Optimal Time-Window Derivation for Human-Activity Recognition Based on Convolutional Neural Networks of Repeated Rehabilitation Motions.

This paper analyses the time-window size required to achieve the highest accuracy of the convolutional neural network (CNN) in classifying periodic upper limb rehabilitation. To classify real-time motions by using CNN-based human activity recognition (HAR), data must be segmented using a time window. In particular, for the repetitive rehabilitation tasks, the relationship between the period of the repetitive tasks and optimal size of the time window must be analyzed. In this study, we constructed a data-collection system composed of a smartwatch and smartphone. Five upper limb rehabilitation motions were measured for various periods to classify the rehabilitation motions for a particular time-window size. 5-fold cross-validation technique was used to compare the performance. The results showed that the size of the time-window that maximizes the performance of CNN-based HAR is affected by the size and period of the sample used.