Vision-based road detection is a key component for autonomous vehicle. Existing techniques could be roughly categorized into two categories: off-line training based algorithms and on-line learning based algorithms. While off-line training based algorithms may not adapt well to the new testing scenario, on-line learning based algorithms may not produce robust results. In this paper, we present a method that combines the merits of both off-line and on-line algorithms. Firstly, we get the likelihood image using road and background detectors based on mixture models. Then, the likelihood image is combined with the result generated by classifier which is trained using off-line booting. And the graph cut segmentation will be performed to get an accurate road region. Experiments on road sequences of unstructured road show that the proposed method provides high road detection accuracy when compared to state-of-the-art methods.