Classification of Gaits with a High Risk of Falling Using a Head-Mounted Device with a Temporal Convolutional Network

The abnormal gaits of the elderly in daily life will increase the risk of falling accidents and lead to the unbalanced development of muscles. This letter presents a method for classifying abnormal and normal gaits based on a head-mounted device with an inertial measurement unit (IMU). The IMU signals of abnormal gaits, which are captured in a 10 m walking experiment, yield acceleration, angular velocity, and rotational angle of head motion. A temporal convolutional network (TCN) model is implemented to classify three abnormal gaits. These gaits may be associated with sudden onset of pain or muscle weakness, causing a risk of falling. The classification results yield average accuracy, recall, precision, and F1-score of 98.33%, 98.27%, 98.2%, and 98.23%, respectively. In another experiment, changes in gait from normal to abnormal during walking are used to evaluate the time delay of abnormal gait detection. The average time delay is 2.52 s, and the detection accuracy is 100%. The overall results indicate that the proposed head-mounted device with TCN model can effectively monitor the abnormal gaits.