2D Video-Based RFD Prediction

Accurate measurement of vertical ground reaction force and rate of force development during countermovement jumps is essential for evaluating explosive strength; however, such assessments are typically limited to laboratory environments due to the cost and complexity of force plates. This study proposes and validates a low-cost, camera-based machine learning framework to reconstruct full vertical ground reaction force waveforms and estimate braking-phase rate of force development using spatiotemporal parameters extracted from single-camera two-dimensional video recordings. A total of 263 healthy adolescents performed 998 valid countermovement jumps while synchronized force plate data sampled at 1000 Hz and lateral video recordings at 60 Hz were collected. Six spatiotemporal features derived from video, along with participant height, were used to train a pointwise Random Forest regression model to predict vertical ground reaction force across 1000 time points. Predicted waveforms were evaluated against force plate measurements using statistical parametric mapping, root mean square error, normalized root means square error, coefficient of determination, and intraclass correlation coefficients. The reconstructed waveforms demonstrated strong agreement with reference data, with median coefficients of determination exceeding 0.60 and intraclass correlation coefficients above 0.70 across most of the pre-flight phase. Rate of force development estimates showed high validity for time windows of 20 milliseconds or longer, with the highest accuracy observed for average braking-phase rate of force development. These findings indicate that a simple two-dimensional video and Random Forest pipeline can provide reliable and clinically meaningful estimates of countermovement jump kinetics in youth athletic and clinical settings.