Abstract: Three-dimensional semantic maps play a key role in tasks such as robot navigation,path planning,intelligent grasping and human-computer interaction.So how to construct 3D semantic maps in real time is especially important.The current SLAM (simultaneous localization and mapping) algorithm can achieve higher positioning and mapping accuracy.However,how to eliminate dynamic objects to obtain higher positioning accuracy in a dynamic environment,and to understand the existence of objects and their location information in the surrounding scenes are still not well solved.This paper presents an algorithm for constructing semantic maps in a dynamic environment.This algorithm is improved on ORB-SLAM2.The dynamic and static point detection algorithm is added to the tracking thread to eliminate the feature points detected as dynamic feature points,which improves the positioning accuracy in dynamic environment.Object detection threads are added to detect key images.The mapping threads are added with the Octo-Map dense map construction.At the same time,the 3D object database is constructed according to the detection results.In order to prove the feasibility of the algorithm,the laboratory is used as the test environment,and the object detection,dynamic point detection,3D target information acquisition,and semantic map construction experiments in the dynamic environment are performed.In the object detection experiment,a high-speed and high-precision object detection network,mobilenet-v2-ssdlite,is trained,which can reach a detection speed of 7 frames/s,which can basically achieve real-time detection.In dynamic point detection,the optical flow method was used to eliminate dynamic point,processing speed is 16.5 frames/s.And this paper creates a data set to evaluate the performance of the algorithm.Compared with the original ORB-SLAM2 algorithm,the positioning accuracy is improved by 5 times after combining with the optical flow method.For the acquisition of three-dimensional object information,two methods based on depth filtering and point cloud segmentation are adopted.The results show that the latter’s 3D object acquisition is more accurate.Finally,the entire laboratory is constructed with a semantic map in a dynamic environment,an Octo-Map dense map is constructed,and a 3D object database is constructed based on the detection results.The detected values of the object size and position are compared with the true values,and the errors are within 5cm.The results show that the proposed algorithm has high accuracy and real-time performance.

Key words: Dynamic point detection, Object detection, Semantic mapping, Visual SLAM