基于ResNet50和通道注意力机制的行人多属性协同识别方法  

Pedestrian Multi-Attribute Collaborative Recognition Method Based on ResNet50 and Channel Attention Mechanism

作　　者：卓力 袁帅[1] 李嘉锋 ZHUO Li;YUAN Shuai;LI Jia-feng(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Beijing Key Laboratory of Computational Intelligence and Intelligent System,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]北京工业大学计算智能与智能系统北京市重点实验室,北京100124

出　　处：《测控技术》2022年第8期1-8,15,共9页Measurement & Control Technology

基　　金：北京市自然科学基金-丰台轨道交通前沿研究联合基金(L211017);北京市教育委员会科技计划一般项目(KM202110005027)。

摘　　要：针对目前行人多属性识别算法识别精度不高的问题,利用行人属性之间的内在关联关系,基于“特征提取+回归”的思想,提出了一种基于深度卷积神经网络的行人多属性协同识别方法。该方法首先对多个行人属性标签进行组合编码,得到一个标签组合向量;然后采用基于通道注意力机制的ResNet50作为主干网络提取行人图像的深度特征;最后,设计了一个包含3个全连接层的神经网络结构来建立标签组合向量与行人深度特征之间的映射模型,在一个统一的网络框架下就可以同时对行人的多种属性进行准确识别。在行人属性公共数据集PETA和RAP上的实验结果表明,采用提出的识别方法在公共数据集PETA上获得的识别准确率为84.08%,而在公共数据集RAP上可以获得高达88.12%的识别准确率。In order to solve the problem of low recognition accuracy of current pedestrian multi-attribute recognition algorithms,by using the intrinsic relationship among pedestrian attributes,a pedestrian multi-attribute collaborative recognition method of the pedestrian based on deep convolutional neural network is proposed,which adopts the basic framework of“feature extraction+regression”.Firstly,pedestrian attribute labels are encoded to obtain a label combination vector.Then,ResNet50 based on channel attention mechanism is used as the backbone network to extract the deep features of the pedestrian image.Finally,a neural network structure with three fully connected layers is designed to establish the mapping model between the label combination vector and the deep features of pedestrians.In a unified network framework,multiple attributes of pedestrians can be recognized accurately at the same time.Experimental results on public datasets of PETA and RAP show that,the recognition accuracy of the proposed method can reach 84.08%on PETA and 88.12%on RAP.

关 键 词：深度学习 ResNet50 通道注意力机制 多属性识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]