Author(s): KOKUDO, S., SANO, T., UEDA, K., KOISO, T., Institution: CHUKYO UNIVERSITY, Country: JAPAN, Abstract-ID: 1110

INTRODUCTION:
Recent advances in pose estimation AI enable quantitative movement analysis, but field-based analysis of children’s running remains technically challenging due to occlusions, tracking instability, and environmental noise. Raw AI outputs often lack sufficient reliability for direct movement evaluation. This study aimed to develop and validate a foundational technical pipeline to stably extract evaluation features from children’s running videos recorded in field settings, with particular emphasis on robustness under real-world conditions.
METHODS:
A total of 59 image sequences (118 runners) were analyzed. Person detection utilized a retrained YOLOv11 model (Precision = 1.000, Recall = 0.994, mAP@0.5 = 0.995), ensuring stable lower-limb detection. Pose estimation was performed with RTMPose to obtain frame-wise joint coordinates. A custom algorithm corrected tracking interruptions, joint outliers, and left–right swaps by applying temporal continuity constraints alongside joint-angle and segment-length consistency checks. To represent key components of overall running movement quality, five categories of evaluation features were extracted: thigh/knee angles at maximal knee lift; trunk inclination from shoulder–hip midpoint slope; maximal forward, upward, and backward elbow positions; peak swing-leg ankle height; and running cycle temporal durations defined by phase transitions in joint trajectories.
RESULTS:
Complete body overlap in six sequences prevented reliable identification of two runners. In five cases involving limb crossing, pose estimation remained accurate after applying the correction algorithm. Right-ankle estimation deteriorated in eight runners under shadow conditions; however, these misestimations occurred outside the specific evaluation frames and therefore had minimal impact on feature extraction. Overall, the system demonstrated robustness under partial occlusion, whereas complete body overlap remained a primary constraint.
CONCLUSION:
The findings indicate that performance limitations arise mainly under severe occlusion rather than moderate limb interaction. Integration of detection refinement, temporal correction, and biomechanically informed filtering substantially improved stability compared to raw pose estimation outputs. The proposed framework mitigated common field-based errors, enabling consistent extraction of running-related features in most real-world conditions. This technical foundation represents a critical step toward the development of an automated digital coaching system capable of assessing movement quality and providing structured feedback in natural training environments. Future work will focus on environment-specific correction strategies, including shadow compensation, occlusion-aware modeling, and real-time implementation to further enhance robustness and scalability.