Date of Award

6-2026

Document Type

Thesis

Publisher

Santa Clara : Santa Clara University, 2026

Degree Name

Master of Science (MS)

Department

Electrical and Computer Engineering

First Advisor

Maria Kyrarini

Abstract

Competitive swimming performance depends on small differences in stroke timing and body motion, but detailed video analysis is difficult to perform consistently at scale because manual review is slow, subjective, and labor-intensive. Automated video-based analysis could make stroke evaluation more repeatable and accessible, but swimming remains challenging because water, splashes, reflections, bubbles, occlusion, and partial submersion can make joint pose estimation for swimmers difficult.

This thesis develops and evaluates a video-based pipeline for freestyle stroke-phase classification and stroke counting using the SwimXYZ synthetic swimming dataset. Using this dataset, a pipeline is developed to automatically generate stroke-phase labels, enabling training without manually labeled real-world data. Freestyle arm motion is modeled using four phases: Recovery, Entry/Catch, Pull, and Push. The contributions of this thesis are: (1) a synthetic swimming dataset pipeline that automatically generates Common Objects in Context (COCO)-17-style pose keypoints and frame-level freestyle phase labels; (2) a task-specific comparison of YOLO and MediaPipe pose-estimation models using localization error, missing-keypoint rate, jitter, wrist-weighted performance, and a composite downstream score; (3) a YOLO extraction pipeline that converts videos into a windowed dataset; (4) a dual-arm Long Short- Term Memory (LSTM) model for frame-level left- and right-arm phase classification; and (5) a stroke-counting method using Hidden Markov Model (HMM)/Viterbi smoothing and rule-based phase-cycle counting. The final pipeline uses YOLO applied to videos from the side-above-water viewpoint, LSTM phase classification, HMM/ Viterbi smoothing, and rule-based phase-cycle counting to estimate completed freestyle strokes. Using perfect (ground truth) COCO-17 pose inputs, the LSTM achieved 0.9469 test accuracy, establishing that the temporal model can learn freestyle phase structure when pose information is accurate. Using YOLO-derived side-abovewater keypoints, the LSTM achieved 0.8363 test accuracy, showing that performance decreases with noisy and missing computer-vision keypoints but remains meaningful for phase classification. In a 20-clip manual validation set, the final stroke-counting pipeline achieved a total mean absolute error of 0.80 strokes and was within one stroke of the manual count in 85.00% of clips. These results show that synthetic pose and phase labels can support an interpretable freestyle swimming analysis pipeline. The system still produces meaningful phase predictions and stroke counts, providing a foundation for future real-world markerless swim analysis.

Download

Available for download on Wednesday, September 01, 2027

Share

COinS