A Low-Cost Monocular Video-Based Framework for Bilateral Plantar Load Estimation During Treadmill Walking

Abstract

Plantar load is an important biomechanical measure in gait analysis, rehabilitation assessment, and human movement evaluation. However, conventional measurement systems, including instrumented insoles and laboratory-based force measurement devices, rely on dedicated sensing hardware and may limit the accessibility of continuous gait monitoring. This study proposes a low-cost, monocular video-based framework for estimating bilateral plantar loads from pose-derived biomechanical features. As an initial application setting, we focused on treadmill walking, where a stationary camera can capture repeated body motion under consistent recording conditions, allowing the relationship between video-derived biomechanical features and plantar loading to be investigated before extending the framework to more variable environments. Video and reference plantar-load data from 14 healthy adults during treadmill walking at 3 km/h were temporally aligned, yielding 50,464 frame-level observations. Body landmarks estimated using MediaPipe Pose were used to construct posture-related, motion-related, and short-term temporal features. Four machine-learning regressors were evaluated, with fold-specific Top-60 feature selection applied to the morphology-excluded feature configuration. ExtraTrees achieved the highest mean performance under blocked temporal five-fold cross-validation, with R2 values of 0.762 ± 0.068 and 0.772 ± 0.073 for the left- and right-foot loads, respectively. Feature-group ablation showed that removing three static morphology variables had only a small effect on blocked-validation performance. Under zero-shot leave-one-subject-out validation, the corresponding mean R2 values were −0.252 ± 0.592 and −0.180 ± 0.573, indicating limited generalization to previously unseen participants. Using the initial 30 s of reference-load data for subject-specific supervised calibration improved prediction performance on the matched post-calibration interval. Regularized temporal, scale, and offset calibration increased the mean R2 values from −0.233 to 0.358 for the left-foot target and from −0.141 to 0.444 for the right-foot target. These findings demonstrate the potential of monocular video-derived biomechanical features for bilateral plantar-load estimation during controlled treadmill walking. Short supervised calibration improved average prediction performance for unseen participants, enabling subsequent video-based estimation without continuous wearable-sensor measurements.