Hardware Accelerator for Real-Time HOG Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computational power constraints in embedded devices make real-time calculation of oriented gradients for applications like histogram of oriented gradients (HOG) in embedded computer vision challenging, consuming most of the available processing power.
Innovation Solution
A hardware accelerator is designed to compute gradients, determine bin identifiers, and calculate magnitudes for pixels, including a gradient computation component, bin identification component, and magnitude component, which are integrated into a system-on-a-chip (SOC) to accelerate HOG computation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If oriented gradient calculations are performed using traditional processors in embedded devices, then computation accuracy is maintained, but computational power is exhausted and real-time processing becomes infeasible
Solution Approach 1:
The patent replaces traditional general-purpose processors with a dedicated hardware accelerator that uses specialized circuitry (logic circuits, arithmetic logic units) to perform gradient calculations. This substitution of mechanical/computational systems enables real-time HOG computation by dedicating specific hardware resources to the calculation task, thereby reducing overall computational power consumption while maintaining accuracy.
Solution Approach 2:
The hardware accelerator is divided into multiple independent calculation units that can process different pixels or gradient components simultaneously. Each unit handles specific aspects of the gradient calculation (Gx, Gy, magnitude, orientation), allowing parallel processing that dramatically improves productivity while keeping each individual unit's power consumption manageable.
2Measurement precision
If HOG computation is performed for large images in embedded applications, then detection accuracy is improved, but computational load increases to the point of exhausting all available processing power
Solution Approach 1:
By replacing software-based HOG computation with dedicated hardware acceleration, the system can process large images at real-time speeds. The hardware accelerator maintains the full computational precision needed for accurate object detection while providing the throughput necessary to handle large image sizes without exhausting processor capabilities.
Solution Approach 2:
The hardware accelerator pre-computes gradient magnitudes and orientations for all pixels in parallel before histogram generation. This preliminary action on gradient calculation enables subsequent histogram computation to proceed efficiently, maintaining detection accuracy for large images while managing overall processing throughput.
Data Source
AI summary
A hardware accelerator for histogram of oriented gradients computation is provided that includes a gradient computation component configured to compute gradients Gx and Gy of a pixel, a bin identification component configured to determine a bin id of an angular bin for the pixel based on a plurality of representative orientation angles, Gx, and signs of Gx and Gy, and a magnitude component configured to determine a magnitude of the gradients Gmag based on the plurality of representative orientation angles and the bin id.