Individualized continuous lower-limb joint kinematics modeling enhanced by discrete cosine transform

• Published in: Robotics and autonomous systems Vol. 195; p. 105226
• Main Authors: Wu, Xiaoguang, Dong, Zhihui, Niu, Xiaochen, Lin, Hongbin, Du, Yihao
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.01.2026
• Subjects: Discrete cosine transform; Fourier series; Gait trajectory; Gaussian process regression; Least squares-support vector machine; Low limb prosthetic; Gaussian process regression; Low limb prosthetic; Gait trajectory; Fourier series; Least squares-support vector machine; Discrete cosine transform
• ISSN: 0921-8890
• DOI: 10.1016/j.robot.2025.105226

Effective joint trajectory planning is critical in prosthetic control systems, as coordinated, continuous locomotion trajectories not only improve wearer comfort but also facilitate the restoration of normative biomechanics across a wide range of locomotor tasks. This paper proposes a novel approach to predict joint trajectories for continuous locomotion tasks. First, Fourier series are used to fit the multi-cycle gait data, addressing the potential phase shift issues and mitigating the influence of outlier data. Next, the human locomotion data are transformed into Discrete Cosine Transform Coefficients (DCTCs) via the Discrete Cosine Transform (DCT). Subsequently, different regression models—including Gaussian Process Regression (GPR) and Least Squares - Support Vector Machine (LS‑SVM)—are then trained to capture the relationship between each DCTC and its associated task. This combination of DCT with either GPR or LS-SVM can significantly reduce the computational complexity, thus improving computational speed and prediction accuracy. Furthermore, based on the consistency of the difference between individual joint locomotion trajectories and inter-subject mean trajectories at the same incline, we propose a gait trajectory personalization method to enable kinematic models to match individual joint kinematics. Using a publicly available gait dataset with multiple subjects, we validate the effectiveness of the proposed approach. The results show that the GPR‑based model yields lower root‑mean‑square error (RMSE), while the LS‑SVM–based model produces fewer clusters of data points with significant deviations from the true data. The personalized model can efficiently and accurately predict gait trajectories over continuously varying tasks, even with limited gait data.