作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (7): 220-226. doi: 10.19678/j.issn.1000-3428.0061884

• 图形图像处理 • 上一篇    下一篇

基于连接注意力的行人重识别特征提取方法

魏紫薇1,2, 屈丹2, 柳聪2   

  1. 1. 郑州大学 软件学院, 郑州 450001;
    2. 战略支援部队信息工程大学 信息系统工程学院, 郑州 450001
  • 收稿日期:2021-06-09 修回日期:2021-08-30 出版日期:2022-07-15 发布日期:2021-09-07
  • 作者简介:魏紫薇(1996—),女,硕士研究生,主研方向为人工智能、智能信息处理、行人重识别;屈丹,教授、博士生导师;柳聪,硕士研究生。
  • 基金资助:
    国家自然科学基金(61673395,62171470);郑州市重大科技攻关项目(188PCXZX773)。

Method for Person Re-identification Feature Extraction Based on Connected Attention

WEI Ziwei1,2, QU Dan2, LIU Cong2   

  1. 1. Software College, Zhengzhou University, Zhengzhou 450001, China;
    2. School of Information System Engineering, Strategic Support Force Information Engineering University, Zhengzhou 450001, China
  • Received:2021-06-09 Revised:2021-08-30 Online:2022-07-15 Published:2021-09-07

摘要: 全民安全意识的逐步提高使得智能监控设备遍布各大公共场所,行人重识别作为视频分析的关键技术之一,被广泛应用于智能安防、自动驾驶等领域。为了提高真实环境下跨摄像头行人检索的识别精度,提出一种基于注意力机制的行人重识别特征提取方法。在数据处理阶段,考虑不同训练数据量下识别效果存在差异的问题,对行人图片采用自动增强方法进行数据增强,以提高数据集规模。在特征提取阶段,将连接注意力模块与ResNet50残差网络相结合构成特征提取网络,提取显著性更强的行人特征。在损失优化阶段,采用三元组损失和圆损失对行人特征进行优化并完成距离度量,最终根据距离的远近得到行人排序结果。实验结果表明,在Market1501数据集上该方法的Rank-1值和mAP值分别达到95.90%和89.66%,在DukeMTMC-reID数据集上Rank-1值和mAP值分别达到91.16%和81.24%,在MSMT17数据集上Rank-1值和mAP值分别达到84.37%和62.73%,与现有经典行人重识别方法PCB、MGN、Pyramid、OSNet等相比,其网络识别性能评价指标均有明显提升。

关键词: 行人重识别, 深度学习, 卷积神经网络, 注意力机制, 自动增强

Abstract: The gradual increase in national safety awareness has promoted the spread of intelligent monitoring equipment across major public places.Person re-identification, a key technique for video analysis, is widely used in intelligent security, automatic driving, and other fields.A method for person re-identification feature extraction based on the attention mechanism is proposed to improve the identification accuracy of cross camera person retrieval in real environments.During the data-processing stage, the automatic enhancement method was adopted for person pictures to enhance the data set size, considering the difference in the identification effect under different amounts of training data.During the feature extraction stage, the connection attention module and ResNet50 residual network were combined to form a feature extraction network to extract more significant person features.In the loss optimization stage, triple and circle losses were used to optimize the person characteristics and complete the distance measurement.Finally, the person ranking results were obtained according to the distance.The experimental results showed that the Rank-1 and mAP values of this method on the Market1501 data set were 95.90% and 89.66%, respectively.The Rank-1 and mAP values on the DukeMTMC-reID data set were 91.16% and 81.24%, respectively, and those on the MSMT17 data set were 84.37% and 62.73%, respectively.Compared to existing classical person re-identification methods(PCB, MGN, Pyramid, and OSNet), the network identification performance evaluation indexes improved significantly.

Key words: person re-identification, deep learning, Convolutional Neural Network(CNN), attention mechanism, automatic enhancement

中图分类号: