A Lightweight Pose Sensing Scheme for Contactless Abnormal Gait Behavior Measurement

Area of Science:

• Biomechanics
• Medical Imaging
• Machine Learning

Background:

• Abnormal gait recognition is crucial for motion assessment and disease diagnosis.
• Current methods using wearable sensors face challenges like data drift and patient discomfort.
• Developing contactless, accurate gait analysis is essential for broader clinical application.

Purpose of the Study:

• To propose a contactless method for abnormal gait behavior recognition using a monocular camera.
• To develop a lightweight OpenPose (OP) model for real-time joint point extraction and 3D gait reconstruction.
• To evaluate the effectiveness of machine learning algorithms for classifying abnormal gait features.

Main Methods:

• A monocular camera captures human pose data, processed by a lightweight OpenPose (OP) model with Depthwise Separable Convolution.
• Extracted 2D joint coordinates are reconstructed into 3D data to extract 11 types of abnormal gait features.
• The XGBoost algorithm is used for feature screening, followed by Random Forest (RF) classification.

Main Results:

• The proposed method successfully recognizes abnormal gait behavior in real time.
• The Random Forest (RF) algorithm combined with 3D gait features achieved the highest precision of 92.13%.
• The system demonstrates robustness against data drift and eliminates the need for wearable sensors.

Conclusions:

• The contactless gait recognition method offers a viable alternative to traditional wearable sensors.
• This approach reduces hardware requirements and overcomes challenges associated with sensor use, particularly in the elderly.
• The real-time, contactless capabilities position this method for wide application in abnormal gait measurement and diagnostics.