A Semantic Visual Slam for Dynamic Environments

EasyChair Preprint 9027

Abstract

For autonomous robotic systems, simultaneous localization and mapping algorithm (SLAM) is one of the core technologies. Traditional simultaneous localization and mapping algorithms usually assume that the system works in a static environment. However, when there are dynamic objects in the environment, For example, people walking around indoors, etc., they will bring false observation data to the system, and the visual SLAM system will become unstable, affecting the system localization and limiting the application of visual SLAM in reality. However, if the dynamic objects occupy most of the image area, it will still affect the pose tracking. To solve this problem, this paper proposes a vision slam system adapted to dynamic environment, based on ORB-SLAM3, adding a lightweight target detection tracking network of YOLO5 at the front end to detect dynamic targets in the tracking environment, and rejecting the feature points of dynamic targets to reduce the influence of dynamic objects on the system. To avoid the lack of feature information due to excessive dynamic feature point rejection, the improved SLAM system in this paper uses Superpoint to replace ORB for feature extraction and descriptor calculation to further improve the robustness of the levy system. The experimental results of the dataset show that the improved ORB-SLAM3 system can effectively improve the robustness and accuracy of the system under dynamic environment.

Keyphrases: Superpoint, Visiual SLAM, dynamic environment, object detection

@booklet{EasyChair:9027,
  author    = {Xuanhui Zhong and Hui Chai and Qian Pu and Chengji Zheng and Meng Hong and Jiaming Tian and Yaodong Wang and Min Guo},
  title     = {A Semantic Visual Slam for Dynamic Environments},
  howpublished = {EasyChair Preprint 9027},
  year      = {EasyChair, 2022}}