结构特征与立体特征协同的建筑物识别研究

发布时间：2018-05-22 17:32

本文选题：结构特征 + 立体特征　；参考：《哈尔滨工业大学》2017年硕士论文

【摘要】：建筑物作为重要的人工地物之一,与人们的生活息息相关。在数据挖掘、无人驾驶、精确制导等诸多领域中对于建筑物的识别有着潜在的需求。传统的利用二维数据的建筑物识别技术面对低分辨率影像、结构复杂或形状相似的建筑物时鲁棒性很差,因此利用三维数据对建筑物进行识别具有着重要的研究价值和深远意义。本论文主要研究建筑物的立体识别方法,根据常见建筑物的顶面类型将其分为脊形、人字形、平顶和锥形四类,利用分辨率为0.1m的DSM数据,通过提取建筑物的结构特征与立体特征对其进行立体识别。论文中首先对四类顶面建筑物的三维结构进行了分析,针对目标建筑物DSM数据中含有大量奇异点与噪声点的现象,利用了一种改进的区域增长算法对顶面进行分割,并通过边缘提取技术得到初始拓扑结构。在此基础之上,利用形态学方法对初始拓扑结构进行了孔洞填充,然后利用道格拉斯-普克法对畸形的轮廓线结构进行了几何校正,最后对拓扑结构分别进行旋转、尺度校正,生成了建筑物的结构特征。针对目标识别中的关键问题,即特征的旋转、尺度与平移不变性,论文研究了一种点对直方图特征。首先应用K维树、K邻近检索以及主成分分析方法对目标表面点的法向量进行了估计。根据大型建筑物表面点数量大、高噪声的特性,对适用于小目标识别的立体特征进行了改进,在含有不同高程、尺度以及多角度的目标样本条件下进行了立体识别,并对比了在不同特征参数下、不同噪声环境中的识别结果,验证了该特征对不同类别建筑物的分辨力以及鲁棒性。对于既定目标的个体识别,论文研究了一种3-D形状上下文特征,并利用球谐变换对特征进行了改进,解决了旋转不变性的问题。最后分别在同样样本条件下和不同特征参数条件下进行了特征匹配,获取最优参数与识别精度,也验证了特征的旋转不变性。论文最后对校正的拓扑结构进行了角点与轮廓线的合并,并生成特征向量。利用主成分分析法对点对直方图特征进行了降维。通过协同结构特征与点对直方图特征,在不同噪声环境下对建筑物的类型进行了立体识别。识别结果验证了协同特征的抗噪能力,相比于仅用立体特征识别具有更高的精度以及鲁棒性。
[Abstract]:As one of the most important human site objects, buildings are closely related to people's lives. In many fields, such as data mining, unmanned driving, precision guidance and so on, there is a potential demand for building identification. The traditional building recognition technology based on two-dimensional data has poor robustness in the face of low-resolution images and complex structures or similar shapes. Therefore, the use of three-dimensional data for building identification has an important research value and far-reaching significance. In this paper, the method of stereoscopic recognition of buildings is studied, which is divided into four types: ridged, herringbone, flat-top and conical according to the types of top surfaces of common buildings. The DSM data with a resolution of 0.1 m are used in this paper. The structure features and stereoscopic features of buildings are extracted for stereoscopic recognition. In this paper, the three dimensional structure of four kinds of roof buildings is analyzed. Aiming at the phenomenon that there are a lot of singular points and noise points in the DSM data of the target building, an improved region growth algorithm is used to segment the top surface. The initial topology is obtained by edge detection. On this basis, the initial topological structure is filled with holes by morphological method, then the contour structure of the deformity is corrected by using the Dogas-Puck method, and the topological structure is rotated and calibrated respectively. The structural features of the building are generated. Aiming at the key problems in target recognition, namely, the rotation of feature, the invariance of scale and translation, a point pair histogram feature is studied in this paper. Firstly, the normal vectors of target surface points are estimated by K-dimensional tree K-neighborhood retrieval and principal component analysis (PCA). According to the characteristics of large number of surface points and high noise of large buildings, the stereo features suitable for small target recognition are improved, and the stereo recognition is carried out under the condition of different elevation, scale and multi-angle target samples. The recognition results in different noise environments under different characteristic parameters are compared to verify the resolution and robustness of the feature to different types of buildings. For individual recognition of a given target, a 3-D shape context feature is studied, and the spherical harmonic transformation is used to improve the feature, which solves the problem of rotation invariance. Finally, the feature matching is carried out under the same sample condition and different characteristic parameter condition, the optimal parameter and recognition accuracy are obtained, and the rotation invariance of the feature is verified. Finally, the corrected topology is combined with the contour and the eigenvector is generated. The principal component analysis (PCA) is used to reduce the dimension of dot pair histogram. Based on the features of cooperative structure and point-pair histogram, the types of buildings are identified in different noise environments. The recognition results show that the anti-noise ability of cooperative features is more accurate and robust than that of only stereoscopic features.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P208

【参考文献】