Study On Fast Image Matching Algorithm With Triangle Constraint And Feature Points Selection Method

Image matching is one of important technologies in the field of computer vision.It is indispensable in many applications such as target location,visual navigation,and 3D reconstruction,etc.As the key to image matching technology is to improve the accuracy and speed of algorithms,existing algorithms still have a large development space.For instance,image matching algorithms based on local feature points have the advantage of robustness,and some algorithms also use the geometric constraints between feature points,but it is difficult to achieve real-time performance.In addition,feature points are generally divided into spots and corners,while practical applications often use a certain type of feature point,rather than for different images to choose more suitable feature point type.Finally,feature points detected in the image matching process often have the problem of excessive number and uneven distribution,which not only consumes more time but also increases the probability of false matching.In order to improve the accuracy and speed of image matching algorithms,this paper studied the above three problems separately:(1)This thesis proposes an improved rapid triange algorithm(RTM)based on binary description.The RTM algorithm used floating point descriptors to describe triangles while the improved algorithms use FREAK and rBRIEF respectively to describe triangles.The improved algorithms are tested by simulation images and actual images.The experimental results show that compared with the original RTM algorithm,improved algorithms can improve the precision and speed,and reduce the memory occupied greatly.Moreover,it can match feature points not matched by the original algorithm in structural scene,and has a certain degree of robustness.Thus,improved algorithms can be used for the actual binocular image matching application.(2)This thesis proposes a feature point type selection method based on edge detection.For an image to be matched,firstly,the edge rate defined in this paper is calculated based on Canny edge detection algorithm.Then the structural degree of the image is judged according to the relationship between the edge rate and thresholds.The thresholds include a high threshold and a low threshold.Finally,the feature point detection method suitable for the image is determined: if the edge rate is greater than the high threshold,it is considered that the structural information of the image is very noticeable,algorithm should use corner detection;if the edge rate is lower than the low threshold,the structural information of the image is less obvious,spot detection should be used;if the edge rate is between the two thesholds,the feature of image is not obvious,corner or spot detection can be used.The high threshold and the low threshold of the edge rate are obtained by testing simulation images.Finally,the validity of the method is verified by the actual images.The method can be used to realize the intelligent selection of the feature points' types at the beginning of the image matching algorithm.It has practical application significance.(3)This thesis proposed a feature point selection method based on image local entropy and feature points' response.In this method,sub-regions are obtained by dividing the target image,and then the local entropy and average entropy of the sub-regions are calculated.In regions where the local entropy is larger than the average entropy,a certain proportion of feature points with larger response are extracted.In the rest,the feature point with the strongest response is extracted.All the feature points selected are described and matched with the feature vector set of the reference image.Results show this method can not only reduce the number of feature points,but also ensure the uniformity of distribution.It can improve the speed and precision of matching algorithms by combining with them.