Human activity recognition (HAR) using smartphones provides significant healthcare guidance for telemedicine and long-term treatment. Machine learning and deep learning (DL) techniques are widely utilized for the scientific study of the statistical models of human behaviors. However, the performance of existing HAR platforms is limited by complex physical activity. In this article, we proposed an adaptive recognition and real-time monitoring system for human activities (Ada-HAR), which is expected to identify more human motions in dynamic situations. The Ada-HAR framework introduces an unsupervised online learning algorithm that is independent of the number of class constraints. Furthermore, the adopted hierarchical clustering and classification algorithms label and classify 12 activities (five dynamics, six statics, and a series of transitions) autonomously. Finally, practical experiments have been performed to validate the effectiveness and robustness of the proposed algorithms. Compared with the methods mentioned in the literature, the results show that the DL-based classifier obtains a higher recognition rate ($\text{95.15}\%$, waist, and $\text{92.20}\%$, pocket). The decision-tree-based classifier is the fastest method for modal evolution. Finally, the Ada-HAR system can monitor human activity in real time, regardless of the direction of the smartphone.

中文翻译：

---

基于智能手机的人类活动自适应识别和实时监控系统

使用智能手机的人类活动识别 (HAR) 为远程医疗和长期治疗提供了重要的医疗保健指导。机器学习和深度学习 (DL) 技术被广泛用于人类行为统计模型的科学研究。然而，现有 HAR 平台的性能受到复杂的身体活动的限制。在本文中，我们提出了一种自适应人体活动识别和实时监控系统（Ada-HAR），有望在动态情况下识别更多人体运动。Ada-HAR 框架引入了一种独立于类约束数量的无监督在线学习算法。此外，采用的层次聚类和分类算法标记和分类 12 个活动（五个动态，六个静态，和一系列转换）自主。最后，通过实际实验验证了所提出算法的有效性和鲁棒性。与文献中提到的方法相比，结果表明基于DL的分类器获得了更高的识别率（$\text{95.15}\%$，腰部，$\text{92.20}\%$，口袋）。基于决策树的分类器是模态进化最快的方法。最后，无论智能手机的方向如何，Ada-HAR 系统都可以实时监控人类活动。和 $\text{92.20}\%$, 口袋）。基于决策树的分类器是模态进化最快的方法。最后，无论智能手机的方向如何，Ada-HAR 系统都可以实时监控人类活动。和 $\text{92.20}\%$, 口袋）。基于决策树的分类器是模态进化最快的方法。最后，无论智能手机的方向如何，Ada-HAR 系统都可以实时监控人类活动。