Étude de l'application des contraintes géométriques et spatiales à la localisation piétonne en intérieur

In indoor localisation, the use of a proprioceptive signals coming from an inertial measurements unit carried by the pedestrian offers a first estimate of the position and heading of the pedestrian. This estimate quickly becomes erroneous because of the accumulation of errors from noisy sensors data.This thesis is divided into two parts. In the first, the study of spatial constraints aims to incorporate the map of a building with an indoor localisation system in order to limit the estimation error. This part proposes the representation of human motion likelihood in interior space in the form of a grid. This representation has the effect of reducing the uncertainty on the position and increasing the integrity of the system.The second part of this thesis concerns the introduction of an exteroceptive signal to correct the estimation. This signal is a detection of landmark in an image by a neural model. This part therefore proposes the methodology for adapting the neural network to the problem of localisation by defining the needs, the hypotheses, the tools and the methods necessary to obtain this signal. Finally, it proposes a probabilistic management of cases where the neural model produces false positives.During each step of its construction, the proposed system is validated by a series of unitary tests and experiments conducted to demonstrate the impact of the propositions of this thesis.