Abstract:Visual feature extraction is a key technology of computer vision for intelligent video processing. Efficient feature extraction is a fundamental problem in computer vision applications. Scale-Invariant Feature Transform (SIFT) is one of the most popular feature extraction algorithms because SIFT features are invariant to image scale and rotation and robust to changes in illumination and noise. However, SIFT is a computationally-intensive and power-hungry algorithm, which needs to be accelerated by efficient hardware design to achieve both high-speed feature extraction and high energy efficiency for many high frame-rate video applications at Artificial-intelligent Internet of Things edges. In this work, an energy-efficient SIFT based feature extraction accelerator is proposed. In the Gaussian pyramid and Differences of Gaussian (DoG) pyramid construction process, three design methods are proposed to reduce power consumption and improve information fidelity: a fast and slow dual clock domain design method with a reconfigurable design strategy is proposed to reduce the computation resources; a partial sum reuse design method is proposed to further reduce the computation resources and the amount of computation; a dynamic padding design method is proposed to solve the problem of information loss at image edges and corners after convolution operation. In the keypoint descriptor generation process, an optimized algorithm using circular region and polar coordinates is proposed to parallelize the main orientation assignment and descriptor generation to achieve high-speed processing, while maintaining a comparable matching accuracy with the state-of-the-art designs. The experiment results show that the proposed SIFT hardware accelerator is able to extract features by up to 162 frames per second ( $640\times 480$ pixels) under 100 MHz, with the power consumption of 364.26 mW and energy efficiency of 2.25 mJ/frame based on 180 nm technology, which is suitable for many high frame-rate AIoT applications including autonomous driving cars and unmanned aerial vehicles.

An Ultra-Fast and Low-Power Design of Analog Circuit Network for DoG Pyramid Construction of SIFT Algorithm

ASP-SIFT: Using Analog Signal Processing Architecture to Accelerate Keypoint Detection of SIFT Algorithm

A Reconfigurable Convolution-in-Pixel CMOS Image Sensor Architecture

An Energy-Efficient SIFT Based Feature Extraction Accelerator for High Frame-Rate Video Applications.

A Precision-Improved Processing Architecture of Physical Computing for Energy-Efficient SIFT Feature Extraction

Implementation of Parallel Acceleration for Real-time Extraction of Visual Features.

Physical Computing With No Clock to Implement the Gaussian Pyramid of SIFT Algorithm

A Fast and Power-Efficient Hardware Architecture for Visual Feature Detection in Affine-SIFT.

Physical Computing Circuit with No Clock to Establish Gaussian Pyramid of SIFT Algorithm

An Efficient Hardware Architecture of the Optimised SIFT Descriptor Generation.

A Hardware Accelerator with Variable Pixel Representation & Skip Mode Prediction for Feature Point Detection Part of SIFT Algorithm.

A Hardware Accelerated Scale Invariant Feature Detector For Real-Time Visual Localization And Mapping

A real-time embedded architecture for SIFT

A 83fps 1080P Resolution 354 Mw Silicon Implementation for Computing the Improved Robust Feature in Affine Space

An Optimized Structure on FPGA of Key Point Detection in SIFT Algorithm

A Novel Hardware-Software Co-Design and Implementation of the HOG Algorithm

Energy Efficient ApproxSIFT Implementation for Image Mosaic with Approximate Computing Technologies.

A 181 GOPS AKAZE Accelerator Employing Discrete-Time Cellular Neural Networks for Real-Time Feature Extraction

Efficient Hardware Optimization Strategies For Deep Neural Networks Acceleration Chip

A 127 fps in full hd accelerator based on optimized AKAZE with efficiency and effectiveness for image feature extraction

Parallelization And Optimization Of Sift On Gpu Using Cuda