A Spatiotemporal Deep Learning Approach for Automatic Pathological Gait Classification.

Area of Science:

• Biomedical Engineering
• Computer Science
• Rehabilitation Medicine

Background:

• Human motion analysis, particularly gait analysis, is crucial for diagnosing and assessing recovery from walking pathologies.
• Current 2D-RGB camera systems offer objective gait assessment, complementing subjective clinical evaluations.
• Existing methods often use compact gait representations (e.g., gait energy images) that may not fully capture temporal dynamics.

Purpose of the Study:

• To develop an advanced spatiotemporal deep learning model for gait analysis using 2D-RGB camera data.
• To improve the accuracy and generalization capabilities of gait pathology classification.
• To address the limitations of compact gait representations in capturing temporal gait information.

Main Methods:

• A spatiotemporal deep learning approach combining convolutional and recurrent neural networks was proposed.
• The system processes gait cycles as sequences of silhouette key frames.
• Temporal patterns were learned from spatial features extracted at different time instants.

Main Results:

• The proposed system demonstrated improved gait pathology classification accuracy.
• It outperformed existing state-of-the-art solutions on the GAIT-IT dataset.
• Enhanced generalization performance was observed in cross-dataset tests.

Conclusions:

• The spatiotemporal deep learning model effectively captures temporal gait dynamics for improved pathology classification.
• This approach offers a more comprehensive and accurate alternative to traditional compact gait representations.
• The system shows promise for objective and reliable clinical assessment of gait disorders.