Census Transform Circuitry for SIMD Processing Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Census transforms, used in stereo-vision and optical flow, are computationally intensive and typically implemented on field-programmable gate arrays (FPGAs, but there is a need for efficient implementation on single instruction, multiple data (SIMD) architectures to enhance processing efficiency.
Innovation Solution
The method involves rearranging the census transform process flow to be more compatible with SIMD architecture by generating and rearranging orientation masks, bit packing, transposing code words, and stitching them to achieve efficient computation on SIMD processors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If census transforms are implemented on FPGAs, then computational intensity is managed, but processing efficiency and compatibility with modern architectures is reduced
Solution Approach 1:
The patent divides the census transform computation into separate horizontal and vertical passes, processing pixels in two distinct directions. This segmentation allows each pass to be optimized independently for SIMD architecture, improving both processing efficiency and architectural compatibility while maintaining the computational intensity management needed for robust feature matching.
2Adaptability or versatility
If traditional census transform process flow is used, then computational accuracy is maintained, but compatibility with SIMD architecture is reduced
Solution Approach 1:
The patent inverts the traditional census transform process flow by first computing horizontal orientation masks and then vertical orientation masks, rather than processing all pixels in a single pass. This inverted approach reorganizes the computation to align with SIMD architecture capabilities while preserving the accurate relative intensity change calculations essential for robust feature matching.
Solution Approach 2:
The patent introduces an additional dimension to the computation by separately processing horizontal and vertical orientation masks before combining them. This dimensional separation allows the data to be rearranged and organized in a way that is more compatible with SIMD architecture, while the final combination step maintains the computational accuracy needed for reliable stereo-vision and optical flow applications.
3Reliability
If census transforms are performed on arrays of pixels, then feature matching robustness is improved, but computational intensity increases
Solution Approach 1:
The patent maintains continuous computation by performing census transforms on entire arrays of pixels rather than processing individual pixels sequentially. The horizontal and vertical passes are executed continuously across the image data, keeping the computational workload sustained and manageable. This continuous processing approach preserves the robustness needed for accurate feature matching while distributing the computational intensity across the entire array operation.
Data Source
AI summary
Methods of stitching data generated by a plurality of census transforms are disclosed. The methods include performing a plurality of census transforms on an array of pixels in a first direction. First and second code words of the census transform results are stored in arrays. The arrays are transposed and interleaved. The first and second code words are stitched by reading a vertical column of the array of interleaved code words.