In order to realize autonomous localization and mapping of mobile robots in dynamic scenes,solve the problems of accumulated errors and poor tracking effect of traditional visual odometer,and improve the accuracy and robustness of loop closure detection,a method of simultaneous localization and mapping fusion of deep learning was proposed.Quadtree algorithm was used to homogenize feature distribution to solve the problem of feature aggregation in dynamic scenes.The dynamic features of scenes were recognized by optimized target detection network,and the interference of dynamic objects for pose estimation was eliminated.The spatial structure information of the scene was fully extracted,and the pose tracking and loop closure detection were realized by combining point features and line features,and the global consistent environment map was constructed.TUM dataset and real scene experimental results show that the improved method improves the accuracy and robustness of mobile robot localization and mapping.