基于注意力增强的行人与头肩级联检测算法

Pedestrian and Head-Shoulders Cascade Detection Algorithm Based on Attention Enhancement

在监控视频中,行人会存在不同视角、不同程度的遮挡问题,导致当前行人检测器漏检率、错检率较高,于是提出了一种注意力增强下的行人整体与行人头肩区域级联检测的行人检测器。提出一种新的通道注意力机制,称为全卷积通道注意力机制;针对分类和回归任务分别融入相适应的注意力机制,来增强有效的检测特征,抑制背景特征信息;设计行人整体与行人头肩区域级联行人检测器,通过行人整体与行人头肩区域的匹配算法,级联地处理检测结果。该算法,尤其针对下半身严重遮挡的情况,极大降低了遮挡行人的漏检率。实验结果表明,在Caltech公开行人检测测试数据集Reasonable(合理子集)的对数平均漏检率降低到5.37%,尤其在Occ=heavy(严重遮挡子集)上的对数平均漏检率降低到23.33%,同时在ETH和CityPersons行人检测数据集上,该算法亦拥有较好的检测效果。

In the surveillance video,pedestrians may have different perspectives and different degrees of occlusion,resulting in a high rate of missed detection and error detection of the current pedestrian detector.Therefore,a pedestrian detector is proposed for cascading detection of the whole pedestrian and the head and shoulder area of the pedestrian with enhanced attention.Firstly,a new channel attention mechanism called fully convolutional channel attention mechanism is proposed.Secondly,channel attention mechanism and spatial attention mechanism are added respectively for classification and regression tasks to enhance effective detection features and suppress background features.Finally,a cascade pedestrian detector is designed for the pedestrian as a whole and the pedestrian head and shoulder area,and the detection results are cascaded through the matching algorithm of the pedestrian as a whole and the pedestrian head and shoulder area.This algorithm also can greatly weaken the missed detection rate,particularly in the case of serious occlusion of the lower body.Experimental results indicate that the log-average missed rate of Reasonable(reasonable subset)pedestrian detec-tion test dataset in Caltech is reduced to 5.37%,and the log-average missed rate of Occ=heavy(severely obscured subset)is reduced to 23.33%.At the same time,the algorithm also has good detection performance on ETH and CityPersons pedestrian detection datasets.