Normal and pathological gait classification LSTM model.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Rehabilitation Technology

Background:

• Clinical gait analysis research is ongoing, but lacks sufficient public, multi-modal datasets.
• Existing gait databases are limited by small sample sizes and single-modality measurements, hindering method comparison.

Purpose of the Study:

• To introduce a novel, open-access, multi-modal gait analysis database using Kinect v.2.
• To adapt skeleton joint orientation data for kinematic parameter calculation comparable to MOCAP systems.
• To develop an unsupervised gait classification tool using Long-Short Term Memory (LSTM) ensemble models.

Main Methods:

• Acquisition of a new multi-modal gait database utilizing Kinect v.2.
• Calculation of kinematic gait parameters from skeleton joint orientation data.
• Feature extraction based on 3D lower-limb flexion dynamics for gait symmetry analysis.
• Development of an LSTM ensemble model for unsupervised gait classification.

Main Results:

• The Kinect v.2 skeleton joint orientation data can be effectively utilized for gait analysis.
• The proposed features enable gait symmetry analysis.
• The LSTM ensemble model achieved moderate performance improvements over state-of-the-art for classifying three gait types (normal/pathological).

Conclusions:

• Kinect v.2 data offers a viable solution for inexpensive clinical gait monitoring.
• The developed system demonstrates potential for accurate gait classification.
• The new database facilitates further research and comparison of gait analysis methods.