Study Of Corner Detection Algorithm Based On Several Contour Curvature Estimation Methods

Corner is the steadily sparse feature of gray image and contains important structure information.Now corner detection has become one of the fundamental topics of image processing,pattern recognition and computer vision,and has played an important role in the image matching,image registration,object recognition and tracking,motion estimation and 3D reconstruction etc.focusing on discrete curvature of digital contour,we design and construct three curvature estimation methods which can well reflect planar curve curvature by related theoretical analysis:(1)angle estimator(two);(2)continuous curvature estimator;(3)relative tangent-to-point distance accumulation curvature estimator.The contributions and results of this thesis are detailed as follows:(1)To overcome the problem in choosing RoS(region of support)of RJ73,we propose a new method to estimate angle(AAE).AAE first parameterize the edge boundary of gray image to two parametric curves based on arc-length,in this way we can convert the angle estimation of edge contour to the slope estimation of arc-length parametric curves respectively.Then by applying LS(least square)or WLS(weighted least square)technique,we present the solution to the slope estimation.(2)AAE presents a new angle estimation method by converting the angle estimation of edge contour to the slope estimation of parametric curves respectively.Also,we can directly estimate the angle of contour.We consider the two half Ro Ss of the target point as two straight lines approximately and treat the angle of the target point as the include angle of the two straight lines.Next we need to calculate the directions of the two straight lines.It is found that the directions of the two straight lines are just the eigenvectors of the two half RoSs respectively.So in this way we propose another new direct angle estimation method-EAE(eigenvetor-based angle estimator).With extensive experiments we found that EAE is not robust to uniform scaling transformation.By some discussions,we present the weighted scheme of EAE(WEAE).(3)We propose a novel approach which addresses the corner detection issue by employing the Chebyshev polynomial fitting to estimate the curvature in a continuous way.First,we parameterize the digital contour of gray image to two parametric curves based on arc-length;then the two digital parametric curves are fitted with two Chebyshev polynomial series respectively which pass the target points.In this way,the curvature estimation of digital edge contour is translated into the derivative calculation of fitted continuous curves.(4)By intuitive observation,we found that the curvature of the target point on a contour is proportional to the distance accumulation of its neighbors to the tangent line of the target point.Based on the observation,we present a new method to estimate the discrete curvature.We first fit the digital curve segments with quadratic polynomials by employing least square technique to derive the tangent line of the target point,then we calculate the distance accumulation of its neighbors to the tangent line,which is a reflection of the discrete curvature.