SimVODIS++: neural semantic visual odometry in dynamic environement

Abstract

As technology advances, more intelligent robots are being used in real world applications. Intelligent robots need an understanding of the surrounding environment to provide people with a high level of ser- vice, and understanding of the environment contains semantic and geometric information. A combination of semantic information and geometric information to intelligent robots can provide a high level of ser- vice, and methods for collecting semantic and geometric information together are presented. SimVODIS was proposed to extract both semantic and geometric information from surrounding environments by training data-driven semantic VO. However, there is a common problem with VO algorithms that do not have robust pose estimation in a dynamic environment and that overfitting occurs by learning only one camera parameter. To overcome these problems we propose SimVODIS++. SimVODIS++ achieved robust pose estimation by applying attention mechanism and prevent overfitting to single camera param- eters by data augmentation and camera parameter estimation. SimVODIS++ achieved very accurate pose estimation results and trajectory estimations compared to baselines and ablation studies.