Spatio-Temporal Analysis During Running Using Magneto-Inertial Sensors: Optimization of Foot Orientation Estimation.

Running is a widely practiced activity providing significant health benefits but also potential injury risks. A thorough biomechanical analysis is essential for optimizing performance and preventing running-related injuries. In running characterization, stride length (SL) and stride velocity (SV) are key running spatio-temporal parameters quantifying foot displacement and velocity within a stride. These parameters can be assessed in indoor and outdoor conditions by wearable magneto-inertial measurement units (MIMUs) through the adoption of ad-hoc algorithms. The most direct approach to estimate spatial-related parameters is to double integrate foot accelerations after gravity removal. Effective gravity removal requires to accurately estimate the foot orientation which can be determined through a sensor fusion algorithm (SFA) using (magneto-)inertial signals. However, to work well, these algorithms require extensive fine-tuning of different parameters and often lack robustness to variations in highly dynamic movements like running. In this thesis, a comprehensive analysis of four different SFAs was conducted to assess their impact on SL and SV estimation and their robustness to changes in their main parameters and in running speed. Then, a novel framework for the automatic stride-by-stride setting of SFA parameters was proposed to handle different running conditions and speeds. The proposed method minimized a cost function based on errors on SL and SV estimates with respect to an available reference, imposing biomechanical constraints and measurement consistency. A total of 20 participants were enrolled to build two datasets at different speeds. The first dataset included both treadmill and overground running at 8-10 km/h with pressure insoles used as reference, while the second one included treadmill running at 14km/h with a stereo-photogrammetric system used as reference. The parameter setting of each SFAs was speed-dependent. Thus, each SFA fine-tuned to select a fixed optimal value of its parameter(s) minimizing SL errors across speeds. The best performance was achieved by the SFA proposed by Madgwick et al. (2011), which resulted in SL errors of 1.6% and 2.5%, and SV errors of 3.5% and 2.6% at 8-10 km/h and 14 km/h, respectively. The proposed framework for stride-by-stride selection of Madgwick’s parameter provided better or comparable results with the previous approach using a fixed Madgwick’s parameter (SL error of 1.8% and 1.6%, and SV error of 0.6% and 0.8% at 8-10 km/h and 14 km/h, respectively). This study demonstrated that using a SFA with a fixed optimized parameter can provide accurate SL and SV estimates. However, the setting of SFA parameters can be speed-dependent and usually influenced by the specific characteristics of the used inertial sensors. Conversely, the proposed framework is a promising solution to automatically adjust SFA parameters regardless of speed, hardware, and running conditions, being suitable for running analysis in different and variable scenarios.