Master'sOpen Access

Examining sprint and jump performance using artificial intelligence techniques in sports

Is this your thesis?

This record came from a bulk archive import. If it’s yours, link it to your profile.

2026
0 views
0 downloads

Abstract (EN)

This study was conducted as a cross-sectional, methodological method-comparison investigation to examine the criterion validity of an artificial intelligence–assisted video analysis system designed for sports performance assessment and its agreement with reference measurement methods. A Python-based, human-in-the-loop, semi-automated video analysis software integrating MediaPipe Pose, YOLO-Pose, and YOLO-Seg models was developed by the research team. The study included 41 volunteer university students with a sporting background. Sprint performance was assessed at distances of 5, 10, and 15 m using four distinct anatomical and algorithmic approaches—Torso Centre, Hip Centre, Front Body, and YOLO—and compared with a physical photocell timing system. The 5- and 10-m split times were additionally compared with manual video analysis performed using Kinovea. Standing long jump distance was analysed on a 0–300 cm ground plane using four-point perspective calibration and an intermediate calibration anchor positioned at 200 cm. The equally weighted mean of the distances generated by MediaPipe and YOLO-Pose at the human-verified first-landing frame was calculated as the MP–YOLO Fusion measurement and compared with a reference distance measurement obtained independently of the artificial intelligence system. Vertical jump performance was evaluated from video recordings obtained simultaneously with the TKK 5406 Jump-MD device using MediaPipe, YOLO-Pose, and confidence-weighted composite pelvis vertical displacement approaches. Because no independent force-platform or contact-mat reference was available for jump outcomes derived from flight time, these results were examined in terms of inter-model agreement rather than absolute validity. Differences between methods were assessed using paired parametric or nonparametric tests; associations were examined using Pearson or Spearman correlation coefficients; absolute agreement was evaluated using intraclass correlation coefficients; systematic bias was assessed through Bland–Altman analysis; and measurement errors were quantified using MAE, MAPE, and RMSE. The findings showed that no single model demonstrated unequivocal superiority across all distances and measurement criteria. That measurement performance varied according to sprint distance and the anatomical reference point used. At 5 m, the Hip Centre approach emerged as the most balanced method, showing a small systematic difference of 0.018 s, a low mean absolute error of 0.049 s, and the highest ICC point estimate of 0.715. In contrast, the Torso Centre and YOLO models took significantly longer than the photocell system, with Bonferroni-adjusted p-values of 0.032 and 0.020, respectively. At 10 m, all models showed strong associations and high absolute agreement with photocell measurements, while the Hip Centre approach yielded the lowest error and the highest ICC, 0.954. At 15 m, the YOLO model was notable for producing valid measurements for all recordings and achieving the highest correlation (ρ=0.909), the highest ICC (ICC=0.905), and, numerically, the lowest error values. Nevertheless, differences in absolute error between YOLO and the Torso Centre and Hip Centre approaches were not statistically significant. In contrast, the Front Body model produced significantly greater absolute errors than the other three models. At the 5- and 10-m split points, the AI-based Hip Centre method consistently produced shorter times than Kinovea measurements (5 m median difference = -0.188 s, p = 0.002; 10 m mean difference = -0.169 s, p < 0.001). Although the associations between methods were strong, the absolute agreement was low, particularly at 5 m (ICC = 0.348). In the standing long jump analysis, 25 recordings that met the predefined technical quality criteria were evaluated. No statistically significant mean differences were found between the reference measurement and the MediaPipe, YOLO-Pose, or MP–YOLO Fusion measurements (p=0.223, p=0.319, and p=0.267, respectively), and effect sizes remained small (d=0.203–0.250). Very high correlations were observed between the reference measurement and the artificial intelligence–based measurements (r=0.975–0.983), while excellent absolute agreement was obtained between the MP–YOLO Fusion and reference measurements (ICC=0.978; 95% CI: 0.952–0.990). The Fusion method produced values that were, on average, 1.82 cm lower than the reference measurements, with Bland–Altman limits of agreement ranging from −17.49 to 13.86 cm. The mean absolute error was 5.44 cm, the mean absolute percentage error was 2.86%, and the root mean square error was 8.04 cm. These findings indicated low systematic bias and high agreement at the group level, while also revealing that larger deviations occurred in some individual measurements. In the vertical jump analysis, no significant mean differences were detected between the TKK 5406 and the MediaPipe, YOLO-Pose, or composite pelvis measurements (all p>0.05), and very strong correlations were obtained (r=0.955–0.959). High absolute agreement was identified between the composite pelvis measurement and the TKK 5406 (ICC=0.959; 95% CI: 0.923–0.978), with an MAE of 2.71 cm and a MAPE of 6.03%. In conclusion, the developed artificial intelligence–assisted video analysis system provided a promising measurement approach, demonstrating strong associations and high absolute agreement with the TKK 5406 for pelvis-based vertical jump assessment and with the independent reference distance measurement for standing long jump assessment. Nevertheless, the individual limits of agreement and maximum error observed in the standing long jump analysis indicate that the method should be used cautiously when evaluating small individual changes in performance. In sprint timing and Kinovea comparisons, systematic differences were identified depending on sprint distance, anatomical reference point, and event definition. Therefore, the findings do not support the conclusion that the developed system can be used interchangeably with the existing reference methods. Owing to the limited number of valid recordings in the 5- and 10-m analyses, missing measurements in some models, and wide confidence intervals, the findings were interpreted as exploratory. Further validation in larger and independent samples is recommended.

Author

Kader Yel

Institution

How to Cite

Kader Yel (Master Thesis). Examining sprint and jump performance using artificial intelligence techniques in sports, 2026, Bayburt University.

Keywords

License

Tüm Hakları Saklıdır

This work is shared under the specified license terms.

More theses from Bayburt University