【作者】 杨珂；

【导师】 叶娅兰；

【作者基本信息】 电子科技大学 ， 工程硕士（专业学位）， 2020， 硕士

【摘要】 近年来,随着传感器技术和智能设备的迅速发展,人体行为识别成为重要的研究热点,在医疗监控、智能安防、老年看护以及室内定位具有重要的研究意义和广泛的应用前景。在人体行为识别领域,基于可穿戴设备传感器信号的方法具有便携性好、功耗低、抗环境干扰等优势,更适用于实际应用场景。随着研究深入,目前基于传感器信号的人体行为识别研究在公开数据集上可达到较高效果,但公开数据集是在实验室理想环境下采集的在服务器上进行精度计算,而面向智能可穿戴设备的实际应用场景更加复杂,存在可穿戴用户动作不规范、速度差异、用户之间存在差异性等问题。针对这些问题,本文面向智能可穿戴设备的轻量级计算平台设计了基于残差网络的人体行为识别深度网络模型,并结合超图学习和迁移学习提高了实际应用场景下用户行为识别精确度。论文主要工作如下:1.针对实际可穿戴场景中由于用户动作不规范且存在速度差异导致精确度不高的问题,本文基于轻量级可穿戴计算平台,设计了基于残差网络的行为识别模型。实验结果表明,在USC-HAD公开数据集下比现有的文献提高2.32%,达到92.72%的精确度。在快递员原始数据9分类下达到94.8%的较高精确度。2.在实际应用场景中,用户的个性因素(身高、体重)的差异对行为识别准确率存在影响,本文为了提高模型的普适性,提出将个性因素与传感器数据融合在一起,并利用超图学习进行行为识别的方法。实验结果表明,在USC-HAD公开数据集下,比现有的文献提高3.1%,达到93.3%的精确度。在快递员数据提取的60维特征的基础下,加入个性因素后提升3.95%,达到90.5%的精确度。3.针对实际场景中由于用户差异性导致测试新用户数据集精确度不高的问题,本文为了提高模型的泛化性能,首先设计了半监督的跨用户行为识别模型,实验结果表明,在快递员数据集跨用户下,识别精确度由74.47%提升到94.38%。其次本文提出了基于余弦加权的无监督域适应方法CORAL用于跨用户的人体行为识别,实验结果表明,在快递员数据集跨用户下,识别精确度提升5.09%,达到86.77%的精确度,在WARD公开数据集上,本文提出的方法比现有的文献提升6.65%,达到90.94%的行为识别精确度。在此基础上,将本文设计的人体行为识别模型应用到计算快递员劳动量的实际应用中,基于层次分析法设计一个计算快递员劳动量模型,并设计和实现面向快递员劳动量计算系统,该系统智能化的工作模式有助于降低快递员离职率。更多还原

【Abstract】 In recent years,with the rapid development of sensor technology and smart devices,human behavior recognition has become an important research hotspot,which has important research significance and wide application prospects in medical monitoring,intelligent security,elderly care and indoor positioning.In the field of human behavior recognition,the method based on the sensor signal of the wearable device has the advantages of good portability,low power consumption,anti-environmental interference,etc.,and is more suitable for practical application scenarios.With the deepening of research,the current human behavior recognition research based on sensor signals can achieve higher results on public data sets,but the public data sets are collected in the ideal environment of the laboratory and the accuracy calculation is performed on the server.The actual application scenarios of wearable devices are more complicated,and there are problems such as irregular movement of wearable users,differences in speed,and differences between users.To solve these problems,this thesis designs a deep network model of human behavior recognition based on the residual network for the lightweight computing platform of smart wearable devices,and combines Hypergraph Learning and Transfer Learning to improve the accuracy of user behavior recognition in actual application scenarios.The main work of the thesis is as follows:1.In order to solve the problem of low accuracy due to non-standard user actions and speed differences in actual wearable scenes,this thesis designs a behavior recognition model based on a ResNet based on a lightweight wearable computing platform.The experimental results show that,under the USC-HAD public data set,it is2.32% higher than the existing literature,reaching an accuracy of 92.72%.It achieves a higher accuracy of 94.8% under the 9 categories of raw data of couriers.2.In actual application scenarios,differences in user’s personality factors(height,weight)have an impact on the accuracy of behavior recognition.In order to improve the universality of the model,this thesis proposes to integrate personality factors with sensor data and use super Hypergraph Learning method for behavior recognition.The experimental results show that,under the USC-HAD public data set,it is 3.1% higher than the existing literature and achieves an accuracy of 93.3%.Based on the60-dimensional features extracted by the courier’s data,the personality factor isincreased by 3.95% to achieve 90.5% accuracy.3.In order to improve the generalization performance of the model,the semi-supervised cross-user behavior recognition model was first designed for the problem of low accuracy in testing new user data sets due to user differences in actual scenarios.The experimental results show that in express under the cross-user data set,the recognition accuracy increased from 74.47% to 94.38%.Secondly,this thesis proposes the unsupervised domain adaptation method CORAL based on cosine weighting for cross-user human behavior recognition.The experimental results show that the recognition accuracy of the courier data set across users increases by 5.09% to86.77%.On the WARD public data set,the method proposed in this thesis is improved by 6.65% compared to the existing literature,and achieves 90.94% accuracy of behavior recognition.On this basis,the human behavior recognition model designed in this thesis is applied to the actual application of calculating the labor volume of couriers,a model for calculating the labor volume of couriers is designed based on the Analytic Hierarchy Process,and a system for calculating the labor volume of couriers is designed and implemented.The system’s intelligent working mode helps reduce the rate of courier turnover.更多还原