This article proposes an automatic fall detection method for a wearable device that can promptly alert caregivers when a fall is detected, which could reduce the injuries of elder people. To do this, we propose a novel generative adversarial network (GAN-) based fall detection method using a heart rate sensor and an accelerometer. Acquiring fall data compared with normal behavioral data can be an arduous process. Instead, we introduce a compelling GAN-based anomaly detection partially surrounded with User Initial information features (UI-GAN). Although GAN-based anomaly detection methods have been previously proposed, each model has its adequate suitability for each anomaly detection application. Therefore, this study firstly evaluates suitable GAN-based anomaly detection models for fall detection from among nine recently proposed GAN-based models. From UI-GAN, performance improvements are observed in the fall detection when using the UI information. To objectively demonstrate the competitive performance of UI-GAN, we compare it with eight other recently presented fall detection studies and have observed that it leads to better results. Lastly, since the target application in this study is the use of UI-GAN with a wearable device, the sufficiently satisfied latency of UI-GAN on the smartwatch is estimated. This study is the first attempt to use the initial information of users in GAN, and we hope that the effectiveness of the UI information is expected to be seen in other applications.

In this letter, we propose a multivariate time-series classification system that fuses multirate sensor measurements within the latent space of a deep neural network. In our network, the system identifies the surface category based on audio and inertial measurements generated from the surface impact, each of which has a different sampling rate and resolution in nature. We investigate the feasibility of categorizing ten different everyday surfaces using a proposed convolutional neural network, which is trained in an end-to-end manner. To validate our approach, we developed an embedded system and collected 60,000 data samples under a variety of conditions. The experimental results obtained exhibit a test accuracy for a blind test dataset of 93%, taking less than 300 ms for end-to-end classification in an embedded machine environment. We conclude this letter with a discussion of the results and future direction of research.

People who exercise may benefit or be injured depending on their foot striking (FS) style. In this study, we propose an intelligent system that can recognize subtle differences in FS patterns while walking and running using measurements from a wearable smartwatch device. Although such patterns could be directly measured utilizing pressure distribution of feet while striking on the ground, we instead focused on analyzing hand movements by assuming that striking patterns consequently affect temporal movements of the whole body. The advantage of the proposed approach is that FS patterns can be estimated in a portable and less invasive manner. To this end, first, we developed a wearable system for measuring inertial movements of hands and then conducted an experiment where participants were asked to walk and run while wearing a smartwatch. Second, we trained and tested the captured multivariate time series signals in supervised learning settings. The experimental results obtained demonstrated high and robust classification performances (weighted-average F1 score > 90%) when recent deep neural network models, such as 1D-CNN and GRUs, were employed. We conclude this study with a discussion of potential future work and applications that increase benefits while walking and running properly using the proposed approach.

Various sensing modalities, including external and internal sensors, have been employed in research on human activity recognition (HAR). Among these, internal sensors, particularly wearable technologies, hold significant promise due to their lightweight nature and simplicity. Recently, HAR techniques leveraging wearable biometric signals, such as electrocardiography (ECG) and photoplethysmography (PPG), have been proposed using publicly available datasets. However, to facilitate broader practical applications, a more extensive analysis based on larger databases with cross-subject validation is required. In pursuit of this objective, we initially gathered PPG signals from 40 participants engaged in five common daily activities. Subsequently, we evaluated the feasibility of classifying these activities using deep learning architecture. The model’s performance was assessed in terms of accuracy, precision, recall, and F-1 measure via cross-subject cross-validation (CV). The proposed method successfully distinguished the five activities considered, with an average test accuracy of 95.14%. Furthermore, we recommend an optimal window size based on a comprehensive evaluation of performance relative to the input signal length. These findings confirm the potential for practical HAR applications based on PPG and indicate its prospective extension to various domains, such as healthcare or fitness applications, by concurrently analyzing behavioral and health data through a single biometric signal.

Humans have the extraordinary ability to identify objects by listening to the sound that is produced when hitting the objects’ surface. In this study, we propose an integrated sensor-actuator system that has the capability to identify the surface on which it is placed. The system contacts (i.e., taps or hits) the surface by generating a mechanical impact and then analyzes the resulting sound waveform to decipher the intrinsic characteristics of the surface on which it is placed. Further, a machine learning pipeline based on recent deep learning techniques was introduced to identify ten different everyday surfaces. The results demonstrated that the proposed system can successfully recognize various surfaces with a test accuracy of 91.73%. Furthermore, it was found that the proposed machine learning pipeline can be implemented in an embedded machine with an inference time of less than 300 ms. We conclude this paper with limitations and discussion of future work.