Search

In the wearable computing environment, a computer provides many kinds of services by using the values from wearable sensors to recognize the user's movements or situations. In the research on activity recognition, accelerometers are attached on the user's body such as wrists, waist and, feet. Though researches on best sensor placement for context aware systems has been released thus far, they do not use enough number of sensors to really decide the best sensor placement. When using these context aware systems in our daily life, we also need to consider the discomfort that the user gets from attaching the sensors. The sensor might get in the user's way or feel uncomfortable for the user, however, as far as we know, the sensor's wearability is not taken into consideration in these researches. This paper proposes an evaluation function that scores sensor placement considering both recognition accuracy and sensor wearability, with twenty sensors on the user's body and thirty kinds of exercises including aerobic exercise, weight training, and yoga. Then we experimentally evaluated sensor placement, resulted in high degree of accuracy achieved without feeling stressful.

In the area of activity recognition, many systems using accelerometers have been proposed. Common method for activity recognition requires raw data labeled with ground truth to learn the model. To obtain ground truth, a wearer records his/her activities during data logging through video camera or handwritten memo. However, referring a video takes long time and taking a memo interrupts natural activity. We propose a labeling method for activity recognition using an execution sequence of activities. The execution sequence includes activities in performed order, does not include time stamps, and is made based on his/her memory. Our proposed method partitions and classifies unlabeled data into segments and clusters, and assigns a cluster to each segment, then assign labels according to the best-matching assignment of clusters with the user-recorded activities. The proposed method gave a precision of 0.812 for data including seven kinds of activities. We also confirmed that recognition accuracy with training data labeled with our proposal gave a recall of 0.871, which is equivalent to that with ground truth.

Mobile phones and video game controllers using gesture recognition technologies enable easy and intuitive operations, such as those in drawing objects. Gesture recognition systems generally require several samples of training data before recognition takes place. However, recognition accuracy deteriorates as time passes since the trajectory of the gestures changes due to fatigue or forgetfulness. We investigated the change in gestures and fast found that several samples of gestures were not suitable for training data. Therefore, we propose two methods of finding appropriate data for training. We confirmed that the proposed methods found better training data than the conventional method from the viewpoints of the number of data collected and the accuracy of recognition.

Mobile phones and video game controllers using gesture recognition technologies enable easy and intuitive operations, such as scrolling a browser and drawing objects. However, usually only one of each kind of sensor is installed in a device, and the effect of multiple homogeneous sensors on recognition accuracy has not been investigated. Moreover, the effect of the differences in the motion of a gesture has not been examined. We have investigated the use of a test mobile device with nine accelerometers and nine gyroscopes. We have captured the data for 27 kinds of gestures for a mobile tablet. We experimentally investigated the effects on recognition accuracy of changing the number and positions of the sensors and of the number and kinds of gestures. The results showed that the use of multiple homogeneous sensors has zero or negligible effect on recognition accuracy, but that using an accelerometer along with a gyroscope improves recognition accuracy. They also showed that some gestures were not consistent among test subjects and interdependent, so selecting specific gestures to use can improve recognition accuracy.

Human activity recognition through the wearable sensor will enable a next-generation human-oriented biquitous computing. However, most of research on human activity recognition so far is based on small number of subjects, and non-public data. To overcome the situation, we have gathered 4897 accelerometer data with 116 subjects and compose them as HASC2011corpus. In the field of pattern recognition, it is very important to evaluate and to improve the recognition methods by using the same dataset as a common ground. We make the HASC2011corpus into public for the research community to use it as a common ground of the Human Activity Recognition. We also show several facts and results of obtained from the corpus., UbiComp '11 Proceedings of the 13th international conference on Ubiquitous computing, September 17-21, 2011, Beijing, China

Mobile phones or video game controllers using gesture recognition technologies enable easy and intuitive operations. However, usually only one of each type of sensor is installed in each device, and the effect of multiple sensors on recognition accuracy has not been investigated. Moreover, the effect of the differences in the motion of a gesture has not been examined. We captured data for 27 kinds of gestures by using a mobile device with 9 accelerometers and 9 gyroscopes, we then experimentally investigated the effects on recognition accuracy of changing the number and positions of sensors, and the number and kinds of gestures. The results showed that the use of multiple sensors and of sensors positioned at specific positions affects accuracy. It was also shown that gestures are interdependent and selecting specific gestures improves recognition accuracy.

Understandings of human activity through wearable sensors will enable the next-generation human-oriented computing. However, most of researches on the activity recognition so far are based on small number of test subjects, and not well adapted for real world applications. To overcome the situation, we have started a project named “HASC Challenge” to collect a large scale human activity corpus. By the end of 2010, by the collaboration of 20 teams, more than 6700 accelerometer data with 540 subjects have been collected through our project. We also developed a tool named “HASC Tool” for management, evaluation and collection of the large number of activity sensor data., Article No.27

Although various systems use multiple accelerometers to recognize minute motions and states in the research area of context-awareness, a conventional architecture can be optimized from the viewpoint of its energy consumption and accuracy. In this paper, we propose a context-aware system that reduces the energy consumption by controlling the sampling frequency of wearable sensors. Even if the sampling frequency changes, no extra configuration is required because the missing data for the controlled sensors are complemented by our proposed algorithm. The evaluation results confirmed a power-reduction of 34.28% with keeping the accuracy. The energy consumption can be reduced without a large loss in accuracy using our system.

Many context-aware systems using accelerometers have been proposed. Contexts that have been recognized are categorized into postures (e.g. sitting), behaviors (e.g. walking), and gestures (e.g. a punch). Postures and behaviors are states lasting for a certain length of time. Gestures, however, are sporadic or once-off actions. It has been a challenging task to find gestures buried in other contexts. In this paper, we propose a method that classifies contexts into postures, behaviors, and gestures by using the autocorrelation of the acceleration values and recognizes contexts with an appropriate method. We evaluated the recall and precision of recognition for seven kinds of gestures while five kinds of behaviors; The conventional method gave values of 0.75 and 0.59 whereas our method gave 0.93 and 0.93. Our system enables a user to input by gesturing even while he or she is performing a behavior.

In wearable computing environments, a wearable computer runs various applications using various sensors (wearable sensors). In the area of context awareness, though various systems using accelerometers have been proposed to recognize very minute motions and states, energy consumption was not taken into consideration. We propose a context-aware system that reduces energy consumption. In life, the granularity of required contexts differs according to the situation. Therefore, the proposed system changes the granularity of cognitive contexts of a user's situation and supplies power on the basis of the optimal sensor combination. Higher accuracy is achieved with fewer sensors. In addition, in proportion to the remainder of power resources, the proposed system reduces the number of sensors within the tolerance of accuracy. Moreover, the accuracy is improved by considering context transition. Even if the number of sensors changes, no extra classifiers or training data are required because the data for shutting off sensors is complemented by our proposed algorithm. By using our system, power consumption can be reduced without large losses in accuracy.

There has been increasing interest in wearable computing. In wearable computing environments, a wearable computer runs various applications with various sensors (wearable sensors). Since conventional wearable systems do not manage the power supply flexibly, they consume excess power resources for unused sensors. Additionally, sensors frequently become unstable for several reasons such as a breakdown. This instability is hard to detect simply from the sensed data. To solve these problems, we propose a new sensor management device CLAD (cross-linkage for assembled devices) that has various functions for power management and sensed-data management. CLAD improves power saving, data accuracy, and operational reliability.