With the continuous development of machine vision, visual sensors has advantages of lightweight and low cost. Thus, visual simultaneous localization and mapping(VSLAM)is attracting moreand more attention and becoming a research hotspot. Deep learning has provided new methods and ideas to deal with VSLAM problenns. This article reviews the deep learning-based VSLAM methods in recent years. Firstly, the development history of VSLAM is reviewed, and the basic principle and composition structure of VSLAM are systematically explained. Then, various methods based on deep learning are summarized and analyzed from three aspects, including visual odometry(VO),loop closure detection and mapping. The application of deep learning in visual odometry is described in three parts,which are feature extraction and feature matching,depth estimation and pose estimation and keyframes selection. Based on the different manner of scene representation, deep learning-based methods in geometric mapping, semantic mapping and general mapping are summarized. Thirdly, it introduces various datasets and performance evaluation metrics commonly used in VSLAM at present. Finally, the challenges of VSLAM are pointed out, and the future research trends and development directions of combining deep learning with VSLAM are forecasted.