A pear orchard scene segmentation based on conditional random fields (CRFs) was proposed. The CRFs modeled posterior probabilities directly, and had an ability to fuse context information of images. Therefore, it was a suitable method to solve images segmentation of the pear orchard scene whose structures are often very complicated. Firstly, labeled images of the pear orchard scene were segmented into superpixels, and feature vectors of the superpixels and their corresponding labels were integrated into a label database as training samples. Secondly, unlabeled images of the pear orchard scene were also segmented into the superpixels, and their features and spatial relationships between these unlabeled superpixels were modeled by using the CRFs. Moreover, parameters of the CRFs model were obtained by taking the label database as the training samples. Finally, labels of the unlabeled superpixels were inferred through the maximum posterior marginal (MPM) algorithm. The experimental results showed that the proposed algorithm could provide more accurate segmentation results of the pear orchard scene compared with the mutual K-nearest neighbor method (MKNN). 