Adaption to the Individual Differences Using an Online Tuning System in the Awakening Behavior Detection System
Hironobu Satoh, Fumiaki Takeda · IEEJ Transactions on Electronics Information and Systems · 2010
To prevent a person from falling of a bed, we have developed an awakening behavior detection system using a neural network (abbreviated as NN). However, the detection ability of unknown persons is not sufficient compared to that of learned persons. In this research, to improve the detection ability of unknown persons, we apply an online tuning system using a continuous learning of the NN to the detection system. In the online tuning system, only a few additional data of a new target person are used for the continuous learning, where the weights of the NN converged in the initial learning are used as the initial weights for the continuous learning. In this paper, first, we verify that the individual differences among persons affect the detection ability. Second, we demonstrate that the detection ability is improved by executing the online tuning. Thus, we verify the effectiveness of the online tuning of the awakening behavior detection system.