Automatic Labeling Method for Human Motion Time Series Using Wavelet Energy Entropy

To enhance labeling efficiency of human motion time series, a new technique of automatic labeling for human motion time series was proposed.9 inertial measurement units placed on each of 6 subjects were used to acquire motion data from 17 human activities in terms of acceleration and angular velocity. Sliding window technique was adopted to segment motion data from each subject while multiresolution analysis was applied to calculate the corresponding wavelet energy entropy. A segmenting threshold and time constraints were chosen to label each subject's motion data automatically. The results showed that the average labeling accuracy reached 95.82%. It took the proposed method approximately 18.6 minutes to label motion data of all 6 subjects, which was 75.76% shorter than the average labeling time by human, 76.75 minutes. The proposed method improved the labeling process significantly with relatively high average labeling accuracy.