Ego Motion Estimation Using Bird-Eye View Transformation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for estimating ego motion of a vehicle using onboard video cameras face challenges with stereo-based approaches in low or no scene overlap scenarios, monocular algorithms requiring severe motion constraints, and direct planar motion estimation issues due to local minima problems and obstacle obstruction.
Innovation Solution
A method employing region matching between consecutive top view images using feature or block matching, with 1D integral transformation to provide 1D integral images for region matching, and a RANSAC estimator to compute global motion parameters, allowing for robust ego motion estimation without requiring scene overlap or high-quality road features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If stereo-based approaches are used for ego motion estimation, then 3D information can be recovered robustly, but the method becomes unsuitable for low or no scene overlap scenarios and monocular cameras
Solution Approach 1:
The patent introduces an intermediary transformation to bird-eye view coordinates as a mediator between monocular image observations and ego motion estimation. This intermediary representation enables the system to work with monocular cameras by converting 2D image coordinates into a 3D-like bird-eye view space, bridging the gap between 2D monocular input and 3D motion recovery without requiring stereo camera pairs.
Solution Approach 2:
The patent replaces the mechanical stereo camera system with a computational approach using monocular cameras. Instead of relying on physical baseline separation between two cameras, the system substitutes this with algorithmic transformation to bird-eye view and residual flow computation, achieving similar 3D motion estimation capabilities using only a single camera.
2Measurement precision
If monocular algorithms with stereo-from-motion approach are used, then 3D coordinates can be obtained, but severe motion constraints such as loop closure are required which are not applicable in automotive applications
Solution Approach 1:
The patent extracts and removes the requirement for severe motion constraints by focusing only on the residual flow component after compensating for ego motion. Instead of requiring loop closure or global consistency constraints, the system extracts local motion residuals that can be computed independently at each pixel location, eliminating the need for complex global motion constraints.
Solution Approach 2:
The patent segments the motion estimation problem into independent pixel-wise residual flow computations. By dividing the global motion estimation task into local, independent computations at each pixel location after ego motion compensation, the system avoids the need for global motion constraints and enables parallel processing without requiring loop closure or other severe motion constraints.
3Adaptability or versatility
If direct planar motion estimation is used, then motion information can be obtained without high-quality road features, but the method suffers from local minima problems when road area is obscured by obstacles
Solution Approach 1:
The patent introduces bird-eye view transformation as an intermediary step that mediates between raw image coordinates and motion estimation. This intermediary representation transforms the problem space such that planar motion constraints can be applied more effectively, and the residual flow computation in this transformed space provides more reliable estimates even when road features are partially obscured, as the bird-eye view geometry naturally accounts for perspective distortions.
4Quantity of substance
If conventional direct planar motion estimation methods are used, then the whole road area can be utilized, but the estimation suffers from local minima problems and requires good initial approximation
Solution Approach 1:
The patent segments the motion estimation into two independent stages: first estimating ego motion using conventional direct planar motion estimation on the whole road area, then computing residual flow at each pixel location independently. This segmentation allows the first stage to utilize the entire road area for robust ego motion estimation, while the second stage processes each pixel independently without requiring good initial approximations, as the residual flow computation is less sensitive to initialization.
Solution Approach 2:
The patent applies partial action by separating the motion estimation into two stages: first obtaining a rough ego motion estimate using all road area, then refining the estimate by computing residual flows. This partial approach allows the system to use the whole road area for the initial estimate without being constrained by local minima problems in the refinement stage, as the residual flow computation operates independently at each pixel.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A method for estimating ego motion of an object moving on a surface, the method comprising generating (S1) at least two composite top view images of the surface on the basis of video frames provided by at least one onboard video camera of said object moving on said surface; performing (S2) a region matching between consecutive top view images to extract global motion parameters of said moving object; calculating (S3) the ego motion of said moving object from the extracted global motion parameters of said moving object.