Automatic Analysis of Gait Cycle Using MediaPipe

抄録

One goal of gait analyses in gait rehabilitation is to determine treatment effects. However, gait analysis, which is mainly based on observation, is limited in determining treatment effects because it is only a qualitative evaluation. Besides, quantitative gait analysis requires special facilities and equipment, including the knowledge and skills to handle them. Therefore, this study constructed a quantitative and simple gait analysis system using MediaPipe, which can automatically analyze gait cycles from gait videos captured by smartphones and other devices. Then, I created a classification model using neural networks to discriminate gait states based on the changes in knee and ankle angles estimated from the gait videos. Investigations revealed that although the model could discriminate 94.2% of the gait state changes in the right leg and 93.5% in the left leg, 68.1% of misclassifications occurred near the change in the gait state. This alteration is hypothesized to be because of ambiguities in the teacher data used to train the classification model. Hence, future work should further improve the discrimination rate and verify the limping and shuffling gait in nonhealthy subjects.