Adaptive RANSAC Threshold for Video Stabilization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video stabilization algorithms, such as RANSAC, face challenges in efficiently estimating movement between successive images due to non-uniform noise and computational variability, particularly in embedded systems like mobile devices, leading to potential inaccuracies in video stabilization and motion estimation.
Innovation Solution
Adapting the threshold calculation in RANSAC-type algorithms by incorporating depth estimation and spatial movement correction terms for each image point, allowing for a variable threshold that accounts for the depth of each point and the apparatus's movement, thereby improving resilience to non-uniform noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a fixed threshold is used in RANSAC algorithm for all image points, then the algorithm is simple to implement, but it cannot account for non-uniform noise and depth variations, leading to reduced measurement precision
Solution Approach 1:
The patent applies local quality by making the threshold adaptive to each image point's depth and position. Instead of using a single fixed threshold for all points, the system calculates a specific threshold for each point based on its depth estimation and spatial location, thereby accounting for non-uniform noise characteristics across different regions of the image.
Solution Approach 2:
The patent changes the threshold parameter from a fixed value to a dynamic value that varies with depth and position. The threshold is calculated as a function of depth estimation and spatial coordinates, allowing it to adapt to local noise characteristics and improve measurement precision for points at different depths.
2Productivity
If RANSAC algorithm processes all image points with uniform threshold, then computational time is predictable, but calculation time becomes very variable depending on image difficulty and number of points tested
Solution Approach 1:
The patent applies preliminary action by performing depth estimation and threshold calculation before the main RANSAC processing. By pre-computing the adaptive thresholds based on depth maps and spatial positions, the system prepares the necessary parameters in advance, which helps make the subsequent processing more predictable and efficient.
3Reliability
If a single threshold is used for all points in the image, then the algorithm is computationally efficient, but it reduces reliability in scenarios with varying noise levels and complex movements
Solution Approach 1:
The patent makes the threshold local to each image point by incorporating depth estimation and spatial position into the threshold calculation. This allows the system to adapt to local noise characteristics and movement patterns, improving reliability in scenarios with varying noise levels and complex movements.
Solution Approach 2:
The patent introduces dynamics by making the threshold adaptive rather than static. The threshold changes dynamically based on the depth and position of each point, allowing the system to respond to varying noise levels and movement complexities across different regions of the image.
Data Source
AI summary
A method determines a movement of an apparatus between capturing first and second images. The method includes testing model hypotheses of the movement by for example a RANSAC algorithm, operating on a set of first points in the first image and assumed corresponding second points in the second image to deliver the best model hypothesis. The testing includes, for each first point, calculating a corresponding estimated point using the tested model hypothesis, determining the back-projection error between the estimated point and the second point in the second image, and comparing each back projection error with a threshold. The testing comprises for each first point, determining a correction term based on an estimation of the depth of the first point in the first image and an estimation of the movement between the first and second images, and determining the threshold associated with the first point by using said correction term.