Measurement Boundaries of Single-Camera Markerless Analysis During Early Sprint Acceleration

The validity of single-camera markerless analysis is often treated as a property of the system, although criterion agreement may also depend on how the same recording is analytically represented and what it is compared against. We tested these dependencies during early sprint acceleration using identical side-view recordings from 14 male athletes, including 12 amateur flag football players (140 trials). Direct two-dimensional MediaPipe and body-model-based Human Skeleton and Mesh Recovery (HSMR) outputs were compared with synchronised Vicon Plug-in Gait kinematics; MediaPipe was also compared with sagittally projected Vicon marker geometry, and video-derived events with force-platform definitions. MediaPipe hip and knee waveform correlations were 0.968–0.991, with root-mean-square errors (RMSEs) of 10.73–12.65°. Relative to MediaPipe, HSMR reduced the hip and knee RMSEs by 4.66° and 2.63° but increased the ankle RMSE by 9.33°. The MediaPipe RMSE was 1.46–2.79° lower against projected marker geometry than against Plug-in Gait, and the mean absolute error (MAE) was estimated to be 21.5% higher at the fifth than at the first analysis-window position across joints and viewing conditions. The video-event timing RMSE was 26.36–34.99 ms. Criterion agreement was therefore conditional on analytical workflow, comparator representation, and movement context rather than a fixed property of the recording; broad proximal waveform description was supported, whereas precise absolute joint angles and force-equivalent timing were not.