Exploring human lower limb kinematics in walking: overcoming challenges with minimal inertial sensor configuration using a robotic modeling approach.

Estimating lower limb joint kinematics is crucial for assessing movement disorders,optimizing rehabilitation, and monitoring sports performance. Traditional clinical setups require marker-based systems with at least three markers per segment. Wearable IMUs offer an alternative by placing sensors on proximal and distal segments but remain bulky, time-consuming, and costly. This highlights the potential of a minimal sensor configuration. Selecting instrumented segments requires careful consideration. Prioritizing the pelvis, representing center-of-mass dynamics, and the feet, as movement end-effectors, aligns with established methods for estimating spatio-temporal parameters in free-living conditions. This thesis extends previous work on a biomechanical model that estimates joint angles using anthropometric measurements and pelvis/feet orientations and positions. That work addressed measurement errors by applying constraints and an optimization framework to fit segment data while limiting hip, knee, and ankle motion. The model employs the Denavit-Hartenberg convention for standardized joint definitions and ISB-recommended rotation sequences. The SQP optimization algorithm minimizes differences between model-derived and sensor data. This thesis aimed to refine the model using marker-based stereophotogrammetry (SP) data as a reference to evaluate the feasibility of estimating joint kinematics with minimal input. The improved model was tested on a healthy subject under different speeds and simulating toe walking and asymmetric steps (PoliTO dataset). After that, a validation was performed in more complex scenarios (Madrid dataset) featuring uneven terrains: flat, zigzag, sponge-like, and irregular surfaces, also assessed on a mechanized tilting platform. During model refinement, fixed anthropometric measurements introduced errors due to marker trajectory uncertainties and soft tissue artifacts. Frame-by-frame adaptation of segment lengths yielded Root Mean Square Error (RMSE) values of 1.9 deg, 4.2 deg, and 2.9 deg for hip, knee, and ankle respectively, comparing estimated joint angles with those from SP. Using fixed lengths resulted in additional errors of 1.8 deg, 3.2 deg, and 0.4 deg for the same joints. Moreover, joint angle plots exhibited saturation. RMSE values on the Madrid dataset were: 1.4 deg (1.3 deg) for right (left) hip, 5.0 deg (3.9 deg) for right (left) knee, 5.3 deg (6.2 deg) for right (left) ankle. Although limited differences were found across terrains, the presence of slope slightly decreased accuracy (errors up 0.7 deg). Range of Motion (ROM) was computed as the difference between maximum and minimum joint angles over each gait cycle, then averaged across all cycles. A one-way ANOVA (α=0.05) on ROM residuals showed significant differences for hip (p=0.0002) and knee (p=0.03), but not ankle (p=0.31), indicating asymmetric model estimation. Thus, averaging ROM values is not appropriate.A two-way ANOVA (α=0.05) revealed no significant influence of both terrain conditions and slope on ROM estimation (p=0.81).This study demonstrated the feasibility of a minimal configuration for gait kinematics estimation, with optimization compensating for measurement errors. Integrating anthropometric variability improved accuracy, though asymmetry and experimental uncertainties persist. Since the ultimate goal remains the use of fixed measurements, the next step will be modeling segment length uncertainty within the optimization to enhance robustness, supporting IMU-based applications.