Adapting MotionBERT for Sports Motion Analysis: Extraction, Comparison, and Generation of Continuous Poses

Sammanfattning

Background: Understanding human motion in sports is a challenging problem incomputer science and sports analysis. Many existing methods focus on estimating 3D human poses from videos, but they are mostly trained on datasets with simple daily activities such as walking or sitting. These models often struggle when applied to sports movements, which are faster, more complex, and involve unusual body positions. Recent research shows that transformer-based models, especially MotionBERT, have improved performance in 3D human pose estimation by learning spatial and temporal patterns together. However, there is still limited work on applying such models specifically to sports motion data. This creates a gap between general pose estimation and real-world sports analysis. Objectives: The main objective of this thesis is to adapt MotionBERT for sports motion analysis through fine-tuning on the SportsPose dataset, which contains 654 motion clips across five sports captured using optical motion capture. The focus is on improving the model to better represent continuous sports movements. The work also aims to evaluate the quality of predicted poses using spatial, temporal, and biomechanical metrics, and to extract useful motion features such as joint angles and movement speed. A further objective is to explore whether simple rule-based methods can identify movement patterns that may indicate possible injury risks, based on known biomechanical principles. Methods: A pretrained MotionBERT model is fine-tuned on sports-specific posedata. Raw SportsPose recordings use the COCO-17 joint convention, which is remapped to the H36M-17 skeleton required by MotionBERT using a custom preprocessing pipeline that synthesises four additional joints and applies torso normalisation. Training uses 40 epochs with a batch size of 4. The model is evaluated using the Mean Per Joint Position Error (MPJPE), procrustes-aligned MPJPE, mean per joint velocity error, and temporal jerk. Motion quality is further assessed through spectral analysis, dynamic time warping, and bone consistency. Clinical validity is assessed via intraclass correlation, Bland-Altman agreement, and limb symmetry indices across six biomechanical indicators. Results: Fine-tuning MotionBERT on the Sports Pose dataset achieves a MPJPE of 0.0187 (approximately 9.4mm) on the in-domain test set, representing a 98.1% reduction in error compared to the H36M zero-shot baseline (0.967 → 0.019). Cross subject evaluation across 127 clips yields a MPJPE of 0.0183, confirming generalisation. Against a learned biGRU baseline, the fine-tuned model reduces error by 75.6%. Biomechanical analysis shows all nine joint angles within 3◦ of ground truth and intraclass correlation coefficients above 0.997 for all six clinical indicators. Rulebased injury-risk classification is correct for 9 out of 10 sport-joint combinations. A motion continuation experiment using tail-mask fine-tuning achieves MPJPE of 0.209 at a 180-frame prediction horizon, outperforming nearest-neighbour retrieval for observation windows of 120 frames or more. The identified temporal limitationis excessive high-frequency jitter: jerk RMS is 14.6 times higher than ground truth, and spectral energy in the 7–15Hz band is 2,778 times above ground-truthlevels. Conclusions: This work demonstrates that a pretrained transformer-based model like MotionBERT can be successfully adapted to sports motion analysis through domain-specific fine-tuning, closing 98.1% of the domain gap from a general human motion checkpoint. It also shows that evaluating motion requires more than positional accuracy, and must include temporal and biomechanical aspects. While the system cannot be used for medical diagnosis, it supports the identification of movement patterns that may require attention, with near-perfect clinical agreement across six injury-risk indicators. The main open limitation is high-frequency temporal jitter in predicted motion, which does not affect positional accuracy but degrades smoothness; addressing this through temporal consistency regularisation or post-processing is identified as the primary direction for future work.