Design Of SIFT Feature Detection And Describe ASIC

The SIFT operator has been widely used because of its good invariance for rotational transformation,scale transformation,affine transformation and gray scale variation.However,with the rapid development of video processing technology,the image processing algorithm has a higher real-time requirements,which can not be satisfied by the traditional SIFT operator due to its large computational complexity.SIFT includes two parts: feature point detection and vector generation.The feature point detection part includes three parts: Gaussian difference space structure,extreme point detection and anomaly point elimination.And the direction and amplitude of the image gradient should be calculated before performing the eigenvector description.The feature vector description part mainly includes the main direction histogram statistics,the main direction of the proposed,the eigenvector histogram statistics and eigenvector normalization of the four parts.In order to meet the real-time nature of the algorithm,the appropriate parallel structure should be applied according to the characteristics of the hardware.In this paper,the principle of SIFT algorithm is analyzed in depth.Based on the existing research of our laboratory,the SIFT feature point detection part is further improved.At the same time,the RTL code of feature vector generation part was developed,and through the analysis of data processing rate of each operation.The main work includes:(1)Proposing and realizing the high precision gradient direction amplitude calculation structure;(2)Perfecting the scanning mechanism of feature statistics;(3)Proposing a normalized structure of multiple frequency multiplexing;(4)Proposing an accelerated feature matching structure;(5)Building a complete simulation test platform based on Modelsim and MATLAB.Experiments show that the average processing time for each feature point is 8.87us(clock frequency is 100Mhz).In this paper,the feature detection and detection structure can reach the processing capability of a feature point in each clock cycle.The processing time of each feature point of the feature vector generation module is 8.87 us,which can be realized in every second 30 frames of the speed of the average number of feature points not more than 3750 image sequence for real-time processing.