Leveraging Markerless Computer Vision for Comprehensive Walking Automated Gait Analysis in Rehabilitation

Abstract

In the fields of sports science, biomechanics, osteopathic medicine, medical diagnostics, etc are being used in gait analysis. Traditional systems like VICON, which use passive markers and advanced video cameras, require multiple cameras for generating 3D images, which can be expensive and obtrusive. Patients suffering from diseases like stroke or spinal cord injuries may find these systems difficult to use. Markerless systems have been developed to address these issues but often lack accuracy, especially when subjects wear long clothing that obscures gait kinematics. This study aims to develop an affordable, user-friendly, and accurate gait analysis system using markerless computer vision techniques. Our approach employs advanced deep learning models for 2D (smartphone RGB camera images) to 3D reconstruction and pose estimation, enhancing the accuracy of key joint points and angle calculations. Additionally, we have introduced a new algorithm that accurately identifies the gait cycle and its phases, providing detailed insights into a patient’s condition and recovery trajectory. Validation through rigorous testing and comparison with existing methods showed an overall accuracy of 98.89% for key joint points angle, compared to the YOLO model. The computational cost analysis indicated a total processing time of 2107.40 seconds and an average of 7.53 seconds per frame. The findings have significant implications for medical and rehabilitation fields, enhancing rehabilitation strategies, optimizing prosthetic designs, and improving patient outcomes. Our proposed system effectively measures the kinematic values of the ankle, knee, and hip, and outperforms models like YOLO, which struggle with varying lighting conditions and subjects wearing long clothing.