We establish novel classification rules for heterogeneous populations under the assumption of multivariate t distribution, which can be treated as a robust generalization of the routinely used multivariate normal one. In this new classification framework, the unknown parameters are estimated by the maximum likelihood method via the expectation maximization (EM) algorithm. The proposed classification technique is particularly useful when the collected data unavoidably contain longer than normal tails or outlying observations. Experimental results show that the new classification rules may outperform the traditional classifiers in some scenarios, especially when the underlying distribution of data is far from normal.