This paper presents the experimental validation of a new method for mobile robot global self-localization in unstructured environments, i.e. that does not need any beacons or other artifacts structuring the environment. The method resorts to a PCA-based positioning sensor, filtered in a Bayesian probabilistic grid and combined with linear Kalman filters to estimate the global pose of mobile robots. In the implemented system, the information of the environment is captured only with onboard sensors installed in a differential drive robot: encoders, compass, and 2D depth sensor pointed to the ceiling. The use of PCA in a Bayesian probabilistic grid allows to fuse the highly compressed PCA database information, obtained with the low computational effort, in an environment where repetitive scenarios can occur. To avoid the negative impact in the localization estimate caused by the corrupted data existing in the 2D depth sensor, an extension to the classic PCA algorithm is suggested. Thus, the proposed method allows the self-localization of mobile robots in indoor environments with bounded accuracy and working in a wide range of illumination conditions.