Detecting Repetitive Human Actions by Neural Networks Trained on Composite Data Only
Yuuki Nishino, Takuya Maekawa, Takahiro Hara · 2023
We propose a neural network-based method for detecting the temporal locations of human repetitive actions in time-series data obtained from a body-worn accelerometer. Because labeling the temporal locations of human repetitive actions in the accelerometer data is costly, resulting in difficulties in preparing training data of the neural network, we generate composite labeled training data using existing accelerometer datasets. The proposed neural network uses the U-Net architecture to acquire a feature map capturing action-independent features of repetitions from the input data and processes the feature map in its 1D Fully Convolution Network, which outputs a binary time-series indicating the temporal locations of repetitions. Our experiments demonstrated that our proposed network is able to detect repetitions with high accuracy even though it is trained only on composite data.