基于级联乘性融合的夜间多尺度行人检测网络

Nighttime Multi-scale Pedestrian Detection Network Based on Cascaded Multiplicative Fusion

摘要:

针对夜间低光照环境下行人检测面临的多尺度特征捕获困难和夜间背景噪声干扰等问题，提出了一种双流乘性融合网络DMFNet；该网络结合CMF模块和DCA模块，有效解决了夜间场景中行人尺度差异与夜间噪声的关键问题；CMF模块采用自顶向下的乘性特征融合机制，通过元素级乘法操作实现了跨尺度信息的高效整合，有效抑制了背景噪声干扰；DCA模块采用一维卷积和全连接层构建双路径注意力机制，通过可学习的混合因子自适应调整特征通道权重，实现了对行人区域特征的精准增强；实验结果表明，DMFNet在NightOwls数据集上的性能超越现有主流方法，Reasonable子集漏检率为6.85%，检测精度更高且鲁棒性更强，能够有效检测夜间复杂场景中的行人目标。

Abstract:

To address the challenges of multi-scale feature capture and severe background noise interference in nighttime low-light pedestrian detection, this paper proposes a DMFNet. This network integrates a CMF module and a DCA module to effectively mitigate the issues of pedestrian scale variation and nighttime noise. Specifically, the CMF module adopts a top-down multiplicative mechanism, which achieves efficient integration of cross-scale information through element-wise multiplication, thereby effectively suppressing background noise. Meanwhile, the DCA module constructs a dual-path attention mechanism utilizing both one-dimensional convolution and fully connected layers. By employing a learnable mixing factor to adaptively adjust feature channel weights, the DCA module realizes precise feature enhancement in pedestrian regions. Experimental results on the NightOwls dataset demonstrate that DMFNet outperforms existing mainstream methods, achieving a Miss Rate of 6.85% on the Reasonable subset. The proposed method exhibits superior detection accuracy and robustness, proving its effectiveness in detecting pedestrians within complex nighttime scenes.