Robust and Accurate Monocular Pose Tracking for Large Pose Shift.
Qiufu Wang,Jiexin Zhou,Zhang Li,Xiaoliang Sun,Qifeng Yu
DOI: https://doi.org/10.1109/tie.2022.3217598
IF: 7.7
2023-01-01
IEEE Transactions on Industrial Electronics
Abstract:Tracking the pose of a specific rigid object from monocular sequences is a basic problem in computer vision. State-of-the-art methods assume motion continuity between two consecutive frames. However, drastic relative motion causes large interframe pose shifts, especially in applications such as robotic grasping, failed satellite maintenance, and space debris removal. Large pose shifts interrupt the interframe motion continuity leading to tracking failure. In this article, we propose a robust and accurate monocular pose tracking method for tracking objects with large pose shifts. Using an indexable sparse viewpoint model to represent the object 3D geometry, we propose establishing a transitional view, which is searched for in an efficient variable-step way, to recover motion continuity. Then, a region-based optimization algorithm is adopted to optimize the pose based on the transitional view. Finally, we use a single-rendering-based pose refinement process to achieve highly accurate pose results. The experiments on the region-based object tracking (RBOT) dataset, the modified RBOT dataset, the synthetic large pose shift sequences, and real sequences demonstrated that the proposed method achieved superior performance to the state-of-the-art methods in tracking objects with large pose shifts.