【作者】 程慧；

【导师】 张健沛；

【作者基本信息】 哈尔滨工程大学 ， 计算机应用技术， 2013， 博士

【摘要】 三维数字城市是一项跨学科的系统工程。尽管在理论研究方面已取得一些进展，但在普通计算机上实现实时显示和交互仍有很大难度。探索低成本、高效率的城市三维模型自动重建途径，对信息技术发展尤其是智能城市建设具有重要的理论意义和实践价值。针对开发数字城市系统存在的挑战，本文从图像重建的角度出发，讨论了建筑物重建、DEM数据重建、建筑物定位等问题，取得的研究成果有：1．基于单幅照片的建筑物重建研究。主要研究了单个建筑物的重建方法，一是提出一种改进的小波ICA滤波器，运用ICA规范小波降维后的信号，发现独立噪声的特征，从图像源中分离噪声与信号。二是基于GA-PSO Hough变换提出一种建筑物平面的重构方法，采用有效的限制搜索空间方法以及分层次搜索策略，有效降低了计算量，提高了重构性能与质量。三是提出从3D建筑物图像中纹理提取与映射的方法，通过校正输入图像，融合透视图提取纹理图片；并提出消除透视扭曲的方法，对纹理图像进行校正，通过补偿光源颜色以估计纹理色彩，避免了3D测量的中间过程。2．基于立体图片的城市重建研究。主要研究了地形与建筑物的重建方法，一是根据线阵推扫式影像的几何性质，分析了线阵推扫式影像的核线模型，提出基于核线曲线的共轭匹配方法，解决了共轭搜索算法、相关域的确定和前后相关匹配策略等问题，并通过实验总结分析了算法的优越性。二是利用自动模糊聚类法进行建筑物抽取，运用Hopfield神经立体匹配方法，对经典匹配进行初始化，提高了匹配率并减小了不明确性。三是针对立体图像匹配时任务调度效率低的缺点，提出了一种启发式任务调度算法HCDDSL，优化任务调度的结果，减少任务完成时间。3．基于特征的模型识别与定位研究。主要研究了大场景图片中建筑物的识别与定位方法，一是提出应用小波变换进行输入图像的边提取方法，提取目标的边缘特征。利用多层次小波性质，实现不同种类目标的检测。二是基于几何不变关系，提出一种三维目标识别算法，通过摄像机投影变换，分析投影参数约束，使用两种不变关系，进行视场不变的目标识别。三是应用联合估计模型和同步联合估计模型，提出了一种基于能量的联合运动和视差估计算法，利用左右运动和前一帧的视差来计算联合估计约束当前帧的初始视差。更多还原

【Abstract】 3D digital city is a systematic engineering concerning with multi subjects. Even thoughsome progresses have been made in theoretical research, it is still very difficult with regard toreal-time display and interactive manipulation on general computers. Therefore, for thedevelopment of information technology and the construction of intelligent city, it ismeaningful both theoretically and practically to explore the three dimensional automaticreconstruction method of cities.As far as the challenges in developing digital city is concerned, from imagereconstruction point of view, building reconstruction, DEM data reconstruction and buildinglocation are elaborated in this thesis. The work and results are expressed as follows.1. Reconstruct the building with single frame photography. The main research is thereconstruction method of sigle building. Firstly, an improved wavelet-ICA filter, which canextract the independent noise feature, is proposed. ICA is employed to regularize the signalfrom the wavelet dimensionality reduction of original image and serves to identify the noisefeature. Secondly, GA-PSO is incorporated to improve the Hough transform (HT)performance. Different strategies are introduced to reduce the computational effort, whichinclude efficient limitation of the search space and a hierarchical search method. A modifiedGA-PSO HT algorithm is developed to improve the performance and quality of reconstructionfor planar building recognition techniques. Thirdly, a method dealing with texture extractionand mapping form3D building image is also proposed, in which the3D-building isreconstructed based on the planar layout. Texture images are then extracted by mergingperspective and correcting input images. A new technique is studied and presented for thepurpose of removing perspective distortion and estimating the texture image color tocompensate for the light source color. The method is free of any intermediate process in3Dmeasurement.2. Reconstruct the city with stereo photograph. The main research is the reconstructionmethod of terrain and buildings. On the first place, according to the geometric properties oflinear pushbroom images, the epipolarity model of Linear CCD pushbroom images isanalyzed. A conjugate matching method is proposed on the basis of epipolar curve. The conjugate search algorithm, the correlation region solution and match strategy are solved withthe conjugate matching method. The advantage of this method is validated with experiments.Then features of the building are extracted with proposed Fuzzy threshold method. Hopfieldneural stereo matching method is set up and used to initialize classical matching problems.The proposed Hopfield neural stereo matching method improves matching rate and decreaseambiguities simultaneously. On the third place, the shortcomings of existing task schedulingalgorithms are analyzed, and a new efficient heuristic task scheduling algorithm, namely,HCDDSL is proposed. Through HCDDSL, the result of task schedule may be optimizedeffectively, and the time span of all tasks is also reduced.3. The model recognition and locating are accomplished with the features. The mainresearch is the identification and location method for building in large scene picture. At first,the wavelet transform method is utilized to process the edge extraction. The edge feature ofthe target is extracted through proposed method. Different kinds of targets detection can berealized with multi-level wavelet properties. Secondly, a new3D target recognition algorithmis proposed based on geometrically invariant relationships. The algorithm employs a cameramodel for projective image formation and the projection parameters constraints are analyzed.Besides, two invariant relationships are utilized to complete the target recognition withview-invariant object. Finally, an energy-based estimation algorithm uniting motion and viewdisparity is proposed on the basis of two energy models, which are the joint estimation modeland the simultaneous joint estimation model. The initial disparity of the current frame iscomputed with the difference between side-to-side motion and previous frame.更多还原