江伟忠,余海铭,陈世峰.基于特征点相关性的行人重识别方法[J].集成技术,2020,9(4):68-78
基于特征点相关性的行人重识别方法
Person Re-identification Method Based on Correlation Between Features
  
DOI:10.12146/j.issn.2095-3135.20200519001
中文关键词:  行人重识别;特征相关性;全局信息
英文关键词:person re-identification; correlation between features; contextual information
基金项目:广东省重点领域研发计划项目( 2 0 1 9B0 1 0 1 5 5 0 0 3 );国家自然科学基金项目(U1 7 1 3 2 0 3 );深圳市孔雀技术创新项目 (KQJSCX20180330170238897);中科院科研仪器设备研制项目(YJKYYQ20190028)
作者单位
江伟忠 中国科学院深圳先进技术研究院 深圳 518055;中国科学院大学深圳先进技术学院 深圳 518055 
余海铭 爱奇艺公司 北京 100010 
陈世峰 中国科学院深圳先进技术研究院 深圳 518055 
摘要点击次数: 187
全文下载次数: 232
中文摘要:
      目前行人重识别算法面临的主要问题包括背景过多、行人区域缺失及图片视角差异等。基 于行人区域中显著性特征之间存在着强相关性及行人区域与背景区域特征之间存在着弱相关性两方 面的观察,该研究提出一种基于特征点相关性的行人重识别方法。其中,通过采用一种基于视觉不 变性与弱检测的上下文信息处理模块,即 CIP(Contextual Information Processing)模块实现该方法。 由于具有强相关性的特征可能分布在不同的粒度之间,所以嵌入 CIP 模块的多粒度行人重识别方法 可以融合粒度之间的相关性信息。实验中,通过以第一配准率(Rank-1)和平均准确率为评价指标, 使用单数据集测试、跨数据集测试两种方法,在 4 个大型的行人重识别数据集上进行了充分的测试 实验,并利用两个可视化的方法——弱检测效果与行人区域中显著特征点的相关性效果,对 CIP 模 块的有效性进行验证。实验结果显示,目前主流的多粒度网络通过嵌入 CIP 模块,性能均有明显的 提升。
英文摘要:
      Person re-identification suffers from some problems such as confusion with excessive background, incomplete targets, and different viewing point etc. There are two basic observations for most person re-identification applications, i.e., strong correlation exists between the discriminative features, and weak correlation exist between feature points of the pedestrian areas and the background areas. Based on such observations, a person re-identification method based on features correlation is proposed in this paper. The CIP (contextual information processing) module based on viewpoint invariance and soft-detection is applied to realize the proposed method. Since strong correlative features distribute at different granularities, the multi-granularity based person re-identification methods can describe relationship between granularities by embedding the CIP module. The experiments are implemented on four large-scale person re-identification data sets. Both single-domain and cross-domain tests are used in the experiments. The Rank-1 and mean average error criterion are used as the evaluation indicators. As the experiment shows, the proposed method enhances the identification performance of several mainstream multi-granularity methods by CIP module.
查看全文  查看/发表评论  下载PDF阅读器
关闭
微信关注二维码 用微信扫一扫 用微信扫一扫 用微信扫一扫

美女

美女图片

美女

美女图片