Vision-based gait impairment analysis for aided diagnosis

Area of Science:

• Biomedical Engineering
• Computer Science
• Kinesiology

Background:

• Gait analysis is crucial for health assessment, but current automated methods often lack interpretability or require specialized equipment.
• Existing approaches frequently use complex spatio-temporal descriptors, hindering clinical adoption and understanding.

Purpose of the Study:

• To develop and validate a set of semantic, normalized gait features for quantifying pathological gait from single-channel video.
• To create features that are invariant to video acquisition parameters (frame rate, image size) for cross-platform comparability.
• To assess the accuracy and sensitivity of these features in detecting gait impairments.

Main Methods:

• Proposed novel semantic and normalized gait features designed to quantify gait impairment, including asymmetry and falling risk.
• Utilized a general-purpose gait dataset for establishing normal feature references.
• Introduced a new dataset with eight walking styles (one normal, seven pathological) for feature validation.
• Conducted statistical analyses to evaluate feature sensitivity and accuracy.

Main Results:

• The proposed gait features demonstrated sensitivity in measuring expected pathologies across different walking styles.
• Statistical studies provided evidence of the accuracy of the developed features.
• Features were designed to be invariant to frame rate and image size, enabling comparisons across different video sources.

Conclusions:

• The developed semantic gait features offer a robust, interpretable method for analyzing pathological gait using low-cost video sensors.
• These features have the potential to assist physicians in objective clinical decision-making.
• The new dataset and validated features advance automated gait analysis for healthcare applications.