The vehicle driving cycle was taken as a stochastic Markov process with time, which was divided into 6 different states by maximum likelihood estimation (MLE). And every driving segment, as the modal event reflecting acceleration variations, finely preserved the speed variation characteristic in tiny time segments of the vehicle driving process. Therefore, the selection approach of candidate cycles was put forward by combination of the states transition probabilities with similarity tests, and the new evaluation criteria with 13 characteristic parameters was established. The actual application analysis was carried out with an example of typical roads driving cycles in Hefei and the independent test was performed to validate the rationality of the Markov method for the driving cycle. The results showed that the similarity levels of the Markov method, on the K-S test of modal events acceleration distribution with experimental data, were 0.998, 0.989, 0.994, contrasting the traditional method, 0.788, 0.036, 0.992. Resultingly, the Markov method more accords with driving cycles of real roads.