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.
Introduction
Visual Odometry (VO) serves as the backbone of autonomous robot navigation, providing real-time pose estimation through camera data. While mainstream "feature-based" methods, such as ORB-SLAM3 or DSO, have become industry standards, they often fail in environments that are feature-deprived, blurry, or contain highly repetitive textures—common scenarios in aerial surveys over water or uniform agricultural fields.
The Fourier-Mellin Transform (FMT) offers a spectral-based alternative that excels in these challenging conditions by operating in the frequency domain rather than relying on discrete point descriptors. However, classic FMT is limited by a planar constraint (assuming a flat world parallel to the camera) and significant drift. While the Extended Fourier Mellin Transformation (eFMT) successfully addressed the planar constraint to allow for multi-depth scenes, it remained computationally expensive and lacked a robust mechanism to handle noise and scale consistency. This paper introduces o-eFMT, an optimized framework designed to achieve real-time performance and superior trajectory accuracy through algorithmic streamlining and local back-end optimization.
Methodology
The o-eFMT approach improves the registration pipeline through three primary technical contributions:
Unified Energy Vector Extraction
In the original eFMT, calculating multiple zoom levels was computationally prohibitive. The optimized approach utilizes a Gaussian-modeled extraction method to generate a single "zoom energy vector" and "translation energy vector." By modeling the energy distribution as a Gaussian block rather than sampling every possible row in the Phase Shift Diagram, the number of required Fourier transforms is drastically reduced. This unifies the processing of rotation, scale, and translation into a single, efficient framework.
Robust Pattern Matching with Laplace Transforms
Identifying the correct scaling factor between frames requires finding the minimum error in a sequence. However, environmental noise often creates "false" minima that lead to registration failure. The authors observed that the "deepest" error is not always the correct one, but the "sharpest" change in the error sequence usually is. By applying a Laplace transform to the error sequence, the algorithm identifies the point with the highest first-order derivative. This "sharpness" metric proves to be a significantly more reliable indicator of the true transformation than simple magnitude.
3-Frame Back-End Optimization
To mitigate the drift inherent in pure VO methods, o-eFMT introduces a local graph-based optimizer. The system considers three consecutive frames ($I_0, I_1, I_2$) and calculates the transformations between the pairs (0,1), (1,2), and the "loop-closing" (0,2) pair. This creates a constraint system that minimizes accumulated drift and allows for the estimation of uncertainty for each pose, providing a more stable trajectory over time.
Experimental Results
The algorithm was evaluated using both real-world and simulated datasets to test its robustness across different environments.
Real-World Performance
Testing on the ShanghaiTech Campus dataset—collected via a DJI Matrice-300 UAV with a downward-looking camera—presented challenges including multi-depth planes like rooftops and bridges. In these tests, o-eFMT achieved a mean error of 15.7m, significantly outperforming the original eFMT (27.5m) and the standard FMT (79.8m). Notably, traditional methods struggled; ORB-SLAM3 reported a mean error of 165.3m, while DSO failed to maintain tracking entirely.
Simulation and Speed
In controlled Blender simulations featuring a "Circle" camera track over varied textures, o-eFMT achieved a median error of 0.58m, a massive improvement over the original eFMT’s 44.3m. Furthermore, the optimizations resulted in a 3x speedup in processing time, with o-eFMT requiring only 0.1432s per registration compared to the 0.4069s required by its predecessor.
Conclusion
The o-eFMT algorithm represents a significant step forward for spectral-based visual odometry. By prioritizing the "sharpness" of error signals and implementing a 3-frame optimization loop, the system provides a robust solution for environments where feature-based SLAM fails. While currently limited to 4-DOF (translation in X, Y, Z and rotation around Z), future work will focus on expanding the system to a full SLAM framework with global loop-closure detection and integration into the SLAM Hive benchmarking suite.
Summary updated