Recent research has shown that better recognition performance can be attained through representing symmetric positive definite matrices as points on Riemannian manifolds for many computer vision tasks. However, most existing algorithms only approximate the Riemannian manifold locally by its tangent space and are incapable of scaling effectively distribution of samples. Inspired by kernel methods, a novel method, called local linear coding based on Riemannian kernel (LLCRK), is proposed and applied successfully to vision classification issues. Firstly, with the aid of recently introduced Riemannian kernel, symmetric positive definite matrices are mapped into the reproducing kernel Hilbert space by kernel method and a mathematical model of sparse coding and Riemannian dictionary learning is constructed by local linear coding theory. Secondly, an efficient algorithm of LLCRK is presented for dictionary learning according to the convex optimization methods. Finally, an iterative updating algorithm is constructed to optimize the objective function, and the test samples are classified by nearest neighbor classifier. Experimental results on three visual classification data sets demonstrate that the proposed algorithm achieves considerable improvement in discrimination accuracy.