Monocular visual odometry in dynamic environments by utilizing motion segmentation and attention mechanism

Abstract Learning-based visual odometry (VO), that estimates frame-to-frame (F2F) translation and rotation in end-to-end network, has attracted much attention in recent years. However, existing methods perform poorly in complex environments, especially when there are moving objects in the scene. To solve the above issues, a novel monocular VO is proposed based on motion segmentation and attention mechanism under dynamic environments. Firstly, semantic information and optical flow residuals are adopted for motion segmentation through iterative optimization, so as to eliminate the influence of dynamic objects. Secondly, a parallel attention mechanism module is used for extracting the most relevant features for pose estimation. Two types of experiment (odometry dataset and tracking dataset) based on Karlsruhe Institute of technology and Toyota Technological Institute (KITTI) are carried out to show the effectiveness of the proposed method. Experimental results show that the accuracy of pose estimation is greatly improved by adding motion segmentation and attention mechanism. Moreover, the comparison results show that the proposed method performs comparably to geometry-based techniques and outperforms other learning-based approaches. This paper lays the foundation for accurate pose estimation in dynamic environments.