Building Extraction From LiDAR Data Based On Marked Point Theory

Building recognition and extraction has always been a hot topic in remote sensing and image processing.Building is the main infrastructure of the city,also the principle element of virtual city,digital city.The most important element,it is the basis of the carrier of geographic information.Therefore,the extraction of building objectives is the key to intelligent city,three-dimensional modeling and other associated fields.The airborne LiDAR technology have advantages of accuracy and efficiency on obtaining three-dimensional data of ground targets,and provides the data source for the study presented in this dissertation..At present,building recognition and extraction have made great progress,but there are still some problems such as the texture feature of the remote sensing image and the spatial relationship,the complexity of the building model parameters and the low computational efficiency.Based on the data of airborne LiDAR,this dissertation proposes a contour extraction method based on the marked point process theory and the use of reversible jump Monte Carlo algorithm to solve the problem.The main research contents include:(1)The introduction of the research background and significance of this paper.Respectively taking a building object as ? the low-level combination and,? the abstract object as a whole,the associated building contour extraction research progress and technical difficulties are summarized.The research content and technical route are put forwarded.(2)Modeling the building extraction from DSM as an optimization problem,i.e.,a cost function minimization subjective to certain constraints.The cost function is defined as the total square error between the building models(the particles)and the images,called energy.The energy term for the building contours is constructed by integrating the spatial information and spectral information of the building object into it.On the other hand,aiming at the problem that the parameters in the existing methods are too many and the efficiency of the algorithm is low,the new energy model definition method is explored in the optimization process to improve the efficiency of the algorithm.The energy model of a building contour can be composed of data energy and geometric a priori energy.The data energy is used to determine the single target contour.This paper defines the data energy from the view of the gradient change,which can set the parameters and the threshold of the algorithm effectively by the point multiplication between the vectors.Geometric a priori energy is used to determine the spatial relationship of multiple targets.This paper introduces the neighborhood relationship between the building contours,and effectively limits the spatial relationship of the targets by penalizing the targets with overlapping in a certain range.(3)The building model is taken as a point or particle with some features,called marked point.The global minimal energy status of the particles is created by random point process,called marked point process.It is a simulated sampling,and here is implemented by RJMCMC.The Simulated Annealing is employed in effectively searching the global optimal solution of the energy model.This method involves many parameters,which can be more efficient through the improvement of the energy model and the optimization of the transfer kernels.The performance of the algorithm,including the accuracy of detection and time efficiency,are analyzed.The performance of the algorithm under different scale data is compared.The problem of parameter setting and applicability of the algorithm are explored on the basis of many experimental results.The main conclusions are as below:(1)The data energy term is built with the gradient integral,which effectively reduces the parameters,thus reduces the uncertainty of the threshold setting process.The constraint of the neighborhood of the building avoids the overlapping of the model.(2)The three transfer nuclei can be better adjusted to reduce the data energy and geometric a priori energy.The geometric shape,direction,and location information of the building object can be adjusted accordingly.(3)In the whole algorithm,the temperature parameter of the simulated annealing plays a role for convergence controlling.The higher the setting value is,the slower the convergence is,while the higher the extraction precision is.The number of iterations plays a role for extraction rate.A too small number of iterations will cause less extraction of the buildings,while a too large number of iterations will on the contrary lead to the over extraction and less accurate location.The time efficiency will get lower.(4)The effectivity of data on the results of the extraction:In the area where the spatial distribution of the building is uniform and the size is close,the applicability is very good and the extraction rate of 90%can be achieved.In the region with big difference in building size,the parameter setting has a great influence on the extraction result,and the convergence time is longer.It is mainly because of the high global energy in this case,which results in difficulty for achieving models ofvery different sizes.thus the birth and death transfer kernel acceptance rate is always higher than non-jump transfer kernals.