基于腕部穿戴传感器移位的行为识别补偿研究  

Study on Compensation for Human Activity Recognition Based on Wrist-wearing Sensor Displacement

作　　者：汪鑫 路成钢 耿颖睿 王燕[1] 温晓许 杨红梅 WANG Xin;LU Chenggang;GENG Yingrui;WANG Yan;WEN Xiaoxu;YANG Hongmei(School of Electric and Information,Zhongyuan University of Technology,Zhengzhou 450007,China;Zhongyuan-Petersburg Aviation College,Zhongyuan University of Technology,Zhengzhou 450007,China)

机构地区：[1]中原工学院电子信息学院,郑州450007 [2]中原工学院中原彼得堡航空学院,郑州450007

出　　处：《赣南师范大学学报》2023年第3期50-54,共5页Journal of Gannan Normal University

基　　金：河南省科技攻关项目(222102210016)。

摘　　要：穿戴人体行为识别的研究结果通常基于设定好的穿戴位置.而传感器移位因为佩戴松动或佩戴习惯不同等问题普遍存在.若不对移位进行相应补偿,行为模型的识别率可能明显下降.文章针对腕部传感器移位问题,建立腕部传感器移位数据集,构建多阶段注意力识别模型,提出局部穿戴位置数据混合补偿和跨位置微调迁移补偿2种方案.实验结果表明2种补偿方法都不同程度地改善了传感器移位对行为识别率的影响,降低了模型训练成本.The results from studies on wearable human activity recognition are usually reported based on the predefined wearing positions.Sensor displacement,caused by the loose wearing during use or users'specific wearing habits,can lead to the performance deterioration of models.This paper focuses on the wrist-wearing sensor displacement and the compensation for it.We thus build a sensor displacement dataset,construct a multi-stage attention recognition model,and propose two compensation schemes,i.e.,mixing data from multiple local positions and the cross-position fine-tuning transfer learning.The experimental results show that both schemes can improve the reduced recognition accuracies caused by sensor displacement.Also,the transfer learning scheme can reduce the model training cost.

关 键 词：穿戴 传感器移位 多阶段注意力 补偿 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]