Efficient contour-based monocular 3D object tracking with two-stage pose initialization for large shifts
Kai Liu, Jing Chen, Jixiang Chen, Yilin Li, Yongtian Wang · Journal of Electronic Imaging · 2025
Contour-based 3D object tracking offers high tracking accuracy and speed but faces challenges with large inter-frame pose shifts. In this paper, we propose a fast and efficient contour-based tracking method to enhance robustness and accuracy under conditions of significant shifts. Specifically, considering large displacements and rotations in the 6D pose space, we propose a two-stage large pose initialization approach to avoid trapping in erroneous local minima. This approach first employs an adaptive shift on the grid foreground probability map and then implements nonlocal rotation search based on Bayesian optimization, enabling more accurate and reliable initial pose generation. For accurate and stable pose estimation under cluttered backgrounds, ambiguous colors, and noise, we develop a robust weight function that integrates distance-based Gaussian and color probability weights. By adaptively adjusting the Gaussian term, we adopt a coarse-to-fine optimization strategy to handle large displacements and speed up convergence. Experimental results on the modified RBOT dataset, a real-world dataset, and a real sequence demonstrate that our method achieves state-of-the-art tracking accuracy under large pose shifts while maintaining real-time performance on only central processing unit.