Dynamic Visual SLAM Based on Efficient Two-Stage Statistical Model

Shaoteng Wu,Ruitao Zhang,Yafei Wang

DOI: https://doi.org/10.1109/cvci56766.2022.9964605

2022-01-01

Abstract:Dynamic objects usually break the static environment assumption which is a crucial principle of visual Simultaneous Localization and Mapping(SLAM), leading to bad tracking quality and high computation burden to improve performances. To promote visual SLAM performances in dynamic environment with less extra time consumption, we propose an efficient robust dynamic landmark removal method based on a two-stage statistical model using static score propagation and conditional random field (CRF) based on ORB-SLAM 2 framework. To detect obviously dynamic landmarks, the static score is firstly estimated according to statistical feature of re-projection error and propagated along multiple observing frames. Then the CRF is constructed to refine the initial static/dynamic labelling results in consideration of neighboring motion and long-term observation quality. With the accurate static/dynamic labels, only static landmarks are employed in tracking process to eliminate disturbance from dynamic objects. Evaluations on TUM RGB-D dynamic datasets indicate that our method improves visual SLAM performances in highly dynamic environment with accuracy at the same level as state-of-the-art methods and a significant advantage on time efficiency.