The current blind road segmentation algorithms are used to segment by extracting the color or texture feature, and by using the clustering method, which are vulnerable to blind road types and the external environment. To solve this problem, a learning approach is introduced, and a blind road segmentation method is proposed based on saliency detection and improved projection dictionary from the consideration of the global feature of blind road. Firstly, the saliency detection algorithms are used to roughly locate the blind road region. Then the image piece is used as the processing unit, and the dictionary is learned through the robust projective dictionary pair learning proposed. And the coarse image after location is divided into blocks to sparsely reconstruct on the dictionary. Finally, the rough-positioned images are reconstructed on the dictionary, and classified according to the reconstruction error to achieve the purpose of segmentation. Experimental results show that the proposed algorithm performs better than the existing algorithms in terms of accuracy and universality in blind road segmentation.