Hybrid Camera Pose Estimation With Online Partitioning for SLAM

Xinyi Li, Haibin Ling

Abstract

This letter presents a hybrid real-time camera pose estimation framework with a novel partitioning scheme and introduces motion averaging to monocular Simultaneous Localization and Mapping (SLAM) systems. Breaking through the limitations of fixed-size temporal partitioning in many conventional SLAM pipelines, our approach significantly improves the accuracy of local bundle adjustment by gathering spatially-strongly-connected cameras into each block. With the dynamic initialization using intermediate computation values, we improve the Levenberg-Marquardt solver to further enhance the efficiency of the local optimization. Moreover, the dense data association between blocks by our covisibility-based partitioning enables us to explore and implement motion averaging to efficiently align the blocks globally, updating camera motion estimations on-the-fly. Experiments on benchmarks convincingly demonstrate the practicality and robustness of our proposed approach by significantly outperforming conventional approaches.

Introduction

At the heart of most visual SLAM and Structure-from-Motion (SfM) systems is the challenge of accurately tracking camera poses and 3D landmarks over time. While Bundle Adjustment (BA) is the gold standard for accuracy, it faces significant hurdles in real-time applications. Incremental BA is prone to drift, and Global BA suffers from cubic-order computational complexity, quickly becoming a bottleneck as maps grow. Furthermore, traditional systems often use fixed-size temporal partitioning, which groups consecutive frames regardless of whether they share visual features, leading to suboptimal convergence.

Proposed Hybrid Framework

The proposed framework addresses these issues by marrying local Bundle Adjustment with global motion averaging through a novel dynamic partitioning scheme.

Overview of the proposed hybrid SLAM pipeline: (a) Input correspondences, (b) Online co-visibility partitioning, (c) Local refinement, and (d) Global alignment.

Online Co-visibility Partitioning

Instead of grouping a fixed number of sequential frames, the system dynamically builds "blocks" based on spatial co-visibility. A block groups cameras that share a high number of 3D points. Crucially, older cameras from previous blocks are pulled into the current block if they share overlapping views, creating dense, overlapping connections between blocks. This ensures that the local optimization has sufficient geometric constraints to find an accurate solution.

Self-Adaptive Local Optimization

For local BA within a block, the framework introduces a modified Levenberg-Marquardt (L-M) solver. Because overlapping cameras carry intermediate computation values from previous blocks, the solver benefits from a highly accurate initialization. The solver dynamically adapts its step size based on the block size, achieving a faster, quadratic convergence rate.

Global Alignment via Motion Averaging

The standout feature of this system is the use of single rotation averaging for global alignment. Because the dynamic partitioning ensures that adjacent blocks share multiple camera poses, the system can globally align the blocks by averaging the relative rotations of these shared cameras.

The transition from local refinement (c) to global alignment (d) using motion averaging, which bypasses the need for expensive point-camera global BA.

This approach bypasses the need for explicit, expensive point-cloud optimization at the global scale, achieving a 60–150x speedup in global optimization compared to conventional methods.

Experimental Results

The framework was evaluated on the TUM RGB-D (indoor) and KITTI Odometry (large-scale outdoor) benchmarks:

  • Massive Speedup: The global alignment step maintains an almost constant run-time even as the number of image frames scales up, solving the computational "wall" of traditional SLAM.
  • Superior Accuracy: On the TUM fr3_office sequence, it achieved an RMSE of 0.8 cm, significantly outperforming ORB-SLAM's 3.5 cm.
  • Robustness at Scale: On KITTI Sequence 08, the system successfully mitigated accumulated drift without explicit loop closures, achieving a trajectory RMSE of 19.7 meters compared to ORB-SLAM's 46.7 meters.

Conclusion

By replacing rigid temporal windows with co-visibility-based partitioning and leveraging motion averaging for global consistency, this hybrid framework provides a scalable and highly efficient solution for real-time SLAM. While the current reliance on SIFT features necessitates GPU acceleration for real-time performance, the underlying architecture demonstrates a significant leap in balancing accuracy and computational efficiency.

Summary updated

Abstract

This letter presents a hybrid real-time camera pose estimation framework with a novel partitioning scheme and introduces motion averaging to monocular Simultaneous Localization and Mapping (SLAM) systems. Breaking through the limitations of fixed-size temporal partitioning in many conventional SLAM pipelines, our approach significantly improves the accuracy of local bundle adjustment by gathering spatially-strongly-connected cameras into each block. With the dynamic initialization using intermediate computation values, we improve the Levenberg-Marquardt solver to further enhance the efficiency of the local optimization. Moreover, the dense data association between blocks by our covisibility-based partitioning enables us to explore and implement motion averaging to efficiently align the blocks globally, updating camera motion estimations on-the-fly. Experiments on benchmarks convincingly demonstrate the practicality and robustness of our proposed approach by significantly outperforming conventional approaches.
BibTeX
@article{li2020hybrid,
  author = {Xinyi Li and Haibin Ling},
  title = {{Hybrid Camera Pose Estimation With Online Partitioning for SLAM}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2020},
  url = {https://openreview.net/forum?id=mP7N-0sHhjH},
}
Figure