Background: Traditional biomechanical analysis of cycling motion relies on expensive laboratory equipment, limiting accessibility for many coaches and athletes. Advances in computer vision offer potential for extracting meaningful data from standard video footage.

Objectives: This thesis develops and evaluates a pipeline using 2D-to-3D pose estimation to analyze cycling biomechanics, investigating its effectiveness in extracting metrics from video data and comparing these across different cycling scenarios.

Methods: We integrated AlphaPose for 2D keypoint detection, MotionBERT for 3D pose estimation, and Savitzky-Golay filtering for temporal smoothing into a user-friendly Streamlit application. The pipeline was applied to 12 cycling videos (3,190 frames) covering flat terrain, uphill, and sprinting scenarios, extracting metrics like joint angles, angular velocities, and range of motion (ROM).

Results: The pipeline achieved a mean absolute error of 7.2° for joint angles compared to literature values. Distinct patterns emerged across scenarios: uphill cycling showed higher knee ROM (52.3° vs. 44.7° for flat terrain, +17%), while sprinting exhibited greater angular velocities (knee extension: 420°/s vs. 310°/s for flat, +35%). The Streamlit app enabled accessible analysis for non-technical users.

Conclusions: Combining pose estimation with biomechanical analysis offers a viable, low-cost alternative to laboratory systems, enhancing accessibility and ecological validity. Future work should focus on validation against gold-standard systems and improving occlusion handling.