Semi-Global Matching Cost Compression for Embedded Stereo
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Semi-Global Matching (SGM) techniques face impracticality in embedded platforms for high-resolution images due to high on-chip memory usage and DDR memory bandwidth issues, making it challenging to implement effectively.
Innovation Solution
The proposed solution involves calculating and sub-sampling SGM path costs based on disparity levels, using disparity-dependent sub-sampling ratios and quantization, and employing bi-linear interpolation and inverse quantization to reduce memory requirements, allowing for efficient storage and processing of SGM path costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If SGM path costs are stored for all pixels at all disparity levels, then stereo correspondence accuracy is improved, but on-chip memory usage increases making implementation impractical
Solution Approach 1:
The patent extracts and stores only a subset of SGM path costs rather than all costs. Specifically, it stores path costs at selected disparity levels and selected pixel locations (e.g., every nth pixel), removing redundant information while preserving essential correspondence data needed for accurate depth estimation.
Solution Approach 2:
The patent segments the disparity space by dividing the full range of disparity levels into multiple ranges, and separately segments the pixel space by selecting subsets of pixels at different disparity levels. This segmentation allows different storage strategies to be applied to different segments, reducing overall memory requirements while maintaining accuracy where needed.
2Reliability
If full SGM path costs are maintained for high-resolution images, then correspondence result robustness is improved, but DDR memory bandwidth requirements increase
Solution Approach 1:
The patent extracts only the essential path cost information needed for robust correspondence results, storing a reduced set of costs at strategically selected disparity levels and pixel locations. This extraction reduces the volume of data that must be transferred via DDR memory, improving bandwidth efficiency while maintaining the reliability needed for accurate stereo matching.
3Quantity of substance
If SGM path costs are sub-sampled, then memory usage is reduced, but information loss occurs
Solution Approach 1:
The patent applies local quality by using different sub-sampling ratios for different disparity levels and different regions of the image. Areas with higher importance for stereo correspondence (e.g., regions with higher gradient magnitude or specific disparity ranges) are preserved with lower sub-sampling ratios, while less critical areas use higher sub-sampling ratios, minimizing information loss where it matters most.
Data Source
AI summary
This disclosure describes techniques for performing semi-global matching (SGM) path cost compression. In some examples, the techniques may perform disparity-dependent sub-sampling of a set of SGM path costs where the sub-sampling ratio is determined based on a candidate disparity level. The sub-sampled SGM path costs may be stored in a memory. When retrieved from memory, the sub-sampled SGM path costs may be interpolated to reconstruct the other path costs not stored in the memory. The reconstructed path costs may be used for further SGM processing. In further examples, the techniques may perform disparity-dependent quantization on the SGM path costs or the sub-sampled SGM path costs, and store the quantized SGM path costs in memory. The techniques of this disclosure may reduce bandwidth as well as reduce the memory footprint needed to implement an SGM algorithm.