This thesis is devoted to the definition and the implementation of multi-sensor systems which are embedded on a walker and dedicated to the capture and estimation of the posture and motion of the user during locomotion. Several experiments were conducted to evaluate quantitatively the quality of resulting reconstructed posture. An anomaly detector during walking has also been proposed to evaluate the system in terms of application. In this thesis, the human motion parameters related to walking (step, posture, balance) generally used for gait analysis, are first introduced. In order to observe the parameters, doctors use different devices. Most of devices are bulky, which limit their uses only in hospitals. The multi-sensor systems embedded on a walker we proposed would provide daily data about longer distance of locomotion and duration. Two architectures have been developed in succession: the first one consists of a 3D camera and two infrared sensors, and the second one is composed of two Kinects. These architectures are used to estimate the posture of the subject and its locomotor activity by fitting on a 3d human body model. The accuracy of the position obtained is improved by integrating a motion prediction module: it uses an estimate of some discrete parameters of gait (e.g. step period and step length) and an assisted walking model. Finally, the development of an algorithm for rupture detection of locomotor rhythm allows us to validate the overall approach according to the final application, which is the diagnosis support. From a set of distance measure corresponding to the Generalized Likelihood Ratio (GLR) calculated on the relative positions of each joint and the speed of the walker determined by an odometer, a multi-dimensional model of a regular walking is then learnt through the One-Class Support Vector Machines (One-Class SVM). This algorithm has been evaluated by using the protocol of walking rhythm change. In the future, this system can be integrated in the robot assisted walker designed within the framework of the ANR project MIRAS (Multimodal Interactive Robot in Support of Strolling).
