基于小样本和随机化的跨域人体动作泛化识别模型

Cross domain human action generalization recognition model based on small sample size and randomization

下载PDF

导出

摘要随着IEEE 802.11bf标准的发布,WiFi感知技术已从学术研究走向工业应用。针对现有人体动作识别在域内能够准确感知,但面对跨域场景时模型识别性能差的问题,提出了一种基于小样本和随机化的跨域人体动作泛化识别模型SSRCD-Fi。首先,使用特征提取器将输入样本映射到向量空间,实现同一动作的样本聚集、不同动作的样本分离;然后针对新的场景域,通过随机化方法和少量被标记样本计算出动作的原型表示;最后,计算查询样本与动作原型之间的距离,从而实现了人体动作的分类。实验结果和分析表明,SSRCD-Fi能够实现鲁棒的跨域人体动作的泛化感知,在不可见的用户和位置上实验准确率分别为92.73%和97.99%。实验代码公开在:https://github.com/4three2one/SSRCD-Fi。 The release of the IEEE 802.11bf standard propels WiFi sensing technology from academic research to industrial applications.Existing human action recognition models perform well within a single domain,but their recognition accuracy significantly deteriorates in cross-domain scenarios.To address this issue,this paper proposed a small sample and randomized cross-domain human action generalization recognition model called SSRCD-Fi to achieve cross-domain human action generalization perception in WiFi.The SSRCD-Fi model operated in several steps.Firstly,it used a feature extractor to map input samples to a vector space,achieved sample aggregation for the same action and sample separation for different actions.Then,for a new scene domain,randomization and a small number of labeled samples could calculate the prototype representation of the action.Finally,it calculated the distance between the query sample and the action prototype achieves the classification of human actions.The experimental results and analysis show that SSRCD-Fi can achieve robust cross domain generalization perception of human actions,with accuracies of 92.73%and 97.99%for unseen users and locations,respectively.The experimental code of this article is publicly available at:https://github.com/4three2one/SSRCD-Fi.

作者胡明许佳炜赵立军王杨欧阳少雄后海伦 Hu Ming;Xu Jiawei;Zhao Lijun;Wang Yang;Ouyang Shaoxiong;Hou Hailun(School of Network Engineering,Wuhu Institute of Technology,Wuhu Anhui 241002,China;School of Computer&Information,Anhui Normal University,Wuhu Anhui 241000,China;Yangtze River Delta Region Hart Robotics Industry Technology Research Institute,Wuhu Anhui 241000,China)

机构地区芜湖职业技术学院网络工程学院安徽师范大学计算机与信息学院长三角哈特机器人产业技术研究院

出处《计算机应用研究》北大核心 2025年第3期849-855,共7页 Application Research of Computers

基金国家自然科学基金资助项目(61871412) 安徽省教育厅自然科学基金重点资助项目(KJ2021A1314) 安徽省中青年教师培养行动项目(JNFX2023116) 机器视觉检测安徽省重点实验室资助项目(KLMVI-2023-HIT-11)。

关键词人体动作识别小样本随机化跨域泛化感知 human motion recognition small sample randomization cross domain generalization perception

分类号 TP391.41 [自动化与计算机技术—计算机应用技术]