Abstract : A large number of cosmological and astrophysical measurements supports the fact that our galaxy should be immersed in a halo of non-baryonic dark matter. Directional detection aims at measuring the direction of recoiling nucleus following an elastic scattering with a Dark Matter particle. This should allow us to show the expected strong angular dependence of the recoil distribution due to the rotation of the Solar System around the galactic center. This thesis presents a comprehensive study of directional detection following three different aspects: phenomenology, experimental and data analysis. The goal of the phenomenological studies is to explore the interest of directional detection in terms of galactic dark matter search. With the use of dedicated statistical tools we show that a detector, as the one proposed by the MIMAC collaboration, should be able to discover dark matter with a high significance down to cross sections 2 to 3 orders of magnitude below current limits. Setting up a new strategy of data analysis is a second goal of this thesis as an efficient 3D track reconstruction is compulsory to achieve an accurate directional detection of dark matter. We present a new method based on a likelihood approach aiming at the optimisation of the estimation of the parameters of each measured track: position in the detector volume and direction. In the context of reducing the electronic background, we present a method based on the analysis of the track topology using a boosted decision tree algorithm which enhances the rejection power by a factor 20 with respect to a sequential analysis. In the context of experimental measurement, we present a new method dedicated to the measurement of the electron drift velocity with uncertainties of the order of the percent and constraining the longitudinal diffusion coefficients. Eventually, we discuss the results of a data analysis obtained during an acquisition using a neutron field which validate the detection strategy of the MIMAC experiment.