To improve the autonomous vehicle's ability to understand the scene of unstructured road driving regions, a semantic segmentation method of unstructured road for autonomous vehicle based on SegNet architecture is proposed. Deep convolutional encoder-decoder architecture is formed by traditional convolutional neural networks, and it can learn the feature map of unstructured roads automatically. By learning and training in the datasets, image semantic segmentation model can be acquired and used to predict the feasible driving area of unstructured roads directly, which is important for autonomous vehicle's scene understanding. The proposed approach outperforms in precision and segmentation consequent, while Dice coefficient reaches more than 80%.