Census Transform Hardware Implementation Reducing Comparator Count
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Census Transform algorithms are computationally intensive and require numerous redundant comparisons in image processing, leading to high hardware and power consumption, especially in applications like Dense Optical Flow and object tracking.
Innovation Solution
A new Census Transform algorithm is introduced that reduces hardware and power consumption by inverting previous comparison results and using a fractional position index to ensure no two pixels are equal, thereby eliminating the need for redundant comparisons.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional Census Transform algorithm is used with full pixel comparisons, then measurement precision is maintained, but device complexity and power consumption increase significantly
Solution Approach 1:
The patent inverts the traditional Census Transform approach by comparing pixels from the bottom-right corner moving leftward and upward, rather than the conventional top-left to bottom-right scanning order. This inversion allows the algorithm to exploit symmetry in the comparison process, where the conjugate comparison results can be directly obtained by inverting the forward comparison results, thereby reducing the number of required comparators while maintaining signature accuracy
Solution Approach 2:
The patent applies partial action by performing comparisons only in one direction (forward comparisons) and deriving the conjugate comparison results through inversion rather than performing all comparisons explicitly. This selective approach reduces the hardware complexity by approximately half while preserving the measurement precision needed for accurate feature matching and optical flow computation
2Measurement precision
If redundant comparisons are performed for each pixel, then measurement precision is maintained, but use of energy increases
Solution Approach 1:
The patent reduces energy consumption by inverting the conventional scanning order and utilizing the symmetry property of comparison results. By comparing pixels in the inverted order (bottom-right to top-left), the algorithm can derive conjugate comparison results by simply inverting the forward comparison outputs, eliminating redundant computations and significantly reducing power consumption while maintaining detection accuracy
Solution Approach 2:
The patent discards redundant comparison operations by deriving conjugate results through inversion rather than performing them explicitly. The symmetry property allows the system to recover conjugate comparison results from forward comparison results, thereby eliminating unnecessary computational energy expenditure while preserving measurement precision for intensity change detection
3Device complexity
If systematic processing scheme is used to track kernel movement, then device complexity is reduced, but measurement precision may be affected
Solution Approach 1:
The patent applies preliminary action by establishing a systematic processing scheme that tracks the movement and position of the kernel window through the image. By pre-determining the scanning path and position of each pixel relative to the kernel center, the system can efficiently identify which pixels require comparison and which can be derived through inversion, reducing hardware complexity while maintaining comparison accuracy through careful positioning logic
Data Source
AI summary
Systems and methods for performing Census Transforms that includes an input from an image, with a support window created within the image, and a kernel within the support window. The Census Transform calculations and comparisons are performed within the kernel windows. A new Census Transform is disclosed which always inverts a previously made comparison. This new approach can be demonstrated to be equivalent to, applying the original Census Transform, on a pre-processed input kernel, where the pre-processing step adds a fractional position index to each pixel within the N×N kernel. The fractional positional index ensures that no two pixels are equal to one another, and thereby makes the Original Census algorithm on pre-processed kernel same as the new Census algorithm on original kernel. The hardware design for this new Census Transform kernel allows for an always invert of previous comparison system resulting in reduced hardware and power consumption.