摘要
行人区域遮挡是影响行人检测精度的重要因素,为提高行人检测精度,提出一种基于多尺度注意力引导的遮挡行人检测方法。首先,设计一种专注于遮挡问题的注意力引导模块,将其应用于特征提取网络中不同尺度的特征图,利用外部监督信息遮罩机制,引导模型关注行人目标可见区域;其次,根据特征图在分辨率与语义特征方面的特性,对注意力引导后的多尺度特征图进行融合;最后,利用融合特征图进行边界框预测。为验证所提方法有效性及泛化性,在不同数据集上进行仿真实验。该方法在Citypersons重度遮挡子集上实现了47.1%的MR^(-2),在Caltech重度遮挡子集上实现了40.62%的MR^(-2),相对于主流的遮挡行人检测方法,检测精度有较为明显的提高。实验结果表明,所提出模块可以有效地处理行人检测中的区域遮挡问题。
Pedestrian area occlusion is an important factor affecting the accuracy of pedestrian detection.In order to improve the accuracy of pedestrian detection,a multi-scale attention guidance-based occlusion pedestrian detection method is proposed.Firstly,an attention guidance module is designed to focus on the occlusion problem,and is applied to feature maps of different scales in the feature extraction network,and the external supervision information masking mechanism is used to guide the model to focus on the visible area of the pedestrian target.Secondly,according to the feature map of resolution and semantic features,the attention-guided multi-scale feature map is fused.Finally,the fusion feature map is used to predict the bounding box.In order to verify the effectiveness and generalization of the proposed method,simulation experiments are carried out on different data sets.This method achieves 47.1%of MR^(-2)on the Citypersons heavily occluded subset and 40.62%of MR^(-2)on the Caltech heavily occluded subset.Compared with mainstream occluded pedestrian detection methods,the detection accuracy is significantly improved.Experimental results show that the proposed module can effectively deal with the area occlusion problem in pedestrian detection.
作者
谢东军
刘志刚
黄朝
田枫
刘苗苗
XIE Dongjun;LIU Zhigang;HUANG Zhao;TIAN Feng;LIU Miaomiao(School of Computer&Information Technology,Northeast Petroleum University,Daqing 163318;Postdoctoral Workstation on Institute of Applied Technology,Northeast Petroleum University,Daqing 163318)
出处
《计算机与数字工程》
2022年第5期983-988,1028,共7页
Computer & Digital Engineering
基金
国家自然科学基金项目(编号:61502094)
黑龙江省自然科学基金项目(编号:LH2020F003,LH2019F042)
黑龙江省优秀中青年科研创新团队项目(编号:KYCXTD201903)资助。
关键词
遮挡行人
目标检测
注意力
多尺度
对数平均漏检率
occluded pedestrian
object detection
attention
multi-scale
log-average miss rate
