Ren Ohmura

Parameter exploration for response time reduction in accelerometer-based activity recognition

In activity recognition techniques, existing wearable sensors have a problem performing the recognition process. Because existing wearable sensors perform recognition process by dividing sensor data into partial sequences, there is lag between the changes in action and the output of the recognition result. Therefore, we focused on the point activities change and have proposed a method to reduce the response time of the activity recognition technique. However, parameters such as window size immediately after the activity changing point and the activity changing point detection in the proposed method have not been studied well. Thus, in this paper, we conducted experiments using the HASC Corpus, which contains large-scale data of human activity. We report results of examining various parameters in the proposed method and features of the proposed method revealed by comparison with a conventional method. To give a concrete example, for IIR band-pass filter bank to be used for activity changing point detection, we clarified the frequency and the appropriate number of filters. In addition, we clarified the relationship between identification accuracy and the size of a special window that is set after activity changing point detection. The proposed method reduced the response time to the 2035ms on average from 2773ms, the of average of the conventional method. In addition, the proposed method can reduce the amount of calculation, achieve both high recognition accuracy and short response time, and output the recognition results in consistent times to reduce the jitter of response time.

Embodiment of an agent by anthropomorphization of a common object

This paper proposes a direct anthropomorphization method to improve interaction between human and an agent. In this method, an artifact is converted into an agent by attaching humanoid parts to it. There have been many studies that have provided valuable information on using spoken directions and gestures via anthropomorphic agents such as computer graphics agents and communication robots. In the direct anthropomorphization method, an artifact is directly anthropomorphized by being fitted with robotic parts shaped in the form of human body parts. An anthropomorphized artifact with such robotic parts can provide information to people through spoken directions and body language. This will persuade people to pay more attention to the artifact, as compared to when using an anthropomorphic virtual or robot agent. The authors conducted experiments to investigate how the response of users to an explanation of the functions of an artifact changes using the direct anthropomorphization method. The results of pre-experiment indicated that participants paid more attention to the target artifact and memorized its functions more quickly and easily when using the direct anthropomorphization method than when using humanoid agent. In following two experiments, the authors compared human-like aspect separately and evaluate what is key element for anthropomorphization. The authors found that “voice” was the key factor for rendering an object as an anthropomorphic agent. Furthermore, the “eyes” were found to be more effective in interactions than the “mouth”.

Wireless sensor network system for supporting nursing context-awareness

We developed a wireless sensor network system for supporting context-awareness of nursing activities in hospitals. Our system is aimed at automated recording of nursing work, providing context-aware services to nurses and visualising analytical results from nursing histories. The system consists of heterogeneous devices utilising three kinds of wireless networks: a ZigBee TM (ZigBee is a registered trademark of the ZigBee Alliance)-based location sensor network, a Bluetooth TM (Bluetooth is a trademark owned by Bluetooth SIG, Inc., USA)-based body-area sensor network for capturing nurse activities and a Wi-Fi TM (Wi-Fi is a registered trademark of the Wi-Fi Alliance)-based network for communication between PDAs worn by nurses and a server PC. We illustrate the layered structure of the entire system as well as algorithms for estimating nurse locations and activities during their operations in a hospital. These estimated results are managed by a real-time database system with which nurse contexts can be visualised in real time on a server PC and/or PDAs. We also show empirical results evaluating the performance of retrieving, analysing and visualising detailed nursing contexts.

Towards concurrent task verification in context-aware applications

For achieving advanced support in a practical context-aware system, the system requires not only to recognize a users activity sequence but also to verify the correctness of the sequence, especially in complex work, such as medical treatment and nursing. However, since those users generally undertake multiple tasks simultaneously and perform them concurrently, the verification of a work increases in complexity. In this study, we propose a method to define a task and to verify multiple tasks that are performed concurrently. We first consider the definition of a single task by regarding it as an ordinal state machine and categorize events into three categories. Then, we propose a verification method for those concurrent tasks by assigning priorities on the transition categories. Our method has been confirmed to perform appropriately by use of a simulation that includes practical vital measurement tasks.

Story creation approach for sensor network application development

This study proposes a story creation approach for developing a sensor/home network application. In this approach, a device on a network is regarded as a character in a story. In this respect, the devices state and action is defined by the description of the character with a natural language corresponding to an ECA rule. Transmitting messages among the devices are dealt as conversation among the characters. This paper will show the detailed concept and an actual application development environment using this approach. The result of the experiment, in which 18 children and six adults were involved, showed the usefulness of the approach.

Exploring combinations of missing data complement for fault tolerant activity recognition

Disrupting the transmission of sensor data due to sensor failure or connection loss significantly decrease accuracy in existing activity recognition techniques. We introduce an approach towards managing missing sensor data which operates at each step of the standard activity recognition, beginning with raw sensor data, feature calculation, classification, and result, as well as their combination methods. Our evaluation showed that the F1-score increased from 0.61 in the case of sensor data loss to 0.68 with the combination of all methods. Moreover, by selecting the combination of methods according to the failed sensor position, the F1-score increased to 0.69.

Preliminary evaluation of feature level compensation for missing data in multi-sensor activity recognition

Activity recognition using multiple body-worn sensors can directly monitor the movement of each body part and can recognize various activities accurately. However, using multiple sensors increases the chance of sensor failure or communication failure, and most current activity recognition algorithms do not work when failure occurs due to the difference (reduction) of the dimension of the feature vector from that of complete sensor data expected in system design time. Therefore, we compared three possible techniques to solves this problem on the feature value level: a classifier trained with reduced feature values, feature value compensation with multiple regression, and feature value compensation with kernel regression, in a no failure situation. All of these techniques do not depend on classification algorithms. While creating a regression model, which is in the training phase, requires relatively high computational power, compensation itself can work with low computational power. As overall results, kernel regression had the best performance that was the closest to the no failure situation. Also, the results imply that each sensor position has its own effective method and more accurate coping can be viable with the appropriate choice of the method.

An annotation tool of layered activity for continuous improvement of activity recognition

Automatic activity logging was recently achieved by combining activity recognition techniques with body area sensor networks. However, collecting labeled data requires a rather high human load and is therefore an obstacle that prevents practical implementation of such systems. There are also cases in which human activity cannot be analyzed by using a simple activity set such as that used with conventional approaches. Therefore, we propose an annotation tool based on an active learning approach. Our tool provides an environment where a huge amount of annotation data can be easily obtained, and the labeled data can be continuously collected by seamlessly linking confirmed and annotated tasks. In addition, the tool allows the user to analyze human activity depending on the purpose by using layered activities. We conducted experiments to evaluate the usefulness of our tool. The experiments showed that our tool was effective for reducing the time needed for labeling and was also effective for improving classifiers.

Sustainable sensor network architecture for monitoring human activities

In this paper, focusing on a sensor network for monitoring human activities, the suitable architecture is discussed to obtain sustain-ability. After briefly showing a general architecture in present, the problems happening with long time usage are described. Then, our ideas are suggested to solve these problems.

Practical Feedback Method for Mobile CPR Support Systems Considering Noise and User's Attention

In this study, with the aim of improving bystander CPR support using a smartphone and a smartwatch, we evaluated six feedback methods considering practical situations. Since CPR is sometimes required to be conducted in a noisy place, each method was evaluated with 50 dB and 80 dB noise environments, which correspond to a silent office and a noisy construction site, respectively. Also, considering the requirement for a bystander to maintain the safety of him/herself and in order to give appropriate care to the patient, the capability of noticing change in patient condition during CPR was evaluated. From the evaluation results, the best feedback method is a method that uses voice, a metronome sound and a graphic display on a smartphone and vibration and graphic display on a smartwatch if both a smartphone and a smartwatch are available. For only a smartphone, the result shows that feedback using only voice is better in the loud environment, while feedback using voice and clicking sounds is the best in the quiet environment. Moreover, with regard to the subjective feeling, feedback using only a smartwatch is worse than other methods, and it is recommended to be used in conjunction with a smartphone.