Vision-Based Gait Events Detection Using Deep Convolutional Neural Networks.

Area of Science:

• Biomechanics
• Computer Vision
• Machine Learning

Background:

• Accurate gait event detection is crucial for clinical analysis but challenging for vision-based methods.
• Existing methods often require complex gait feature extraction using silhouettes or human pose estimation.
• A practical, efficient, and accurate vision-based solution is needed.

Purpose of the Study:

• To develop and evaluate a multi-view deep convolutional neural network (CNN) approach for gait event detection.
• To enable gait event detection (heel-strike, toe-off) from both frontal and lateral video views.
• To eliminate the need for manual gait feature engineering.

Main Methods:

• Utilized four different deep CNN models trained on a custom dataset of 11 healthy participants.
• Input consisted of stacked video frames (9 subsequences).
• Output was a probability vector for gait events (toe-off, heel-strike) per frame.

Main Results:

• Deep CNN models achieved over 93% accuracy in detecting gait events.
• High accuracy was maintained across frontal and lateral views.
• Performance was consistent for walking straight and walking around actions.

Conclusions:

• Deep CNNs can accurately detect gait events from video without feature engineering.
• The multi-view approach is effective for both frontal and lateral perspectives.
• This method offers a practical solution for vision-based gait analysis.