Stereo Camera Disparity Matching Without Pre-Rectification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing stereo camera systems require computationally intensive image rectification, which blurs pixel information and increases memory usage, and often discard or increase resolution without adding valuable data, complicating disparity computation.
Innovation Solution
A method using a trained convolutional neural network for 2D-matching of unrectified stereo camera images to generate pixel-level disparity information, eliminating the need for pre-rectification and reducing algorithmic complexity while preserving image information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If image rectification is performed before stereo matching, then the matching algorithm complexity is reduced, but pixel information is blurred and memory usage increases
Solution Approach 1:
The patent performs rectification after matching rather than before, which is the opposite of conventional approaches. The CNN performs stereo matching on unrectified images first, then rectification is applied to the disparity map. This eliminates the need for preliminary rectification of high-resolution images, preserving pixel information while still achieving the complexity reduction benefit through the CNN's efficient processing.
Solution Approach 2:
The patent replaces traditional geometric rectification mechanisms with a learned CNN-based matching approach. Instead of using fixed geometric transformations that blur pixels, the CNN learns to perform matching directly on unrectified images, substituting mechanical rectification with an intelligent learning-based system that preserves information.
2Measurement precision
If rectification is performed on high resolution images, then disparity computation accuracy is improved, but computational intensity and memory requirements increase
Solution Approach 1:
The patent extracts only the essential disparity information using a CNN that processes images efficiently. Instead of performing computationally intensive rectification on entire high-resolution images, the CNN extracts disparity maps directly, taking out only the necessary information while avoiding unnecessary computational overhead and memory usage.
Solution Approach 2:
The patent changes the processing parameters by working with unrectified images in the native coordinate system rather than transforming to a rectified coordinate system. This parameter change eliminates the need for high-resolution rectification while maintaining disparity accuracy through the CNN's learned matching capabilities.
3Loss of information
If traditional stereo matching is performed on unrectified images, then pixel information is preserved, but matching complexity increases significantly
Solution Approach 1:
The patent substitutes complex traditional stereo matching algorithms with a CNN-based approach. The CNN automatically learns the matching patterns from data, replacing complex mechanical matching procedures with an intelligent system that achieves both information preservation and computational efficiency through learned feature representations.
Solution Approach 2:
The patent changes the matching approach by using a learned model that operates directly on unrectified images. Instead of complex geometric transformations and multiple processing stages, the CNN uses learned parameters and features to perform matching efficiently while preserving all pixel information from the original images.
Data Source
AI summary
The disclosure relates to a method and system for extracting disparity information out of a pair of stereo camera images for processing images of camera sensors for Automated Driving or Advanced Driver Assistance Systems in a vehicle. The method includes: a) receiving a pair of unrectified stereo camera images, b) performing a 2D-matching of the received pair of images using a trained convolutional neural network and providing the 2D-displacement on a pixel-by-pixel level as disparity information, and c) outputting the 2D-displacement on pixel level. The method is robust which does not require a high precision rectification. Applying rectification on the disparity values after the matching process leads to reduced algorithmic complexity with simultaneous preservation of all information.