Walking gait features extraction and characterization using wearable devices

New wearables devices are introduced with novel options for observing personal transport and mobility in indoor and outdoor spaces. This hardware includes low cost MEMS sensors: accelerometer, gyroscope, magnetometer, which provide continuously available data contrary to existing solutions that are based on radio signals. In order to mitigate the propagation of sensor errors in the position estimate, a pedestrian dead reckoning strategy is commonly adopted. The processing requires parametric step length models relying on some physiological parameters, displacement features and acceleration statistical properties. The coefficients of these models need frequent adjustment to limit cumulative errors induced by alteration of gait pattern. A large experimental database providing information about human locomotion variability is required for this calibration. However, the development of such database is costly in terms of time and effort, and several gait-affecting factors should be considered, which highly increases the number of measurement trials. In this thesis, we propose an alternative way of generating locomotion data that consists in simulating human gait motion under different conditions. In this scope, a 3D multibody system simulator based on parametric optimization technique was developed, and improvements were made throughout this work to get a more realistic walking motion prediction. Joint trajectories during one step were optimized by minimizing an energy criterion based on actuated torques. Validation with inertial data from overground walking experiments on one healthy subject showed an asymmetry in experimental acceleration signals from one step to the next. This suggests that asymmetric movement are likely to result in a step-level asymmetry of displacement features. This defeats the general assumption in PDR strategy: the presence of a device in hand does not impact the gait symmetry and all steps are identical for a fixed walking speed. This hypothesis is investigated with motion capture experiments with several subjects, designed to study the influence of a mass carried in hand on the walking gait cycles. Analysis of variance tests have shown that the presence of a mass in hand changes the gait symmetry at the step level, and then the proposed optimization process is extended to the stride level in order to allow observing asymmetric acceleration patterns. Overall, our simulator reproduces similar fundamental patterns of walking, and the same variation trend of acceleration related items found in experiments. However, it shows limitations when predicting acceleration data related to the hand, due to some modeling assumptions and numerical issues. Therefore, our simulation approach partially solves for the direct modeling problem in pedestrian navigation, and improvements on model assumptions are foreseen to predict acceleration signals over a complete gait cycle more reliably. Keywords: Handheld devices, MEMS sensors, Multibody system, Parametric optimization, Step-level symmetry, Walking gait patterns, Human motion analysis, Analysis of variance