The lack of much labeled data in the real world affects the training of supervised model for pedestrian re-identification. Besides, applications of pedestrian re-identification in pedestrian retrieving and criminal tracking are limited because of the lack of semantic representation. This paper presents a pedestrian re-identification method based on the combination of deep learning and attributes learning, which extracts essential features with unsupervised deep learning model and enhances the semantic representation of features with ‘attributes’. Firstly, a convolutional auto-encoder (CAE) is used to extract features of unlabeled pedestrian images, and the extracted features are then input into several attribute classifiers to judge whether the pedestrian owns the attributes. Lastly, with a table of ‘attributes-classes mapping relations’, we can get the final dassification result. Tests of the proposed algorithm and comparisons with other algorithms on the VIPeR and i-LIDS datasets are shown, and results prove that our algorithm indeed strengthens the semantic representation and improves the accuracy of pedestrian re-identification, achieving good ‘zero-shot’ re-identification performance as well.