps for accurate crowd counting. The method is experimentally validated on three challenging datasets, namely ShanghaiTech, NWPU and QNRF, and the experimental results show that the proposed method significantly outperforms state-of-the-art approaches in both counting accuracy and robustness.