Research On Electronic Image Stabilization Algorithm Based On ORB Feature Matching

Coal mine camera system is easy to be disturbed by the carrier's own jitters and the external environment when shooting videos,resulting in the instability of the obtained video surveillance picture.Such unstable video sequence also brings some difficulties to mine safety production,alarm after the accident and subsequent image processing etc.Therefore,improving the stability of the video for coal mine safety operations has the innegligible significance.The aim of image stabilization technology is to reduce the interference that caused by camera equipment's shake itself on the captured video,In this paper,the basic principle of the electronic image stabilization algorithm is studied,and the classical algorithms used in the main links are analyzed.The electronic image stabilization algorithm based on feature matching is the key research object,the main contents and results of the study are as follows:In the first two chapters,this paper deduces the imaging principle of camera under free movement and discusses the basic principle of image stabilization system and the transformation model of image motion.In the third chapter,this paper has carried on the principle analysis as well as the summary of the advantages and disadvantages of the classical algorithms used in the motion estimation,motion filtering and motion compensation.The precision and the real-time of the motion estimation algorithm have an important influence on the image stabilization system.In chapter 4,this article focuses on the study on the estimation of local motion vectors and global motion vectors,and proposes an improved motion estimation algorithm based on ORB feature matching.For video sequences containing foreground interference,the algorithm first partitions the image and then detects and marks the motion region using continuous three-frame difference method based on the image block.Then,using the ORB algorithm to detect the feature points on the reference frame and the current frame and removing the feature points located in the foreground motion area in combination with the result of the foreground mark.For the uneven distribution of feature points caused by ORB algorithm's feature extraction,this paper proposes an improved method to extract significant feature points by counting the Harris response values in its sub-blocks,keeping the points with larger R value as the global significant feature points.After obtaining the significant feature points,the ORBalgorithm is used to describe and match the feature points.RANSAC algorithm is used to eliminate the pairs of mismatched points.Finally,the least square method is used to obtain the global motion vector.In the fifth chapter,this paper studies the motion filtering and the image compensation in the stable image.The main research is the application of the Kalman filter algorithm in the image stabilization system.The influence of the noise parameter setting in the Kalman filter on the filtering effect is analyzed.Aiming at the problem of the filter divergence in the traditional Kalman filter,a Sage-Husa filtering method is proposed.This algorithm uses the observed value to make continuous correction and real-time estimation to the predicted value,and based on this,an adaptive smoothing filtering algorithm is proposed.By correcting the minimum mean square error matrix in the prediction stage,the tracking ability of the filtering algorithm under the sudden change of the motion state finally achieves the purpose of reducing the error of the traditional model and improving the filtering accuracy.Then,the bilinear interpolation method is used to compensate the filtered image to obtain a stable video sequence.Finally,the method of evaluating the performance of the image stabilization is given and the real-time performance and validity of the image stabilization algorithm proposed in this paper are verified through a series of related experiments.