Center for MicroElectroMechanical Systems (CMEMS), University of Minho, 4800-058 Guimarães, Portugal.
LABBELS-Associate Laboratory, 4710-057 Braga, Portugal.
Sensors (Basel). 2022 May 26;22(11):4028. doi: 10.3390/s22114028.
The recognition of Activities of Daily Living (ADL) has been a widely debated topic, with applications in a vast range of fields. ADL recognition can be accomplished by processing data from wearable sensors, specially located at the lower trunk, which appears to be a suitable option in uncontrolled environments. Several authors have addressed ADL recognition using Artificial Intelligence (AI)-based algorithms, obtaining encouraging results. However, the number of ADL recognized by these algorithms is still limited, rarely focusing on transitional activities, and without addressing falls. Furthermore, the small amount of data used and the lack of information regarding validation processes are other drawbacks found in the literature. To overcome these drawbacks, a total of nine public and private datasets were merged in order to gather a large amount of data to improve the robustness of several ADL recognition algorithms. Furthermore, an AI-based framework was developed in this manuscript to perform a comparative analysis of several ADL Machine Learning (ML)-based classifiers. Feature selection algorithms were used to extract only the relevant features from the dataset's lower trunk inertial data. For the recognition of 20 different ADL and falls, results have shown that the best performance was obtained with the K-NN classifier with the first 85 features ranked by Relief-F (98.22% accuracy). However, Ensemble Learning classifier with the first 65 features ranked by Principal Component Analysis (PCA) presented 96.53% overall accuracy while maintaining a lower classification time per window (0.039 ms), showing a higher potential for its usage in real-time scenarios in the future. Deep Learning algorithms were also tested. Despite its outcomes not being as good as in the prior procedure, their potential was also demonstrated (overall accuracy of 92.55% for Bidirectional Long Short-Term Memory (LSTM) Neural Network), indicating that they could be a valid option in the future.
日常生活活动（ADL）的识别一直是一个备受争议的话题，其应用领域非常广泛。ADL 识别可以通过处理来自可穿戴传感器的数据来完成，这些传感器专门位于下躯干，这在不受控制的环境中似乎是一个合适的选择。许多作者已经使用基于人工智能（AI）的算法来解决 ADL 识别问题，取得了令人鼓舞的结果。然而，这些算法识别的 ADL 数量仍然有限，很少关注过渡活动，也没有解决跌倒问题。此外，文献中还发现了数据量少和缺乏验证过程信息等其他缺点。为了克服这些缺点，总共合并了九个公共和私人数据集，以收集大量数据来提高几种 ADL 识别算法的鲁棒性。此外，本文还开发了一个基于 AI 的框架，以对几种基于机器学习（ML）的 ADL 分类器进行比较分析。特征选择算法用于从数据集的下躯干惯性数据中提取仅相关的特征。对于 20 种不同的 ADL 和跌倒的识别，结果表明，使用 Relief-F 排名前 85 位特征的 K-最近邻（K-NN）分类器的性能最佳（准确率为 98.22%）。然而，使用 PCA 排名前 65 位特征的集成学习分类器的整体准确率为 96.53%，同时保持每个窗口的分类时间较低（0.039 毫秒），显示出在未来实时场景中具有更高的应用潜力。还测试了深度学习算法。尽管其结果不如前一种方法好，但也证明了它们的潜力（双向长短期记忆（LSTM）神经网络的整体准确率为 92.55%），表明它们在未来可能是一个有效的选择。
