This paper proposes a novel 3D face recognition method using the local covariance descriptor and Riemannian kernel sparse coding in order to accurately evaluate the intrinsic correlation of the extracted features and further improve the 3D face recognition accuracy. Firstly, the keypoints are detected by the farthest point sampling method, and the corresponding keypoint neighborhood is extracted by the specified radius associated with geodesic distance. Then, different types of the efficient features are selected to construct the local covariance descriptor with inherent property. Finally, the appropriate Riemannian kernel sparse coding is used to identify the faces in probe. Experimental evaluation has been performed on two challenging 3D face datasets, FRGC v2.0 and Bosphorus, which indicates that the proposed approach can significantly improve the identification accuracy comparing with other state-of-the-art methods.