基于注意力和多级特征融合的铁路场景小尺度行人检测算法  

Small-scale Pedestrian Detection Algorithm Based on Attention and Multi-level Feature Fusion for Railway

作　　者：石瑞姣 陈后金[1] 李居朋[1] 李艳凤[1] 李丰 万成凯 SHI Ruijiao;CHEN Houjin;LI Jupeng;LI Yanfeng;LI Feng;WAN Chengkai(School of Electronic and Information Engineering,Beijing Jiaotong University,Beijing 100044,China;Beijing Century Real Technology Co.,Ltd.,Beijing 100085,China)

机构地区：[1]北京交通大学电子信息工程学院,北京100044 [2]北京世纪瑞尔技术股份有限公司,北京100085

出　　处：《铁道学报》2022年第5期76-83,共8页Journal of the China Railway Society

基　　金：山东省重大科技创新工程项目(2019TSLH0206);北京交通大学教育基金会重点项目(9907005501);国家自然科学基金(61872030)。

摘　　要：行人入侵是影响铁路行车安全的重要因素。为有效解决短焦距摄像机在大视场中小尺度行人检测精度低的问题,提出一种注意力机制引导下的多级特征融合网络模型。首先,将YOLOv3作为主干网络,针对多次降采样后行人特征丢失的问题,设计四倍降采样分支以利用高分辨率特征有效提取小尺度行人信息。其次,特征融合阶段引入通道-空间注意力机制以抑制低层特征中背景噪声干扰。最后,引入CIoU损失函数用于行人目标框的回归,解决均方误差损失函数存在的优化不一致及尺度敏感问题。实验结果表明,相较于经典YOLOv3以及现阶段主流目标检测算法,本算法具有更高的检测精度,在自建铁路私有数据集和Caltech公开数据集的各子集上对数平均漏检率均有明显降低。Pedestrian intrusion is an important factor affecting the safety of railway operation.Short focus camera is widely used in pedestrian intrusion monitoring in railway scene.To solve the problem of low-precision detection of small-scale pedestrian detection in large field of view,a multi-level feature fusion network model guided by an attention mechanism was proposed.Firstly,YOLOv3 was adopted as the backbone network.Aiming at the problem of pedestrian feature loss after multiple down sampling,a quadruple down sampling branch was proposed to extract small-scale pedestrian information using high resolution features.Secondly,in the feature fusion stage,channel and spatial attention mechanism was introduced to suppress the interference of background noise in the low level features.Finally,the CIoU loss function was introduced to the bounding box regression to solve the optimization inconsistency and scale sensitivity during the use of the mean square error loss function.The experimental results show that,compared with traditional YOLOv3 and the existing mainstream target detection algorithm,the proposed algorithm has better detection performance.In the private dataset and different subsets of the Caltech Pedestrian Dataset,the logarithmic average missed detection rate is significantly reduced.

关 键 词：铁路行车安全 小尺度行人检测 多级特征融合 通道-空间注意力 CIoU损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]