This PhD thesis, written in english, studies the non-Gaussianity (NG) of extragalactic foregrounds to the Cosmic Microwave Background (CMB), the latter being one of the golden observables of today's cosmology. In the last decade has emerged research for deviations of the CMB to the Gaussian law, as they would discriminate the models for the generation of primordial perturbations. However the CMB measurements, e.g. by the Planck satellite, are contaminated by several foregrounds. I studied in particular the extragalactic foregrounds which trace the large scale structure of the universe : radio and infrared point-sources and the thermal Sunyaev-Zel'dovich effect (tSZ). I hence describe the statistical tools to characterise a random field : the correlation functions, and their harmonic counterpart : the polyspectra. In particular the bispectrum is the lowest order indicator of NG, with the highest potential signal to noise ratio (SNR). I describe how it can be estimated on data, accounting for a potential mask (e.g. galactic), and propose a method to visualise the bispectrum, which is more adapted than the already existing ones. I then describe the covariance of a polyspectrum measurement, a method to generate non-Gaussian simulations, and how the statistic of a 3D field projects onto the sphere when integrating along the line-of-sight. I then describe the generation of density perturbations by the standard inflation model and their possible NG, how they yield the CMB anisotropies and grow to form the large scale structure of today's universe. To describe this large scale structure, I present the halo model and propose a diagrammatic method to compute the polyspectra of the galaxy density field and to have a simple and powerful representation of the involved terms. I then describe the foregrounds to the CMB, galactic as well as extragalactic. I briefly describe the physics of the thermal Sunyaev-Zel'dovich effect and how to describe its spatial distribution with the halo model. I then describe the extragalactic point-sources and present a prescription for the NG of clustered sources. For the Cosmic Infrared Background (CIB) I introduce a physical modeling with the halo model and the diagrammatic method. I compute numerically the 3D galaxy bispectrum and produce the first theoretical prediction of the CIB angular bispectrum. I show the contributions of the different terms and the temporal evolution of the galaxy bispectrum. For the CIB angular bispectrum, I show its different terms, its scale and configuration dependence, and how it varies with model parameters. By Fisher analysis, I show it allows very good constraints on these parameters, complementary to or better than those coming from the power spectrum. Finally, I describe my work on measuring NG. I first introduce an estimator for the amplitude of the CIB bispectrum, and show how to combine it with similar ones for radio sources and the CMB, for a joint constraint of the different sources of NG. I quantify the contamination of extragalactic point-sources to the estimation of primordial NG ; for Planck it is negligible for the central CMB frequencies. I then describe my measurement of the CIB bispectrum on Planck data ; it is very significantly detected at 217, 353 and 545 GHz with SNR ranging from 5.8 to 28.7. Its shape is consistent between frequencies, as well as the intrinsic amplitude of NG. Ultimately, I describe my measurement of the tSZ bispectrum, on simulations and on Compton parameter maps estimated by Planck, validating the robustness of the estimation thanks to realist foreground simulations. The tSZ bispectrum is very significantly detected with SNR~200. Its amplitude and its scale and configuration dependence are consistent with the projected map of detected clusters and tSZ simulations. Finally, this measurement allows to put a constraint on the cosmological parameters : sigma_8*(Omega_b/0.049)^0.35 = 0.74+/-0.04 in agreement with other tSZ statistics.