车载移动测量系统数据配准与分类识别关键技术研究

发布时间：2018-06-23 00:01

本文选题：车载移动测量系统 + 数据配准　；参考：《武汉大学》2014年博士论文

【摘要】：车载移动测量系统集成了激光扫描仪、数码相机、GPS、里程计、惯性测量单元等多种传感器,不仅能快速获取包含物体空间坐标信息的激光扫描数据,还能获取包含丰富纹理信息的光学影像,具有数据获取速度快、场景目标丰富的特点,已成为一种新型、集成、高效的空间信息获取的技术手段,广泛应用于地理国情监测和智慧城市的各项建设。由于光学影像和激光扫描数据对目标物的描述存在诸多互补性,为了提高城市场景三维可视化效果,需要发展高效的光学影像与激光点云的配准方法。另外,车载移动测量系统获取的点云数据具有海量特性,点云数据处理耗时长、计算量大,再加上场景复杂,不同目标分类的自动化和智能化程度低,这些问题限制了车载移动测量系统在移动测图、基础测绘等应用的实际功效。本文针对上述存在的问题,重点研究和探讨车载激光点云和全景影像的精确配准方法,以及激光点云的高效分类识别技术,主要进行了以下几方面的工作： 1.总结了当前国内外车载移动测量系统在光学影像与激光点云数据配准、车载激光点云分类方面的研究进展。对现有的相关技术和研究方法进行归纳并分析其优缺点,针对现有数据配准及分类中的不足和难点,确定了本文的研究目标与内容。 2.阐述了车载移动测量中的相关基本原理。包括系统的组成及工作原理,分析了各传感器间数据的流转关系；介绍了全景影像的成像原理,对鱼眼镜头成像的光学基础、成像过程、鱼眼图像变形特征及纠正算法进行分析；阐述了车载激光扫描仪测距的原理、方式及其对三维空间的表征方式。 3.提出了一种基于投影回归的全景影像采集和激光扫描配准方法。根据光路传播的可逆原理,把鱼眼镜头成像理解为一个以投影中心为光源点的光线发射过程,提出通过把鱼眼镜头的光源和激光扫描仪的光源进行归一重合,实现鱼眼图像与激光点云配准的思想,对投影回归方法的正确性进行了理论证明。结合具体的实验数据,对鱼眼图像纠正、配准、拼接、分割并最终纹理映射成球形全景,对激光点云数据进行了配准、除噪和简化处理,并把处理好的点云投影到与球形全景图共球心的虚拟投影球面上,实现了全景影像与激光点云的配准。该配准方法不受区域灰度及几何特征信息的制约。 4.配准量测精度的验证研究。在全景影像与激光点云配准的基础上,提出了一种采用角度逼近法来获取球面投影中距离待测点最近邻的激光点的方法,把对影像上的量测转换为对最近邻激光扫描点坐标的查询与计算,对角度逼近法的正确性进行了数学证明,并针对不同角度分辨率下激光点云与全景影像的配准精度进行了验证与分析,结果表明采用投影回归原理的配准方法具有较好的精确度。 5.基于知识与特征图像的点云分类研究。采用横轴圆柱投影和正射投影分别生成了点云的空间特征图像、回波强度图像和颜色特征图像,基于知识和特征图像对行道树点云提出了分层投影、叠加分析的分类方法,即通过高程阈值对树干点云和树冠点云分别投影,得到相应的特征图像,再根据灰度值把特征图像转换成二值图像,并对二值图像进行叠加与分析,最后通过知识滤波来提取行道树点云数据。论文结合具体数据进行了分类实验,实验结果表明,该方法对一些混杂在行道树的噪声信息处理的比较好,且该方法把三维点云转化为二值图像进行处理,有效避免了大量的几何运算,显著降低了分类算法的复杂度。说明采用分层投影、叠加分析的方法对行道树点云提取是行之有效的。 6.基于机器学习的点云分类研究。以激光点云对象的原始特征为基础,通过对其周边点群的上下文语义环境进行分析,充分利用点云的空间分布特征及其局部几何特征,归纳计算了点云对象的新特征,最终构建了由17个特征组成的点云特征向量。采用支持向量机和人工神经网络模型对行道树分类识别进行实验,在支持向量机分类过程中,为了提高模型的泛化能力,分别采用粒子群优化算法和遗传算法对模型参数进行寻优,分析了不同算法的学习曲线特征,最终采用粒子群优化算法,针对不同训练样本、不同特征向量进行了点云分类识别的系列实验。论文选取支持向量机最优分类结果所对应的实验条件,采用人工神经网络的方法对点云分类进行实验对比验证,两种方法均取得整体较为满意的实验结果,表明了机器学习方法在车载激光点云自动分类中的适用性,为提高车载激光点云分类自动化程度和智能化水平提供了新的思路。
[Abstract]:The vehicle mobile measurement system integrates a variety of sensors, such as laser scanner, digital camera, GPS, odometer, inertial measurement unit and so on. It can not only quickly obtain the laser scanning data containing the space coordinate information of the object, but also obtain the optical image containing rich texture information, which has the characteristics of fast data acquisition and rich scene target. As a new, integrated and efficient technology for obtaining spatial information, it is widely used in geographical conditions monitoring and the construction of intelligent cities. The description of objects in the optical image and laser scanning data has many complementarities. In order to improve the effect of the three-dimensional visualization of the city scene, it is necessary to develop high efficient optical images. In addition, the point cloud data obtained by the vehicle mobile measurement system have massive characteristics, the point cloud data processing is time-consuming, the amount of computation is large, the scene is complex, and the automation and intelligence of different target classification are low. These problems restrict the application of mobile measurement system in mobile mapping and basic surveying and mapping. Practical effect.
Aiming at the above problems, this paper focuses on the research and Discussion on the accurate registration method of the vehicle laser point cloud and panoramic image, as well as the efficient classification and recognition technology of the laser point cloud.
1. summarize the research progress of the vehicle mobile measurement system at home and abroad on the registration of optical image and laser point cloud data and the classification of the on-board laser point cloud. The existing related technologies and research methods are summed up and analyzed their advantages and disadvantages. In view of the shortcomings and difficulties of the existing data registration and classification, the research objectives of this paper are determined. And content.
2. the basic principles of the vehicle mobile measurement are described, including the composition and working principle of the system, the flow relationship between the various sensors is analyzed, the imaging principle of the panoramic image is introduced, the optical foundation of the fish eye lens imaging, the imaging process, the distortion characteristic and the correction algorithm of the fish eye image are analyzed, and the vehicular excitation is expounded. The principle and method of optical scanner ranging and its representation of three-dimensional space are discussed.
3. a method of panoramic image acquisition and laser scanning registration based on projection regression is proposed. According to the reversible principle of optical path propagation, the fish eye lens imaging is understood as a light emitting process with the center of projection as the light source. By combining the light source of the fish eye lens and the light source of the laser scavenger, the fish eye is realized. The idea of registration of image and laser point cloud is used to prove the correctness of the projection regression method. Combined with the specific experimental data, the fish eye image is corrected, registered, spliced, segmented and mapped into a spherical panoramic view, and the laser point cloud data are registered, denoising and simplifying, and projecting the treated point clouds to the ball. The registration of the panoramic image with the laser point cloud is realized on the virtual projection sphere of the central panorama. The registration method is not restricted by the information of the gray and geometric features of the region.
4., on the basis of the registration of the panoramic image and the laser point cloud, a method of using angle approximation to obtain the laser point from the nearest neighbor in the spherical projection is proposed, and the measurement of the image is converted to the query and calculation of the coordinates of the nearest neighbor laser scanning point, and the angle approximation method is used. The correctness is proved by mathematics, and the registration accuracy of the laser point cloud and the panoramic image at different angle resolution is verified and analyzed. The results show that the registration method using the projection regression principle has good accuracy.
5. the study of point cloud classification based on knowledge and feature images. Using horizontal cylindrical projection and orthophoto projection, the spatial feature images of point clouds, echo intensity images and color feature images are generated respectively. Based on knowledge and feature images, the hierarchical projection of the road tree point cloud is put forward, and the superposition classification method is used, that is, to the tree trunk through the elevation threshold. The point cloud and the crown point cloud are projected separately, and the corresponding feature images are obtained. Then the feature images are converted into two value images according to the gray value, and the two value images are superimposed and analyzed. Finally, the data of the street tree point clouds are extracted by the knowledge filtering. The paper is classified by the concrete data. The experimental results show that the method is mixed with some of the data. The noise information processing of the miscellaneous tree is better, and this method transforms the three dimensional point cloud into two value images, which effectively avoids a large number of geometric operations and significantly reduces the complexity of the classification algorithm. It is proved that the method of stratified projection and superposition analysis is effective for the extraction of the road tree point cloud.
6. based on the point cloud classification based on machine learning. Based on the original features of the laser point cloud objects, the context semantic environment of the surrounding point groups is analyzed. The spatial distribution features and local geometric features of the point clouds are fully utilized and the new characteristics of the point cloud objects are summed and calculated. Finally, a point cloud consisting of 17 features is constructed. In order to improve the generalization ability of the model, in order to improve the generalization ability of the model, the particle swarm optimization algorithm and genetic algorithm are used to optimize the model parameters, and the characteristics of the learning curve of different algorithms are analyzed. Finally, the characteristics of the learning curve of different algorithms are analyzed. The particle swarm optimization algorithm is used to carry out a series of experiments on the classification of point clouds for different training samples and different eigenvectors. The paper selects the experimental conditions corresponding to the optimal classification results of support vector machines, and uses artificial neural network to test the point cloud classification by experiment. The two methods have obtained the whole more satisfactory experiment. The result shows the applicability of machine learning method in automatic classification of vehicle laser point cloud, which provides a new idea for improving the automation and intelligence level of vehicle laser point cloud classification.
【学位授予单位】：武汉大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：P208;P234;TP391.41

【相似文献】