Gait anomaly detection based on video-derived 3D pose estimation.

With the increase of age, the lower limb strength and function of the elderly gradually decline. Timely detection of motor dysfunction is of great significance for the prevention of disability, disease intervention, and improvement of living quality. Focusing on gait monitoring of the elderly living in groups, such as nursing homes, an abnormal gait recognition network based on daily walking information is proposed. We improve a multi-view 3D pose estimation network to extract gait parameters from the TUG exercise for monitoring, and design the abnormal gait recognition network to solve the problems of late evaluation of movement ability, large subjectivity, and the balance between accuracy and speed of the elderly living in groups. At a frame rate of 21.75 fps, the pose estimation accuracy is stable above 96.53%, and the joint error is controlled within 3.63°. In gait anomaly detection, the sensitivity reaches 96.71% and the inference speed reaches 512 ms; the F1 score reaches 0.9680, which is very close to the optimal value of the participant-comparison model, and the AUROC reaches 0.9694. This humble gait monitoring technology has great potential to provide assisted care and improve the overall well-being of the elderly.