Abstract:  Visual SLAM has been the hot research topic in recent years, which treats visual image as processing objects. Deep learning shows the prominent advantages in image processing, which makes it possible to combine the visual SLAM and deep learning. The characteristics and properties of traditional SLAM and SLAM based on deep learning are summarized. The prominent achievements on visual odometry and loop closure detection incorporated with deep learning are introduced. The future research directions of advanced SLAM based on deep learning are discussed.