An Implementation Of FPGA Algorithm Base On FPGA

Object detection has always been an important research direction in the field of machine vision.Researchers have proposed many excellent image processing algorithms,such as SIFT,SURF and HOG.Among them,the object detection algorithm based on local adaptive regression kernel(Local Adaptive Regression Kernel,LARK)has the advantages of high recognition accuracy,no training and high robustness,which make it have a broad application prospect.But the calculation progress of LARK algorithm is complicated.The demand of computing resources is huge,so it is difficult to implement in real time.This paper proposes a complete LARK algorithm implementation architecture based on the FPGA platform,covering all the implementation steps of the algorithm to achieve real-time computing ability while keeping other performance of the algorithm unchanged.This paper focuses on three architectural modules: feature extraction module,feature dimensionality reduction module and similarity measurement module.(1)The feature extraction module mainly includes image gradient extraction,covariance matrix calculation and local steering kernel extraction.(2)The feature dimension reduction module mainly includes singular value decomposition and feature remapping.In the feature dimension reduction module,the feature value decomposition hardware architecture based on one-sided Jacobi method is designed.(3)In the similarity measure module,the implementation architecture of norm solving is designed.The out-of-chip DDR multi-column buffer recovery technique is proposed,which reduces the requirement of two-dimensional FFT transformation of large-scale images for on-chip storage space.Through functional simulation and experiment verification,the FPGA implementation architecture designed in this paper can maintain the same recognition performance as the original algorithm.With the query image resolution up to 160×160,the real-time processing speed of 35 frames per second is achieved for the 320×320resolution video,which effectively solves the real-time problem of the algorithm.