.png){kind=icon}
.png){kind=icon}
Return
R-SDSO: Robust stereo direct sparse odometry
.png){kind=icon}
PRE
AI
DOI：10.1007/s00371-021-02278-0.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This paper presents a robust stereo direct visual odometry method with improved robustness against drastic brightness variation and aggressive rotation. It is achieved by incorporating a direct sparse odometry based on image preprocessing, a depth initialization module, and an abend recovery module into the visual odometry framework. The image preprocessing enhances raw camera images, which facilitates more accurate pixels detecting. Meanwhile, a new error function based on image preprocessing is proposed for making direct visual odometry robust to brightness variation in the environment. In the depth initialization module, the Delaunay triangulation algorithm and feature point matching are combined together for efficient and robust depth estimation. Furthermore, in the abend recovery module, we design a lost/abnormal detection method and a robust state restoration strategy to address the tracking lost/abnormal problem in harsh conditions. Evaluation results on KITTI and EuRoC datasets and a light-switch experiment demonstrate that, with the aid of these three modules, the proposed method can achieve state-of-the-art performance, even compared with visual-inertial fusion methods.
Keywords:
Visual odometry
Image preprocessing
Localization and mapping
Robust