A Fast and Light-weight Non-Iterative Visual Odometry with RGB-D Cameras

2025-07-24

AI Summary

Traditional RGB-D visual odometry (VO) methods rely on iterative optimization and feature extraction/matching, resulting in high computational cost, poor real-time performance, and insufficient robustness in low-texture environments. To address these limitations, this paper proposes a non-iterative RGB-D VO framework that decouples rotation and translation estimation for the first time: rotation is directly solved via plane overlap constraints inherent in scene geometry, while translation is efficiently estimated using a kernelized cross-correlator (KCC). The approach entirely eliminates feature detection, description, matching, and iterative optimization. Evaluated on a resource-constrained Intel Core i5 CPU, it achieves 71 Hz—significantly outperforming state-of-the-art methods—while maintaining high accuracy and strong robustness even in texture-deprived, degenerate scenarios. This work establishes a new paradigm for efficient, reliable six-degree-of-freedom pose estimation on computationally limited platforms.

Technology Category

• Computer Vision: Motion & Tracking
• Intelligent Robots: State Estimation
• Search and Optimization: Non-convex Optimization

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Experiences and lessons learnt from Web-based algorithms and system deployments
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

In this paper, we introduce a novel approach for efficiently estimating the 6-Degree-of-Freedom (DoF) robot pose with a decoupled, non-iterative method that capitalizes on overlapping planar elements. Conventional RGB-D visual odometry(RGBD-VO) often relies on iterative optimization solvers to estimate pose and involves a process of feature extraction and matching. This results in significant computational burden and time delays. To address this, our innovative method for RGBD-VO separates the estimation of rotation and translation. Initially, we exploit the overlaid planar characteristics within the scene to calculate the rotation matrix. Following this, we utilize a kernel cross-correlator (KCC) to ascertain the translation. By sidestepping the resource-intensive iterative optimization and feature extraction and alignment procedures, our methodology offers improved computational efficacy, achieving a performance of 71Hz on a lower-end i5 CPU. When the RGBD-VO does not rely on feature points, our technique exhibits enhanced performance in low-texture degenerative environments compared to state-of-the-art methods.

Problem

• Estimates 6-DoF robot pose efficiently without iterative methods
• Reduces computational burden by decoupling rotation and translation estimation
• Improves performance in low-texture environments without feature points

Innovation

• Decouples rotation and translation estimation
• Uses planar elements for rotation calculation
• Employs kernel cross-correlator for translation