Optimizing the Extended Fourier Mellin Transformation Algorithm

Wenqing Jiang, Chengqian Li, Jinyue Cao, Sören Schwertfeger

TL;DR

The paper introduces o-eFMT, an optimized version of the Extended Fourier Mellin Transformation for visual odometry. By streamlining energy vector extraction and adding a 3-frame back-end local optimizer, the authors achieved a 3x speedup and significantly higher trajectory accuracy in challenging, multi-depth, and feature-deprived environments compared to state-of-the-art methods like ORB-SLAM3 and DSO.

What Problem Does This Paper Address?

Visual Odometry (VO) is the backbone of robot navigation, but mainstream "feature-based" methods (like ORB-SLAM3) often fail in environments that are feature-deprived, blurry, or contain highly repetitive textures (e.g., over water or uniform fields).

The Fourier-Mellin Transform (FMT) is a spectral-based alternative that excels in these tricky conditions because it operates in the frequency domain rather than relying on discrete points. However, classic FMT has two major flaws:

  1. Planar Constraint: It assumes the world is perfectly flat and parallel to the camera.
  2. Drift: As a pure VO approach, it accumulates errors over time without the benefit of optimization.

A previous iteration, eFMT, solved the planar constraint but was computationally expensive and lacked a robust way to handle noise and scale consistency. This paper introduces "o-eFMT" to make the algorithm fast enough for real-time robotics while improving its robustness through local optimization.

Methods

The authors improve the eFMT pipeline through several key technical contributions:

  • Unified Energy Vector Extraction: Instead of calculating multiple zooms (which is computationally heavy), they use a Gaussian-modeled approach to extract a single "zoom energy vector" and "translation energy vector." This unifies the processing of rotation, scale, and translation into a single efficient framework.
  • Robust Pattern Matching with Laplace Transforms: To find the correct scaling factor between frames, the algorithm looks for the minimum error. However, noise often creates "false" minima. The authors apply a Laplace transform to the error sequence; they look for the point that is not only a minimum but also has the highest first-order derivative (the "sharpest" drop), which proves to be a much more reliable indicator of the true transformation.
  • 3-Frame Back-End Optimization: This is the "o" in o-eFMT. The authors implement a local graph-based optimizer that considers three consecutive frames ($I_0, I_1, I_2$). By calculating the transformations between (0,1), (1,2), and the "loop-closing" (0,2) pair, they create a constraint system that minimizes accumulated drift and estimates uncertainty for each pose.

Datasets

The algorithm was evaluated on two distinct types of data:

  1. Real-World Dataset: The ShanghaiTech Campus dataset, collected via a DJI Matrice-300 UAV with a downward-looking camera. This dataset is challenging due to multi-depth planes (rooftops, bridges, ground).
  2. Simulated Datasets: Created using Blender, featuring "Circle" and "Analemma" (figure-eight) camera tracks over a scene with three distinct depth planes and varied textures.

Key Insights and Ideas

  • Sharpness over Magnitude: In pattern matching, the "deepest" error isn't always the right one, but the "fastest changing" error usually is. Using the Laplace transform to find the sharpest change in the error sequence is a clever way to bypass noise.
  • Bypassing Multi-Zoom: By modeling the energy distribution as a Gaussian block rather than sampling every possible row in the Phase Shift Diagram, the authors drastically reduced the number of Fourier transforms required.
  • Spectral VO as a "Middle Ground": The paper reinforces that while feature-based SLAM is great for general rooms, spectral methods like o-eFMT are the superior choice for "unstructured" environments where traditional descriptors (like ORB or SIFT) fail to find unique anchors.

Key Results

  • Speed: o-eFMT is significantly faster, taking 0.1432s per registration compared to eFMT’s 0.4069s (nearly a 3x speedup).
  • Accuracy (Real World): On the ShanghaiTech dataset, o-eFMT achieved a mean error of 15.7m, outperforming eFMT (27.5m) and FMT (79.8m). Notably, ORB-SLAM3 struggled significantly with a mean error of 165.3m, and DSO failed to track entirely.
  • Accuracy (Simulation): In the "Circle" simulation, o-eFMT achieved a median error of 0.58m, while the original eFMT sat at 44.3m. This highlights the massive impact of the new back-end optimization and Laplace-based matching.

Limitations and Future Work

  • Degrees of Freedom: The algorithm is currently limited to 4-DOF (translation in X, Y, Z and rotation around Z). It cannot handle significant roll or pitch, which limits it to downward-looking or stabilized camera systems.
  • Loop Closure: While the 3-frame optimization helps locally, it isn't a full SLAM system. The authors plan to add global loop-closure detection to turn o-eFMT into a complete SLAM framework.
  • Integration: The authors intend to integrate the algorithm into the "SLAM Hive" benchmarking suite for broader community evaluation.

Summary updated

Abstract

With the increasing application of robots, stable and efficient Visual Odometry (VO) algorithms are becoming more and more important. Based on the Fourier Mellin Transformation (FMT) algorithm, the extended Fourier Mellin Transformation (eFMT) is an image registration approach that can be applied to downward-looking cameras, for example on aerial and underwater vehicles. eFMT extends FMT to multi-depth scenes and thus more application scenarios. It is a visual odometry method which estimates the pose transformation between three overlapping images. On this basis, we develop an optimized eFMT algorithm that improves certain aspects of the method and combines it with back-end optimization for the small loop of three consecutive frames. For this we investigate the extraction of uncertainty information from the eFMT registration, the related objective function and the graph-based optimization. Finally, we design a series of experiments to investigate the properties of this approach and compare it with other VO and SLAM (Simultaneous Localization and Mapping) algorithms. The results show the superior accuracy and speed of our o-eFMT approach, which is published as open source.
BibTeX
@inproceedings{jiang2023optimizing,
  author = {Wenqing Jiang and Chengqian Li and Jinyue Cao and Sören Schwertfeger},
  title = {{Optimizing the Extended Fourier Mellin Transformation Algorithm}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  url = {https://openreview.net/forum?id=9YrkjgIPa4},
}

Loading related content...

Figure