Available at: https://digitalcommons.calpoly.edu/theses/3396
Date of Award
6-2026
Degree Name
MS in Computer Science
Department/Program
Computer Science
College
College of Engineering
Advisor
Alex Dekhtyar
Advisor Department
Computer Science
Advisor College
College of Engineering
Abstract
This study evaluates a diverse range of computer vision pipelines for the automated assessment of two standard clinical mobility tests: Sit-to-Stand (Sit-to-Stand (S2S)) and Timed Up and Go (Timed-Up-and-Go (TUG)). Using a dataset of 132 trials, the research compares the efficacy of kinematic-based approaches against deep spatiotemporal architectures across multiple detectors, trackers, and action models. The results demonstrate that the optimal framework is highly task-dependent. For S2S, a kinematic-based approach utilizing MMPose and proximity-stitched tracking proved superior, achieving a Median Absolute Error (MedAE) of 0.00 for stand repetitions. This suggests that direct motion analysis is more reliable for discrete event counting than deep learning models like SlowOnly, which often undercounted transitions. Conversely, for the multi-stage TUG test, spatiotemporal models provided more robust performance for temporal segment detection. Pipelines utilizing You Only Look Once (YOLO) paired with StrongSORT and SlowOnly consistently achieved lower error rates than Faster R-CNN alternatives. The study identifies significant environmental hurdles to automated assessment, including inconsistent handheld camera work, varying perspectives, and occlusions caused by clinicians assisting patients. While trackers like StrongSORT and BoTSORT were expected to resolve these occlusions through ReID capabilities, they offered only fractional improvements in high-clutter environments. These findings suggest that while kinematic models are ideal for repetition-based tasks, complex mobility assessments require spatiotemporal features to overcome environmental noise.