| （英） |
With the widespread use of smartphones, there have been efforts to classify human behavior using built-in sensors. However, most of these efforts are limited to a single location where the smartphone is held and are considered insufficient to be incorporated into actual smartphones as a system. If it can be confirmed that it is possible to classify behavior and possession location at the same time, it will be possible to change the notification method of the smartphone according to the user's situation. In this study, we acquired data from the accelerometer of a smartphone, trained it using deep metric learning, and classified the user's behavior and possession position using cosine similarity during inference. As a result, not only did we obtain the same accuracy as in the previous study even when classifying both actions and possession positions at the same time, but we also confirmed that it was possible to output untrained data as an unknown class. |