Aiming at the shortcomings of low precision and poor robustness of traditional simultaneous localization and mapping（ SLAM） in dynamic environment due to the disturbance of dynamic objects

a robust visual SLAM algorithm based on semantic is proposed. Deep learning is used to construct object detector based on convolutional neural network（ CNN）

combined with prior knowledge

so as to realize the detection of dynamic object on the semantic level. Propose a compensation model for missed detection in adjacent frames based on the invariance of speed

which further improve the detection rate of object detecting network. A feature point-based visual SLAM system is constructed

in tracking line path

feature points on dynamic objects are removed so that the error of pose estimation is reduced. The experiment verifies that in extreme dynamic environment test

the system keeps positioning without losing. In TUM dynamic environment data set test

compared with ORB-SLAM2

the positioning precision is improved by 22. 6 %

and the performance is improved by 10 %.