A monocular video-based framework using pose-derived biomechanical features and machine learning achieved R2 values of 0.762 and 0.772 for bilateral plantar load estimation during treadmill walking, with limited zero-shot generalization that was substantially improved by short supervised calibration.
Key Findings
Results
The ExtraTrees regressor achieved the highest mean performance for bilateral plantar load estimation under blocked temporal five-fold cross-validation.
R2 values of 0.762 ± 0.068 and 0.772 ± 0.073 for left- and right-foot loads, respectively
Four machine-learning regressors were evaluated: ExtraTrees was the top performer
Blocked temporal five-fold cross-validation was used to prevent data leakage across time
Fold-specific Top-60 feature selection was applied to the morphology-excluded feature configuration
Dataset comprised 50,464 frame-level observations from 14 healthy adults walking at 3 km/h
Results
The framework showed limited generalization to previously unseen participants under zero-shot leave-one-subject-out validation.
Mean R2 values were -0.252 ± 0.592 and -0.180 ± 0.573 for left- and right-foot loads, respectively
Negative R2 values indicate predictions worse than a simple mean baseline for some participants
Large standard deviations (±0.592 and ±0.573) indicate high variability in performance across subjects
Zero-shot validation used no data from the test subject during training
Results
Short supervised calibration using 30 seconds of reference-load data substantially improved prediction performance for unseen participants.
Regularized temporal, scale, and offset calibration increased mean R2 from -0.233 to 0.358 for the left-foot target
Calibration increased mean R2 from -0.141 to 0.444 for the right-foot target
Only the initial 30 seconds of reference-load data were used for subject-specific calibration
Performance was evaluated on the matched post-calibration interval
This approach enabled subsequent video-based estimation without continuous wearable-sensor measurements
Results
Removing three static morphology variables had only a small effect on blocked cross-validation performance.
Feature-group ablation analysis was conducted to assess the contribution of different feature groups
Features were grouped into posture-related, motion-related, and short-term temporal categories
Static morphology variables were excluded from the primary feature configuration evaluated
The small effect of morphology exclusion suggests the framework can function without body measurement inputs
Methods
Body landmarks estimated using MediaPipe Pose were used to construct three categories of biomechanical features for plantar load estimation.
Feature categories included posture-related, motion-related, and short-term temporal features
MediaPipe Pose was used as the pose estimation backbone, enabling low-cost monocular video input
Video and reference plantar-load data from 14 healthy adults were temporally aligned
Data were collected during treadmill walking at 3 km/h under consistent recording conditions
A stationary camera setup was used, exploiting the controlled and repeated nature of treadmill locomotion
Background
The study focused on treadmill walking as an initial controlled application setting before potential extension to more variable environments.
Treadmill walking was chosen because a stationary camera can capture repeated body motion under consistent recording conditions
14 healthy adults participated in the study
Walking speed was fixed at 3 km/h
The authors explicitly noted this as an initial step prior to extending the framework to more variable environments
What This Means
This research suggests that it may be possible to estimate how much force a person places on each foot while walking using only a standard video camera and machine learning software, without requiring expensive pressure-sensing insoles or laboratory equipment. The researchers recorded 14 healthy adults walking on a treadmill at a slow speed, simultaneously capturing video and actual foot pressure measurements. They used a free pose-estimation tool (MediaPipe Pose) to extract body position and movement features from the video, then trained machine learning models to predict foot pressure from those features. The best model explained about 76–77% of the variation in foot pressure during cross-validation within the same recording session.
However, the system struggled when tested on people it had never seen before, performing worse than simply guessing the average pressure for every frame. This is a significant limitation for real-world use. The researchers found that giving the system just 30 seconds of actual pressure data from a new person — as a brief calibration step — dramatically improved its predictions for the rest of that person's session, bringing the explained variance up to roughly 36–44%. This means the system could potentially be used in practice with a short initial calibration period, after which the video alone could drive the estimates.
This research suggests that low-cost video-based tools could one day make gait and foot-loading analysis more accessible outside of specialized laboratories, for example in rehabilitation or sports settings. However, the current findings are limited to slow treadmill walking in healthy adults, and the poor generalization to new individuals without calibration highlights that further development is needed before such a system could be reliably used across diverse populations or real-world walking conditions.
Check Your Own Numbers
Upload your bloodwork. We'll cross-reference your results against this study and 4,700 others.
Jiao Y, Huang G, Huang X, Bian Y, Wu B. (2026). A Low-Cost Monocular Video-Based Framework for Bilateral Plantar Load Estimation During Treadmill Walking.. Sensors (Basel, Switzerland). https://doi.org/10.3390/s26196307