A novel dataset and deep learning-based approach for marker-less motion capture during gait

Area of Science:

• Biomedical Engineering
• Computer Vision
• Human Motion Analysis

Background:

• Deep learning human pose estimation shows promise but is limited for gait studies due to homogeneous datasets and marker placement errors.
• Existing datasets lack diversity, hindering clinical applications requiring accurate gait analysis in varied populations.

Purpose of the Study:

• To introduce a novel marker-less motion capture system and a specialized dataset (ENSAM) for advancing gait analysis.
• To evaluate the efficacy of deep learning-based pose estimation for clinical gait studies using the new dataset.

Main Methods:

• Developed a marker-less motion capture system utilizing deep learning pose estimation.
• Collected the ENSAM dataset comprising walking trials from diverse participants, including those with scoliosis, spondylolisthesis, and bone disease.
• Evaluated and refined the pose estimation model by training and testing on the ENSAM dataset, comparing results with a marker-based system and medical imaging.

Main Results:

• Fine-tuning the pose estimation model on the ENSAM dataset significantly improved Bland-Altman confidence intervals for joint positions (e.g., hip joint improved from 106.9mm to 17.4mm).
• The marker-less system achieved mean joint position errors ranging from 6.2mm (ankles) to 21.1mm (shoulders) on the ENSAM test set.
• The refined system demonstrated enhanced accuracy for human pose estimation in a clinical context.

Conclusions:

• The proposed marker-less system and ENSAM dataset show significant potential for accurate human pose estimation in gait studies.
• The findings suggest improved applicability of marker-less motion capture for clinical gait analysis, especially in diverse patient groups.
• Further research is warranted to assess the system's performance on specific gait parameters.