Details

Autor(en) / Beteiligte

• Han, Bingxin
• Li, Xiao
• Yu, Qinghua
• Zhang, Lilian
• Lu, Huimin

Titel

A Novel Visual Odometry Aided by Vanishing Points in the Manhattan World

Ist Teil von

• 2019 2nd China Symposium on Cognitive Computing and Hybrid Intelligence (CCHI), 2019, p.252-257

Ort / Verlag

IEEE

Erscheinungsjahr

2019

Quelle

IEEE Electronic Library Online

Beschreibungen/Notizen

• Visual odometry makes continuous pose estimation using visual information to achieve robots' incremental localization. Most visual odometry algorithms realize pose estimation by single local visual feature matching or direct aligning pixel information, which will cause a large pose drift during long-term operation, especially in weakly textured scenes, where incorrect pose estimation may be caused due to the inability to extract enough points or feature matching errors. Considering that multiple visual information provides multiple geometric constraints on pose estimation, we propose a novel visual odometry aided by vanishing points for structural environments with weak texture that satisfy the Manhattan world assumption. The algorithm extracts line features from the building structure lines, and further extracts vanishing points as global constraints, suppressing the drift of pose estimation. During the optimization, we derive the projection residuals based on vanishing points and minimize the residuals to estimate the camera pose in the nonlinear optimization framework. The experimental results show that the method effectively suppresses the cumulative error of the visual odometry and improves the localization accuracy.