基于特征选择方法的无锚框多光谱行人检测

Anchor-free Multispectral Pedestrian Detection Based on Feature Selection Method

为解决低分辨率及遮挡行人目标难以检测的问题,提出一种基于特征选择的无锚框多光谱行人检测方法。该方法以双通道CenterNet网络为基础网络架构,将特征选择注意网络应用于两个卷积模块之间,对语义信息及细节特征进行筛选,最终获取更有利于检测低分辨率及遮挡行人目标的信息。实验结果表明,相比于传统的通道融合多光谱行人检测算法,该方法降低了3%的行人检测漏检率,对于较小的行人目标以及严重遮挡的行人目标分别降低了15%、9%的漏检率,具有一定的应用价值。

In order to solve the problem of low resolution and blocked pedestrian targets,an anchor-free multispectral pedestrian detection method based on feature selection is proposed.This method takes the dual channel CenterNet as the basic network architecture,applies the feature selection attention network between the two convolution modules,filters the semantic information and detailed features,finally obtains the information that is more conducive to detecting low resolution and occluding pedestrian targets.The experimental results show that compared with the traditional channel fusion multispectral pedestrian detection method,this method reduces the miss-rate of pedestrian detection by3%,and reduces the miss-rate of small pedestrian targets and seriously occluded pedestrian targets by 15% and 9% respectively,which has a certain application value.