SMPL-Based Estimation of Gait Joint Angles for Rehabilitation Applications

SMPL-Based Estimation of Gait Joint Angles for Rehabilitation Applications Beatrice Lagomarsino DIBRIS University of Genoa Genoa, Italy [email protected]  Maura Casadio DIBRIS University of Genoa Genoa, Italy [email protected] Matteo Moro DIBRIS University of Genoa Genoa, Italy [email protected] Ronald Poppe ICS Utrecht University Utrecht, The Netherlands r.w[email protected] Abstract—Gait analysis is a key tool in rehabilitation, but both marker-based and markerless systems require biomechanical modeling to derive standardized kinematic variables. This process is computationally demanding, time-consuming, and dependent on expert supervision, while the sparse keypoints typical of markerless pipelines further limit their applicability. To address these challenges, we introduce a unifying approach based on the Skinned Multi-Person Linear (SMPL) model, which provides a dense, anatomically consistent body representation from different input modalities. Based on this representation, a neural network was trained to directly predict clinically relevant joint angles, eliminating the need for traditional biomechanical software. Evaluation on the AMASS dataset demonstrated errors within clinically acceptable ranges, supporting the feasibility of SMPLbased pipelines as robust, scalable, and clinically translatable tools for gait assessment in rehabilitation. Index Terms—Biomechanics, Gait Analysis, Markerless Motion Capture, Neural Networks, Rehabilitation, SMPL I. INTRODUCTION Gait is a fundamental motor function often impaired by pathological or traumatic events, such as neurological, orthopedic, or degenerative conditions. Accurate gait assessment is therefore essential in clinical and rehabilitative practice, as it enables monitoring of functional recovery, supports personalized intervention planning, and provides quantitative metrics to evaluate therapeutic effectiveness [1]. Gait can be analyzed using marker-based systems, in which reflective markers are tracked by multiple cameras, or markerless systems, which estimate body motion directly from video recordings. Although different in acquisition, both approaches ultimately provide a set of body points that must be processed through biomechanical modeling software, such as OpenSim, in order to extract standardized joint kinematics and clinically meaningful variables [2]. This processing step is timeconsuming, computationally demanding, and requires expert knowledge, which limits the adoption of gait analysis outside controlled laboratory environments. This challenge is even greater for markerless methods, as the limited number of This work was funded by the European Union—NextGenerationEU and by the Ministry of University and Research (MUR), National Recovery and Resilience Plan (NRRP), Mission 4, Component 2, Investment 1.5, project “RAISE—Robotics and AI for Socio-economic Empowerment” (ECS00000035). detected keypoints further reduces the reliability and applicability of biomechanical modeling [2]. This study aims to address these limitations by proposing a unifying solution. The Skinned Multi-Person Linear Model (SMPL) [3], which offers a compact representation of body pose, shape, and motion, provides a dense and anatomically consistent mesh of the human body that can be derived from both marker-based and markerless inputs. Building on this common representation, we introduce a neural architecture that transforms raw pose parameters obtained from different input modalities into a comprehensive set of clinically meaningful joint angles. Our goal is to assess the feasibility of this approach and evaluate its potential as a scalable, generalizable, and clinically translatable tool for human movement analysis. II. MATERIALS AND METHODS The overall ideal workflow is illustrated in Figure 1. Fig. 1. Proposed workflow. Input data are converted into SMPL parameters. The model predicts joint angles (solid blue line), while OpenSim-derived ground truth is used only for training and test (black dashed lines). •Input sources: Human motion can be captured using either a stereophotogrammetric system or video data (RGB or RGB-D). •SMPL estimation: Data are converted into SMPL parameters, providing a compact and standardized representation of body pose and shape. 2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629778 147 •Neural network: The architecture takes the SMPL parameters as input and predicts the corresponding rehabilitation-related outputs. •Output: The system produces clinically meaningful parameters to support standardized and interpretable assessment in rehabilitation contexts. Dataset and Pre-processing We used the AMASS dataset, which contains multiple walking trials captured with stereophotogrammetric systems. Each trial includes both video recordings and synchronized SMPL parameters. We selected 10 participants, each performing several walking trials. The stereophotogrammetric data were processed using the OpenSim biomechanical software to extract reference values for four key gait joint angles (GT Joint Angles, Fig. 1): knee flexion, hip flexion, hip abduction, and ankle dorsiflexion. Network Architecture, Training, and Data Splits The neural network was implemented in Keras. The final architecture consisted of five fully connected layers with 512, 512, 256, 256, and 256 neurons, each followed by a ReLU activation function, and a linear output layer. The Adam optimizer was used with a learning rate of 0.0001, a batch size of 32, and training was monitored using mean absolute error (MAE) and mean squared error (MSE). Early stopping was applied to prevent overfitting. Hyperparameters were explored with Bayesian optimization using Keras Tuner, and the best configuration was selected based on validation loss. The dataset was initially divided into 70% training, 15% validation, and 15% testing. The validation set was used for hyperparameter tuning and early stopping, while the test set was used only for final evaluation. Cross-Validation and Evaluation Metrics In addition to the fixed split, we performed a leave-2-subjects-out crossvalidation to further assess robustness and subject-independent generalizability. All 45 possible pairs of participants were used, with the remaining 8 subjects serving as training data in each fold. Model performance was evaluated by comparing the predicted values with the OpenSim-derived ground truth using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the correlation coefficient for each joint angle. III. RESULTS AND CONCLUSIONS The cross-validation procedure demonstrated an acceptable performance of the model across all analyzed joints (Table I and Figure 2). TABLE I CROSS-VALIDATION RESULTS:MEAN AND STANDARD DEVIATION OF THE ERROR BETWEEN THE ESTIMATED AND GROUND-TRUTH JOINT ANGLES (MAE, RMSE, AND CORRELATION COEFFICIENT). ANGLE MAE (°) RMSE (°) CORRELATION Knee Flexion 3.45±0.60 3.80±0.70 0.90±0.05 Hip Flexion 3.20±0.55 3.60±0.65 0.93±0.04 Hip Adduction 3.70±0.50 4.10±0.60 0.90±0.06 Ankle Dorsiflexion 3.95±0.75 4.32±0.80 0.87±0.07 Specifically, the Mean Absolute Error (MAE) ranged from 3.20° for hip adduction to 3.95° for ankle dorsiflexion, while the Root Mean Square Error (RMSE) ranged from 3.60° to 4.32°. Correlation coefficients between the predicted joint angles and the OpenSim-derived ground truth were consistently Fig. 2. Comparison between predicted (blue) and ground-truth OpenSimderived (black) joint angle trajectories for one representative subject. high, exceeding 0.85 for all joints and reaching 0.93 for hip flexion. These results confirm that the model can capture the temporal dynamics and waveform characteristics of joint angle trajectories with fidelity (Fig 2.). Among the analyzed joints, ankle dorsiflexion showed slightly higher prediction errors than the others. This outcome is not unexpected, as this movement involves complex interactions between multiple anatomical segments and is known to exhibit substantial inter-subject variability. Nevertheless, even for this joint, the prediction accuracy remained within a range that could be considered acceptable for clinical applications, suggesting that the model might be able to generalize across different gait patterns and individual anatomical characteristics [4]. Overall, these preliminary results suggest that, by leveraging the SMPL representation, our approach can automatically transform raw pose parameters obtained from different input modalities (marker-based or markerless) into interpretable and standardized joint kinematics, thereby reducing common limitations of current gait analysis methods. This preliminary framework represents a promising step toward scalable and accessible gait analysis tools with potential impact on rehabilitation practice, enabling routine, quantitative, and objective assessments even with low-cost and widely available recording devices. Furthermore, the proposed framework could be integrated with robotic rehabilitation technologies, such as lower-limb exoskeletons, where the estimated joint kinematics may serve as reference trajectories or adaptive control signals to support more personalized and responsive assistance. An additional consideration for clinical translation concerns patient data privacy. While this study used publicly available datasets, real-world deployment would involve sensitive patient data and thus require strict compliance with privacy regulations (e.g., GDPR), including anonymization, secure storage, and possibly on-device processing. Future work will focus on expanding the dataset to include a more diverse population and pathological conditions, as well as investigating how different SMPL parameters (pose, shape, and dynamics) contribute to prediction accuracy, to guide the design of more specialized and efficient architectures. REFERENCES [1] M. Moro, G. Marchesi, F. Hesse, F. Odone, and M. Casadio, “Markerless vs. marker-based gait analysis: A proof of concept study,” Sensors, vol. 22, 3 2022. [2] R. J. Cotton, “Posepipe: Open-source human pose estimation pipeline for rehabilitation research,” Archives of Physical Medicine and Rehabilitation, 2022. [3] M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black, “Smpl: a skinned multi-person linear model,” ACM Trans. Graph., 2015. [4] J. Mcginley, R. Baker, R. Wolfe, and M. Morris, “The reliability of threedimensional kinematic gait measurements: A systematic review,” Gait posture, 2008. 148