Accurate navigation capabilities can help escort robots interact better with humans.Traditional navigation algorithms are usually used in static environments. When dynamic objects appear in indoor environments, the matching of dynamic points on the object can interfere with the camera pose and leave stains on the map. Although semantic segmentation methods can remove potential dynamic objects, dynamic objects will not be detected for some movable objects, such as books held by humans. In this regard,a Dynamic SLAM framework based on human Contact Experience (CE-DSLAM) is proposed for escort robots. This framework effectively adapts to indoor dynamic settings and can predict the actual status of certain movable objects based on contact experience. We assume that the movement of movable objects is caused by humans, and propose an adaptive frame strategy for detecting human masks. For information from multiple consecutive frames, we use effective contact to update the prior level of indoor movable objects. This article first assigns an initial level to all indoor objects based on human experience, and enters the adaptive frame thread based on whether a person’s mask is detected or not. Then, the dynamic level of indoor movable objects is updated based on the experience of contact between humans and movable objects. When the next frame of the image arrives, the high dynamic level feature points will be removed, and the remaining static points will be used for localization. We evaluated CE-DSLAM on the TUM RGB-D dataset and compared it with the DynaSLAM system in a dynamic environment. The experimental results indicate that CE-DSLAM can operate stably in dynamic environments and can be well applied in indoor environments.