Author(s): LIN, S., MING, L., HAUTIER, C., ZHANG, Q.S., Institution: SHANGHAI UNIVERSITY OF SPORT, Country: CHINA, Abstract-ID: 1111

INTRODUCTION:
The optical marker-based motion capture (Mocap) is the gold standard for kinematic assessment. However, its high cost, complexity, and restriction to laboratory settings limit routine use.Deep learning-based markerless motion capture has emerged as a practical, non-invasive alternative for field application. For now, Optimizing rowing biomechanics is crucial for enhancing performance; thus, its accuracy in specific rowing tasks requires rigorous verification. This study aims to evaluate the reliability and validity of a single-camera, deep learning-based markerless system against the marker-based gold standard during indoor rowing.
METHODS:
Twenty-five experienced male rowers performed a maximal rowing protocol on a Concept2 ergometer, consisting of five resistance levels with two sets of 20 strokes per level. Kinematic data were simultaneously recorded using a tripod-mounted DJI action camera and a Qualisys Mocap system. Video data were processed using a custom deep learning algorithm to compute rowing metrics. Reliability and validity were assessed using linear mixed-effects models, Correlation Coefficient（ICC）, Bland-Altman analysis, and Root Mean Square Error（RMSE）
RESULTS:
The results demonstrated excellent agreement with the Mocap system for the assessed kinematic parameters across temporal, spatial, and velocity domains. For example, Peak Drive Velocity showed a negligible bias of 0.041 m/s (ICC = 0.979, RMSE = 0.067 m/s, MAPE = 2.96%, LoA: -0.06 to 0.14 m/s). Similarly, spatial parameters such as Drive Distance exhibited high concurrent validity, showing a bias of only -0.048 m (ICC = 0.962, RMSE = 0.050 m, MAPE = 5.17%, LoA: -0.08 to -0.02 m). Furthermore, temporal metrics showed the strongest consistency; specifically, Stroke Rate demonstrated near-perfect agreement with a bias of 0.082 spm (ICC = 0.998, RMSE = 0.807 spm, MAPE = 0.59%, LoA: -1.49 to 1.66 spm).
CONCLUSION:
This study demonstrates that a single consumer-grade camera, powered by deep learning algorithms, offers a highly valid and reliable alternative to traditional marker-based systems for analyzing rowing kinematics. By achieving high agreement with the gold standard Mocap across multiple resistance levels, this markerless approach effectively overcomes the logistical and financial barriers of laboratory-grade equipment. Consequently, it provides a practical, low-cost tool that enables coaches and researchers to conduct routine biomechanical monitoring and technical analysis directly in rowing.