Egomotion Estimation Using Patch-Based Projected Correlation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing egomotion estimation methods are computationally intensive and not robust enough for real-time operation on low-power platforms, such as unmanned autonomous vehicles, due to high computational costs and sensitivity to variations in illumination, weather, and occlusions.
Innovation Solution
A method that divides images into overlapping patches and uses projected correlation to estimate local 2D translation vectors, followed by trimmed-mean estimation to calculate global rotation and translation, reducing computational complexity and improving robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If feature-based egomotion estimation methods (SIFT, SURF) are used, then feature detection and matching can be performed, but computational complexity becomes too high for real-time operation on low-power platforms
Solution Approach 1:
The patent divides the image into multiple overlapping patches and processes each patch independently to estimate local translation vectors. This segmentation approach reduces the computational burden compared to processing the entire image with feature-based methods, while still capturing sufficient motion information for robust egomotion estimation.
Solution Approach 2:
The patent replaces the mechanical feature detection and matching process with a correlation-based approach in the frequency domain. By using Fourier transforms and correlation operations, the system achieves efficient computation without relying on complex feature extraction algorithms like SIFT or SURF.
2Reliability
If random sampling techniques (RANSAC) are used for feature matching, then robust motion model estimation is achieved, but the number of random samplings required makes it computationally intractable for real-time low-power operation
Solution Approach 1:
The patent performs preliminary processing by dividing the image into patches and computing their Fourier transforms before correlation operations. This preliminary action organizes the data in a way that enables efficient processing and reduces the need for extensive random sampling, allowing robust estimation without compromising real-time performance.
Solution Approach 2:
The patent changes the parameter space by working in the frequency domain instead of the spatial domain. By transforming images to the frequency domain using Fourier transforms, correlation operations become simpler and more efficient, enabling robust motion estimation without requiring numerous random samples.
3Reliability
If Fourier-Mellin transformation is used for holistic egomotion estimation, then motion model estimation is achieved, but the 2-D Fourier Transform of the whole image is computationally expensive and not robust to outliers
Solution Approach 1:
Instead of applying Fourier-Mellin transformation to the entire image, the patent segments the image into multiple patches and processes each patch independently. This segmentation dramatically reduces the computational cost of Fourier transforms while still providing robust motion estimation through the aggregation of local translation vectors.
Solution Approach 2:
The patent applies the computationally intensive Fourier transform operation to only partial regions (patches) of the image rather than the whole image. This partial action approach maintains the benefits of frequency-domain processing for motion estimation while reducing overall computational requirements to levels suitable for low-power platforms.
Data Source
Figure 1A~2D
Figure 3
Figure 4A~4B
AI summary
A method for improving computational efficiency for egomotion estimation includes detecting a pair of consecutive images, and estimating a rotation angle and a two dimensional translation vector that best models the motion between the image frames. The pair of images are each divided into overlapping image patches, and then an efficient projected correlation approach is used to estimate a local two dimensional translation vector for each image patch. The translation vectors from all the patches form a vector field, and then trimmed-mean method is used to robustly and efficiently estimate the rotation angle and the two dimensional translation that best models the egomotion of the detector, which may be a camera.