Neural network-based user-independent physical activity recognition for mobile devices

Activity recognition using sensors of mobile devices is a topic of interest of many research efforts. It has been established that user-specific training gives good accuracy in accelerometer-based activity recognition. In this paper we test a different approach: offline userindependent activity recognition based on pretrained neural networks with Dropout. Apart from satisfactory recognition accuracy that we prove in our tests, we foresee possible advantages in removing the need for users to provide labeled data and also in the security of the system. These advantages can be the reason for applying this approach in practice, not only in mobile phones but also in other embedded devices.