Automated Objective Video Quality Assessment for the Automated Edinburgh Visual Gait Score (EVGS)

Abstract

This thesis presents a framework for automated video quality assessment in gait analysis, addressing critical challenges in clinical gait evaluation. The research focuses on developing robust algorithms for detecting multiple persons, plane orientation, zoom artifacts, computing overall video quality score, and generating feedback suitable for automated Edinburgh Visual Gait Score (EVGS) scoring.

The methodology involves extracting skeletal keypoints from video frames using the MoveNet Lightning pose estimation model. The algorithms use these keypoints to detect multiple persons, track the person of interest, detect the plane of motion, identify overlapping people, detect camera zooming, and evaluate video resolution. These components are integrated into a unified quality classification system using a Random Forest classifier.

Results demonstrated exceptional performance across various metrics. The plane detection algorithm achieved excellent classification in both training and validation datasets, ensuring only appropriate videos are used for analysis. Multiple person detection and overlap assessment algorithms show strong performance, with accuracies of 96% and 95% respectively in the validation dataset. The zoom detection algorithm achieved 92% accuracy in identifying sudden zoom events in the video. The overall video quality assessment framework demonstrates a 95% accuracy in categorizing videos as either suitable for immediate analysis or requiring manual editing. This high level of accuracy showcases the effectiveness of the proposed methodology in automating the quality assessment process. The system also provides specific suggestions for improvement when videos fail to meet quality standards, enhancing the overall efficiency of the gait analysis workflow.

In conclusion, this research makes significant contributions to automated gait analysis by addressing the crucial aspect of video quality assessment. The developed framework demonstrates high performance score across various quality metrics, potentially removing the need for manual quality check for automated EVGS scoring in clinical practice.