用于位置信息辅助复杂人体行为识别的新型深度学习框架

A Novel Deep Learning Framework for Location Information Assisted Complex Human Activity Recognition

随着近年来智能生活理念的普及和可穿戴终端技术的快速发展,基于传感器数据的人体行为识别(human activity recognition,HAR)已引起广泛关注,并且具有重要的学术研究和商业应用价值。该文研究了增强HAR模型对用户日常简单行为(simple activity,SA)和复杂行为(complex activity,CA)的识别,并提出了一个深度学习(deep learning,DL)模型。首先,使用两个可公开获取的数据集UCI HAR和Shoaib CHA,并对其进行标准化处理。其次,使用所提出的模型提取各种动作的特征,进行人体行为识别。鉴于用户行为和位置之间的高度关联,通过独热编码技术将位置信息集成到数据集中,以提高模型的分类性能。此外,将所提出的模型与8种经典机器学习(machine learning,ML)算法和6种DL算法进行了对比。最后,评估了不同行为类型对HAR模型识别性能的影响。实验结果表明,所提出的模型在UCI HAR和Shoaib CHA数据集上的最高分类准确率分别达到了96.77%和99.13%。通过向数据集添加位置信息,HAR模型对SA和CA的分类准确率得到了显著提高。

With the popularization of smart living and the rapid development of wearable terminal technology in recent years,sensor-based human activity recognition(HAR)has attracted widespread attention and has significant academic research and commercial application value.This paper focuses on enhancing the HAR model’s recognition of users’daily simple activities(SAs)and complex activities(CAs),and proposes a deep learning(DL)model.Firstly,two publicly available datasets,UCI HAR and Shoaib CHA,are normalized.Then the characteristics of distinct activities are retrieved by the proposed model for HAR.Given the high association between users’activities and locations,location information is integrated into the dataset by the one-hot encoding technique to boost the model’s classification performance.In addition,the proposed DL model is evaluated against eight traditional machine learning(ML)algorithms and six DL algorithms.Finally,the effect of various types of activities on the HAR model’s recognition ability is studied.The experimental findings reveal that the proposed model achieves the highest classification accuracy on UCI HAR and Shoaib CHA datasets,with 96.77%and 99.13%,respectively.The classification accuracy of the HAR model is also greatly enhanced for both SAs and CAs by adding location information to the datasets.