Software to Support Layout and Data Collection for Machine-Learning-Based Real-World Sensors.

There have been many studies of gesture recognition and posture estimation by combining real-world sensor and machine learning. In such situations, it is important to consider the sensor layout because the measurement result varies depending on the layout and the number of sensors as well as the motion to be measured. However, it takes time and effort to prototype devices multiple times in order to find a sensor layout that has high identification accuracy. Also, although it is necessary to acquire learning data for recognizing gestures, it takes time to get the data when the user changes the sensor layout. In this study, we developed software that can arrange real-world sensors. In this time, the software can handle distance-measuring sensors as real-world sensors. The user places these sensors freely in the software. The software measures the distance between the sensors and a mesh created from measurements of real-world deformation recorded by a Kinect. The classifier is generated using the time-series of distance data recorded by the software. In addition, we created a physical device that had the same sensor layout as the one designed with the software. We experimentally confirmed that the software could recognize the gestures on the physical device by using the generated classifier.