GMVO: graph-CNN-based trajectories’ instance-level segmentation for multi-motion visual odometry

Abstract

Multi-motion visual odometry (MVO) based on geometric features is widely used. However, the accuracy and efficiency of the MVO algorithm are limited to instance-level segmentation of landmarks with inconsistent motion. To address this challenge, this study proposes a two-stage landmark instance segmentation network, GMVO, based on geometric features. First, a supervised method is used to classify static trajectories for camera odometer calculations and dynamic trajectories for rigid body motion estimation. Next, the dynamic trajectories are clustered and the instance-level segmentation of their motion consistency is restricted within them. Finally, bundle adjustment is applied to optimize the observation of landmarks within a sliding window and complete the MVO calculation. Extensive experiments were conducted on the OMD and KITTI data sets to evaluate the performance of the GMVO algorithm in terms of its classification performance, camera odometer accuracy, and real-time performance. The related results were visualized, which verified that the GMVO algorithm is more effective than and superior to state-of-the-art approaches.

Keywords

• SLAM
• multi-motion visual odometry
• graph CNN
• trajectory instance segmentation