The Ground Penetrating Radar (GPR) consists in an electromagnetic signal which is transmitted at different positions through the ground in order to obtain an image of the subsoil. In particular, the GPR is used to detect buried objects like pipes. Their detection and localisation are intricate for three main reasons. First, the noise is important in the resulting image due to the presence of several rocks and/or layers. Second, the wave speed and the response of the pipe depend on the characteristics of the different layers. Finally, the signal attenuation could be important because of the depth of pipes. In this paper, we propose to derive an adaptive detector where the steering vector is parametrised by the wave speed in the ground and the noise follows a Spherically Invariant Random Vector (SIRV) distribution in order to obtain a robust detector. To estimate the covariance matrix, we propose to use robust maximum likelihood-type estimators called M-estimators. To handle the large size of data, we consider regularised versions of such M-estimators. Simulations will allow to estimate the relation Probability of False Alarm (PFA)-Threshold. Application on real datasets will show the relevancy of the proposed analysis for detecting buried objects like pipes.