researchmap

Publishing type：Research paper (international conference proceedings)   Publisher：ACM International Conference Proceeding Series  

In this demo, we will show our proposed system that can count nodding without either a camera or any sensor attached to the person. Our proposed system capitalizes on the fact that the upper body moves in conjunction with nodding and that this body motion slightly shakes the chair. We explore the challenge of recognizing nodding from the extremely subtle sway of a chair. To recognize nods in real-Time, we employed a supervised learning approach using acceleration data from sensors attached to the chair's backrest. Ultimately, the Support Vector Machine (SVM) achieved a nodding recognition accuracy of 0.990. Further testing of the accuracy of nodding frequency measurements yielded an accuracy of 0.947, suggesting that the optimal position for the accelerometer is the backrest. These results indicate that simply placing the accelerometer on the backrest can effectively quantify the frequency of nods from seated participants.

DOI： 10.1145/3627050.3630740

Scopus

researchmap

Language：Others  

Language：Others  

Language：Others  

Language：Others   Publisher：IEEE Access  

The recent advancements in sensing technology have opened up the possibilities for various services that support daily life, such as energy-saving home appliance control. To realize such services, accurate and cost-effective daily living activity recognition in a wide range is essential. To actualize such a system, it is imperative to address the following requirements: the acquisition of sensors entails very high costs (Issue 1), it is hard to achieve precise recognition for location-independent activities like reading a book (Issue 2), a burden of wearing devices from the perspective of residents (Issue 3), and the preservation of residents' privacy is compromised by using image data from the camera (Issue 4). In this paper, we propose a method for recognizing daily living activities utilizing Doppler sensors in a relatively longer detection range than other motion detection sensors that can be used for dynamic objects. We assess the proposed system by optimizing recognition accuracy, evaluating ensemble methods, and examining sensor reduction's impact. In the first assessment, the logistic regression achieved the highest accuracy of 65.99% in the leave-one-person-out cross-validation. The second assessment revealed an accuracy of 59.39% for the parallel activity recognition method and 57.24% for the joint recognition method of location and activity. In the third assessment, logistic regression achieved a recognition accuracy of 65.26% when four sensor nodes were used: two sensors were placed on both sides of a participant, another was diagonally behind the participant, and the other was installed on the ceiling.

DOI： 10.1109/ACCESS.2023.3330895

Web of Science

Scopus

researchmap

Language：Others  

researchmap

researchmap

researchmap

researchmap

Language：Others   Publisher：IEEE Access  

Quadriplegics face a communication obstacle as their physical abilities are restricted, leaving them unable to speak or use their limbs, with only their upper neck being mobile. So, we propose a recognition system and a new communication language utilizing Morse code and head movements, to break this barrier. We aim to overcome the limitations of camera-based and wearable-sensor methods, including occlusion, privacy concerns, and user inconvenience. The goal is to passively detect quadriplegics' head movements and map them to their corresponding character. The dataset including all 26 alphabet letters, was gathered in various settings, including single-user and multi-human environments, with multiple locations for each setting. For evaluation, 2% samples are randomly selected from the unseen environment to be used with the seen environment as a training dataset. Based on the results, our system demonstrates practical feasibility for real-world implementation, with accuracy rates of 94% and 80% achieved in single-user and multi-human environments, respectively.

DOI： 10.1109/ACCESS.2023.3326259

Web of Science

Scopus

researchmap

Language：Others   Publishing type：Research paper (other academic)   Publisher：IEEE Wireless Communications and Networking Conference, WCNC  

Recently, the studies of wireless device-free human sensing technology have dramatically advanced with enabling a variety of applications, from activity recognition to vital sign monitoring. In this paper, we propose Wi-Nod which leverages the Wi-Fi Channel State Information (CSI) to detect head nodding gestures for each Morse code symbol based on time-frequency features for accurate recognition accuracy in multi-human context environment. The system consists of three basic modules: data collection, data preprocessing, and learning part based on the inception model. The model was trained to perform the head movement detection based on the CSI spectrogram collected by the ESP32 nodes. We evaluated the performance of the system on four different data sets collected in two different sessions. Our system achieves over 95% recognition accuracy that reveals the feasibility of Wi-Nod system for real-life deployment.

DOI： 10.1109/WCNC55385.2023.10118666

Web of Science

Scopus

researchmap

Publisher：2023 14th International Conference on Mobile Computing and Ubiquitous Network, ICMU 2023  

Nowadays more and more people are wearing smart-watches in their daily lives. The various sensors embedded in smartwatches bring the ability to evaluate users' status as well as the risk of privacy issues. For example, if users are typing on key-boards while wearing smartwatches, the attacker could know the typing contents from the sensor data collected by the malicious applications that are installed on the targets' smartwatches. In this paper, we propose WatchLogger, the framework using audio and accelerometer signals to recognize the English words being typed, for demonstrating how to implement the smartwatch-based side-channel attack. Different from the previous studies that focused on the recognition of each key or pair of keys being pressed, WatchLogger aims to perform recognition on the scale of words. To achieve this goal, WatchLogger exploits the audio signals for segmentation and the accelerometer signals for classification. In addition, we propose an ensemble classification model to deal with the problem caused by too many words. At last, we build the dataset WTW-100 with 100 classes of words and 100 samples for each class, and we conduct experiments on the dataset. The experimental results show an accuracy of 98.5 % for keystroke recognition and 91.5 % for word classification, showing a considerable performance of WatchLogger.

DOI： 10.23919/ICMU58504.2023.10412218

Scopus

Web of Science

Publisher：2023 14th International Conference on Mobile Computing and Ubiquitous Network, ICMU 2023  

Nowadays, there is a growing demand to understand the mental well-being of office workers, driven by increased awareness of its impact on productivity and the need for healthier work environments. Recently, the use of Wi-Fi channel state information (CSI) for activity recognition has received significant attention due to its wide availability and privacy protection. In this paper, we propose a passive desk body gesture recognition system that utilizes Wi-Fi CSI from an ESP32 toolkit to automatically detect the worker's mood and emotions. The system is designed to operate within the Internet of Things ecosystem, employing a low-energy device to collect and compress CSI measurements, resulting in improved energy efficiency and cost-effectiveness. The proposed system demonstrates high recognition accuracy of over 98 % in-session and 72 % out-session evaluations.

DOI： 10.23919/ICMU58504.2023.10412222

Scopus

Web of Science

Publisher：2023 14th International Conference on Mobile Computing and Ubiquitous Network, ICMU 2023  

In recent years, many studies explore CSI-based Wi-Fi sensing produce high accurate results in device-free localization, human activity recognition. Channel state information (CSI) represents how wireless signals propagate from transmitter to receiver and provides rich information to identify human presence. By collecting CSI from the specific devices, the machine learning model can be trained to recognize the human activity. However, in multi-room residential settings where walls and furniture obstruct signals, effective coverage with a limited number of transceivers becomes a crucial challenge, underlining the importance of their optimal placement. In this paper, we deployed multiple transceivers in a smart home environment in our university and studied transceiver arrangement patterns for CSI-based indoor localization and activity recognition. We compared the accuracy of localization and activity recognition for a total of 18 patterns consisting of up to five transceivers. In the activity recognition, we used Support Vector Machine (SVM) to classify whether or not a person is moving. The results show that the pattern using only one pair of transceivers achieved an accuracy 85% and covered the entire house. Meanwhile, in the localization we used Light Gradient Boosting Machine (LGBM) to classify which room the person is. The results show that accu-racy decreases as the number of devices is reduced. Therefore, we investigated deployment pattern to achieve accuracy with smaller number of transceivers.

DOI： 10.23919/ICMU58504.2023.10412230

Scopus

Web of Science

Publisher：2023 14th International Conference on Mobile Computing and Ubiquitous Network, ICMU 2023  

In this paper, we propose a method to measure human head motion, especially nodding, without attaching any sensors to the person. Our proposed system focuses on the fact that the upper body moves along with nodding and that the body motion slightly shakes the chair. We challenge the problem of whether it is possible to recognize a nodding from the extremely slight sway of a chair. To reveal the optimal position of sensors, we collected data by attaching multiple accelerometers to various positions on a chair, including the backrest, the seat's underside, and the legs. Using a supervised learning approach, we determined the best positions and combinations of sensors for recognizing nodding more collectively. The Support Vector Machine (SVM) achieved a nodding recognition accuracy of 0.990. Further testing of the accuracy of nodding frequency measurements resulted in an accuracy of 0.947, suggesting that the best position for the accelerometer is the backrest. These results suggest that simply placing the accelerometer on the backrest can effectively quantify the nod frequency of seated participants.

DOI： 10.23919/ICMU58504.2023.10412249

Scopus

Web of Science

Language：Others   Publishing type：Research paper (scientific journal)   Publisher：IEEE Access  

Wireless sensing represented by WiFi channel state information (CSI) is now enabling various fields of applications such as person identification, human activity recognition, occupancy detection, localization, and crowd estimation these days. So far, those fields are mostly considered as separate topics in WiFi CSI-based methods, on the contrary, some camera and vision-based crowd estimation systems intuitively estimate both crowd size and location at the same time. Our work is inspired by the idea that WiFi CSI also may be able to perform the same as the camera does. In this paper, we construct Wi-CaL, a simultaneous crowd counting and localization system by using ESP32 modules for WiFi links. We extract several features that contribute to dynamic state (moving crowd) and static state (location of the crowd) from the CSI bundles, then assess our system by both conventional machine learning (ML) and deep learning (DL). As a result of ML-based evaluation, we achieved 0.35 median absolute error (MAE) of counting and 91.4% of localization accuracy with five people in a small-sized room, and 0.41 MAE of counting and 98.1% of localization accuracy with 10 people in a medium-sized room, by leave-one-session-out cross-validation. We compared our result with percentage of non-zero elements metric (PEM), which is a state-of-the-art metric for crowd counting, and confirmed that our system shows higher performance (0.41 MAE, 81.8% of within-1-person error) than PEM (0.62 MAE, 66.5% of within-1-person error).

DOI： 10.1109/ACCESS.2022.3155812

Scopus

researchmap

Language：Others   Publishing type：Research paper (other academic)  

In the field of crowd estimation, most non-visual approaches confine their objective to only crowd counting, whereas there are a number of vision-based researches which can estimate both the number and location of people. By observation, we figured out that the WiFi channel state information (CSI) also contains the potential characteristics for both estimations. In this paper, we propose a user-device-free simultaneous crowd estimation system that enables both crowd counting and localization simultaneously, by WiFi CSI and Machine Learning. The originality of this study is that we leverage the CSI bundles as the source for extracting features that contain characteristics depending on the dynamic state (counting) and static state (localization). By experiments during three different-time sessions, we confirm that we could achieve up to 94% counting accuracy and 95% localization accuracy by k-fold cross-validation.

DOI： 10.1109/IPIN51156.2021.9662572

Language：Others   Publishing type：Research paper (other academic)  

In Japan, the demand for nursing homes is increasing due to the rapid aging of the population, while the shortage of caregivers has become a serious problem. This problem has been recognized as a social issue because it leads to an increase in the workload per caregiver. In response, we have been developing a platform that can easily collect nursing care activities in an effort to reduce the workload. In the process, we thought that the mental state (i.e., stress) of caregivers, which changes due to their care activities, might dramatically affect their work efficiency. The objective of this study is to obtain new knowledge for reducing the workload of caregivers by visualizing and analyzing his/her stress. In this paper, we ask caregivers to wear a heart-rate sensor, and measure objective stress indicators such as R-R interval (RRI) and low frequency (LF) /high frequency (HF) ratio obtained from each sensor, as well as subjective stress indicators obtained from questionnaires administered "before work,""during lunch breaks,"and "after work."To be specific, we analyzed and visualized the changes in stress associated with care activities based on psychological indicators of caregivers collected in an actual nursing home. As a result, we found that tended to increase during certain care activities and there were some relationships between those activities and stress indicators.

DOI： 10.1145/3458744.3473346

Language：Others   Publishing type：Research paper (other academic)  

Floor vibrations caused by walking person (hereinafter, gait vibrations) have recently been explored as a way to determine their locations and identities, and the technologies for measuring such gaits may enable low-cost elderly monitoring services and crime prevention systems. In this paper, we report on the development of a system that can accurately capture both high- and low-level signal gait vibrations using a piezoelectric sensor, and then propose a new system that can identify a walking person based on a small number of footsteps. Our proposed system uses two key approaches to accurately obtain such gait vibrations. The first uses a combination of a source follower circuit in parallel and a piezoelectric sensor. The second involves widening the dynamic range by the use of a dual power supply drive. We then show how we can increase the accuracy of our system by combining multiple footsteps rather than using a single footstep, and thus achieve a more robust system regardless of the distance between the sensor and the target. In experiments comparing five different machine learning (ML) models conducted with six test participants to evaluate our system, person identification results obtained using only a single footstep showed accuracy levels up to 70.8% of average F-measure when using the Light Gradient Boosting Machine (LightGBM) classifier, while for other methods, the average F-measures were 63.1%, 75.9%, and 87.1% in cases of using the first, first and second, and from first to third footsteps from each back-and-forth walk test, respectively.

DOI： 10.1007/978-3-030-75100-5_63

Language：Others   Publishing type：Research paper (other academic)  

Crowd estimation is considered as an attractive technique for indoor energy saving, route guidance, etc. Generally, crowd estimation consists of crowd density estimation and crowd counting. In vision-based approach of crowd estimation, there are several researches addressing both crowd counting and localization. However, most of sensor or radio-based approaches are focusing on crowd counting only. In this study, we assess the potential of simultaneous crowd counting and localization by using WiFi CSI and machine learning.

DOI： 10.1145/3427796.3430000

Language：Others   Publishing type：Research paper (other academic)  

In recent years, various approaches for smart home technology have been developed, such as home appliances control, services for energy saving and support of daily life. In order to realize such services, we need a system which is able to accurately recognize various human activities using low-cost devices. To realize such a system, we need to address several problems: The required sensors are too expensive (P1); it is difficult to precisely recognize place-independent activities like reading (P2), and putting on a device causes a burden to people (P3) the information such as images infringe on the privacy of the occupants (P4). In this paper, we propose a method for activity recognition by utilizing a doppler sensor as a motion detection sensor and a machine learning technique to solve the problems above (P1-P4). Specifically, frequency characteristic is obtained from the signals of the doppler sensor and we construct a machine learning model using effective features, which is presented by Anguita, and speed of target calculated from the doppler frequency. In order to examine the usefulness of the proposed method and find out critical issues of realizing activity recognition, we have collected sensor data of 6 kinds of activities(stationary, smartphone operation, PC operation, reading, writing, and eating) performed by 10 participants. For leave-one-session-out cross-validation, the maximum average recognition accuracy was 95.7%, and the average for 10 participants was 81.0%. For leave-one-person-out cross validation, the average recognition accuracy of logistic regression shows maximum accuracy of 42.1%.

DOI： 10.1145/3427477.3429463

Language：Others   Publishing type：Research paper (other academic)  

In recent years, many studies and development of Cyber Physical Systems (CPS) have been carried out to feed back analysis results to human users in a physical space by using machine learning, aiming to analyze a huge amount of information obtained from physical space in a cyber space. To apply CPS to sports, a lot of studies have been conducted on sensing and recognizing actions and movements of athletes using machine learning. In this study, we focus on fishing as a sport, and propose a fishing CPS that recognizes anglers' actions in real-time and provides information on the past useful actions that are linked to fishing results depending on time and place as a decision support when the anglers do not make catch. In addition, this paper reports on the development of an IoT (Internet of Things) device that acquires positional information, acceleration and gyroscope information, and a web system that displays results of real-time activity recognition along with the place and time by animation for realizing the fishing CPS. We have evaluated the developed IoT device and web system from the viewpoint of practical use. As a result, we have confirmed that the GPS and acceleration sensors, in the actual breakwater environment, were constantly transmitting data to a server via UDP communication for 4 hours and 40 minutes.

DOI： 10.1145/3384419.3430447

Language：Others   Publishing type：Research paper (other academic)  

In wireless positioning systems, there are several techniques and features for location estimation of device or user. Global navigation satellite system (GNSS) is representative, but because of some advantages, Bluetooth low energy (BLE) beacon devices are also attracting attention as an alternative of conventional positioning systems. It is the reason that BLE beacons are small, inexpensive and power-efficient device that appropriate for internet of things (IoT) which is a promising future technology. However, there is a problem that the received signal strength (RSS) is not stable but fluctuating. Moreover, these RSS measurements are measured quite differently in same distance by beacon device, even the beacons are from the same manufacturer. We call this problem in RSS bias problem, and we proposed a method how to find the bias level and how to compensate these biased RSS measurements.

DOI： 10.1109/ICUFN.2017.7993833

Language：Others  

researchmap

researchmap

researchmap