Available at: https://digitalcommons.calpoly.edu/theses/3243
Date of Award
3-2026
Degree Name
MS in Computer Science
Department/Program
Computer Science
College
College of Engineering
Advisor
Jonathan Ventura
Advisor Department
Computer Science
Advisor College
College of Engineering
Abstract
Gait analysis provides valuable insight into health status, mobility impairments, injury prevention, and fall risk. Traditional approaches require controlled laboratory environments with marker-based motion capture systems or fixed camera configurations, imposing substantial burdens: reflective markers must be attached to anatomical landmarks, participants must walk along specified paths or step on designated foot placements, and detailed protocols require instruction and practice. These requirements make traditional gait analysis expensive, time-consuming, and infeasible for daily use. Recent advances in markerless pose estimation have permitted video-based gait analysis from ordinary footage of natural movement without these participant demands. Yet existing models remain largely confined to laboratory settings with treadmill or fixed-path walking, constrained viewing angles, and stationary cameras. These constraints limit applicability to real-world health monitoring, where individuals move freely throughout natural environments.
This thesis addresses step detection, step counting, and activity classification tasks through the use of unconstrained, real-world video captured in third person by a following camera operator during participants' natural daily activity. This setting introduces challenges absent in typical gait analysis tasks: camera motion that impacts perceived limb movement, frequent occlusions, varying viewing angles, and non-linear motion.
A data processing pipeline was designed to extract pose keypoints via Mediapipe and visual embeddings via a Convolutional Vision Transformer (CvT); these features were then used to train multiple deep sequence models over temporal windows centered on each frame. We also apply a "peaky focal'' loss function to address severe class imbalance and push the system towards localized peaks during step detection. The resulting system offers a scalable, low-cost, and lightweight solution for real-world gait analysis applications and establishes a baseline for future work in unconstrained biomechanics analysis from video.
Included in
Artificial Intelligence and Robotics Commons, Biomechanics Commons, Exercise Science Commons, Other Computer Sciences Commons