Region covariance descriptor has been employed in many computer vision applications such as texture classification, object detection and object tracking. It provides a natural way of fusing multiple features based on a set of pixels of a given region. However, the discriminative capacity of covariance descriptor can vary dramatically regarding different combination of feature sets that are fused and thus gives rise to the problem of discriminative feature selection when given a specific application. In this work, we propose a PCA-based feature selection approach in the construction procedure of the covariance descriptor. We show that covariance descriptor computed in a minutely-learned subspace can be adaptive to a specific target and thus results in a more compact and potentially more discriminative descriptor. Comparing experiments on real world video sequences demonstrate superior representational ability of the proposed method with respect to the conventional region covariance descriptor.