For the defect that traditional path deviation detection algorithms apply only to structured roads, a kind of deviation control algorithm for unstructured roads based on active discrimination dictionary learning is proposed. The active discrimination dictionary is obtained from unlabeled samples by using the active leaning algorithm. According to its reconstruction error, the deviate direction of vehicles could be clear and as a basis for the next step direction control for vehicles. The definition of price weighted error is used to measure the correcting ability of the proposed algorithm. The experimental results show that the proposed algorithm has strong validity and reliability for structured and unstructured environments.