Extracting Building Footprints From Digital Measurable Images

Building footprint extraction reported in the current literature is mainly based on methods using remote sensing or Li DAR data.Little effort has been made to extract the building footprint from the closerange image.Using remote sensing data to extract the building footprint requires high resolution images which are often expensive and not easy to obtain.Digital measurable image is easier to access and is cheaper to collect than Li DAR data.This thesis develops a novel method to extract building footprints from the digital measurable images(DMI).It is a muti-stage process that integrates many algorithm,mainly including identifying foot-points from the original images,transforming the image coordinates of the indentified foot-points into the three-dimension(3-D)ground coordinate system,removing the outliers among the foot-points if necessary and finally approximating the building footprint in the form of rectangle.First,it locates the foot-points(the points that represent the outline of the building)on the original left image,then it determines the corresponding points of these solved foot-points on the corresponding image based on SIFT image matching algorithm.After that,it uses the direction linear transformation(DLT)model to compute the ground coordinates of the solved foot-points.Finally,after removing the outliers(if necessary)from the foot-points(which are in 3-D ground coordinate system),it then encloses these points with a minimal rectangle,which is considered as the final building footprint.To identify the foot-points,the method uses two different ways,edge-projection and LSD detector.Then it forms the co-relationship between the image pair with the aid of SIFT registration method and computes the image coordinates of the corresponding points related to the right image using the imagepair coefficients.Once it solves the foot-points on both images in the image coordinate system,it uses the DLT model to calculate the three-dimension ground coordinates of the foot-points with the aid of ground control points(GCPs)obtained from the CAD map and determined by ground survey using a total station(TS06).Since the result of the previous stage probably contains outliers,the next step is to remove those outliers from the original points cluster once it transforms the solved foot-points into the 3-D ground coordinate system.After that it choses one from these three different methods to remove the outliers of the foot-points,which are SD remover outlier-remove(see section 3.4.1)method,the modified Thompson's Tau method(Thompson 1985)and the cluster-Thompson remover.Once all the process mentioned above are done,the foot-points are closed to form a rectangle that is subject to a minimal area,since the footprint of the studied building is almost rectangle,see section 3.4.2.The result varies if it uses different methods to determine the foot-points or different means to remove the outliers,see section 4.4.The general results of the proposed algorithm indicate the relative accuracy of the building footprint varies from around 30% to 60%(width),from about 70% to 90%(length)and from 20% to 50%(area).The coordinate accuracy spikes from 3m to 16 m in the X direction and from about 15 m to 30 m in the Y direction.Finally,from the final building footprint and its accuracy,it concludes that this preliminarily study demonstrates a viable solution to extract the building footprint from DMI images or other close-range images.But as shown above,the results of the proposed method indicate the accuracy from 3m to 30 m,which means its further applications might be limited.Therefore,further studies need to be done to ….