Visual Semantic SLAM Based on Examination of Moving Consistency in Dynamic Scenes

Abstract

Simultaneous Localization and Mapping (SLAM) has always been a hot topic in the field of intelligent mobile robots in recent years. The existing SLAM algorithms have been developed relatively maturely under the assumption of static environment, and have achieved good performance. However, assuming a static environment limits the application of SLAM in dynamic scenarios to a certain extent. To solve the problems existing in the application of robots in dynamic scenes, this paper proposes a semantic SLAM system for dynamic scenes. We add depth information to adjust the contour of semantic segmentation results and combine it with an improved optical flow-based examination of moving consistency to form a dynamic detection algorithm that classifies feature points into three categories. Different types of features are selected according to different robot tasks, which weakens the effect of dynamic objects on the camera pose estimation and makes full use of static features to improve the localization accuracy of the robot in dynamic scenes. Meanwhile, a semantic octree map constructed from the static scene is generated, which can be used for advanced tasks. We set up experiments on the TUM RGB-D dataset. Compared with ORB-SLAM2, our system achieves up to 97.43% improvement in dynamic scenes and also outperforms four other advanced dynamic SLAM systems. Experiments show that our system reduces the pose estimation error of the camera in dynamic scenes to different degrees and improves the localization accuracy of the camera.