A visual dynamic-SLAM method based semantic segmentation and multi-view geometry

Abstract

Robust localization in dynamic scenes is one of the bottlenecks in the development of visual SLAM. In recent years, many researchers have built new slam systems to avoid the interference of dynamic objects. However, it is not perfect to provide the original object mask with the existing semantic segmentation methods, which can’t completely cover the moving objects, especially the object boundary. In order to solve this problem, we proposed a dynamic SLAM method combining semantic information and multi view geometry in this paper. In our algorithm, semantic network is used to recognize and segment each object in the image frame. And then, the moving object processing link in the dynamic scene is realized based on semantic tags and multi-view geometry, so as to reduce the influence of moving objects on the pose estimation of SLAM algorithm and improve the scene adaptability of the algorithm. We evaluate the performance of our system with TUM dataset. Results are compared with the typical visual SLAM system to show that the absolute trajectory accuracy in the algorithm proposed has been greatly improved.