A Novel Method for Distinguishing Indoor Dynamic and Static Semantic Objects Based on Deep Learning and Space Constraints in Visual-inertial SLAM.