监控管理中基于交叉姿态平滑的行人重识别
Person re-identification based on cross-posture smoothing in surveillance management
陈小慧 1何宜庆2
作者信息
- 1. 南昌大学人文学院,江西南昌 330031
- 2. 南昌大学经济管理学院,江西南昌 330031
- 折叠
摘要
现有行人重识别缺乏对局部上下文信息之间的关联,且由于局部图像中存在许多与行人身份无关的信息,使用与全局图像相同的标签去学习局部信息会产生许多噪声.为缓解以上问题,提出一种基于交叉姿态平滑的行人重识别方法.整个方法分为两个模块,提出一种关键点语义交叉划分策略获取具有交叉语义关联的局部上下文信息;设计一个基于相对姿态上下文的局部伪标签平滑方法,使模型学习到更多局部区域的细微姿态差异.实验结果表明,所提方法在公开数据集上的表现优于其它最新方法,有效提升了模型对行人姿态细微变化的鲁棒性与鉴别能力.
Abstract
Existing person re-identification lacks the correlation between the local contextual information.The same labels are used as the global image to learn the local information generates massive noises,because there is irrelevant information to the pedestrian identity in the local image.To alleviate the above problems,a person re-identification method based on cross-posture smoothing was proposed.The whole method was divided into two modules.A keypoint semantic intersection partitioning strate-gy was presented to obtain local context information with cross semantic associations.A local pseudo label smoothing method based on relative posture context was designed.A local pseudo label smoothing method based on relative posture context was designed,which made the training model learn more subtle information of the local region.Experimental results show that the proposed method outperforms other state-of-the-art methods on the public datasets.The proposed method effectively improves the robustness and discriminative ability of the re-identification model to pedestrian posture changes.
关键词
行人重识别/关键点语义交叉划分策略/相对姿态偏移量/局部伪标签/标签平滑/交叉语义/智能监控管理Key words
person re-identification/keypoint semantic intersection partitioning strategy/relative posture offset/local pseudo la-bel/label smoothing/cross semantics/intelligent surveillance management引用本文复制引用
基金项目
国家自然科学基金项目(72163021)
江西省社会科学"十四五"重点基金项目(21YJ01)
江西省教育科学"十四五"规划重点基金项目(21ZD001)
出版年
2024
NETL
NSTL
万方数据