Automatic motion analysis of a gymnast in rhythmic gymnasticsKolar, Maja (Avtor) Skočaj, Danijel (Mentor) deep neural networkscomputer visionrhythmic gymnasticspose estimationmovement classificationautomatic scoringmachine learningRhythmic gymnastics requires precise execution of complex movements, yet objective motion analysis and evaluation remain limited and largely dependent on human judgment. This thesis addresses this challenge through automatic motion analysis of Body Difficulty (BD) elements, defined by measurable body positions. We propose a modular pipeline that integrates 2D keypoint detection, 3D pose estimation, action recognition, and temporal segmentation to provide a structured breakdown of routines. On top of motion analysis, we explore grading of execution quality with a regression model. Experimental results show that the pipeline reliably detects and classifies BD elements, achieving strong performance at both the movement type and subtype level. Implemented methods robustly segment full routines, while the grading model, though limited by data, demonstrates the potential of extending motion analysis toward more objective evaluation in rhythmic gymnastics.20252025-09-30 15:25:00Magistrsko delo/naloga174285sl