Research Of Data Registration And Classification For Vehicle-borne Mobile Measurement System

Vehicle-borne Mobile Measurement System consists of laser scanner, digital camera, GPS, odometer, inertial measurement unit and other sensors. It can collect the spatial coordinates of the object quickly and obtain the optical images as high resolution texture. Mobile Measurement System has become an important technique for national geographical monitoring and smart city construction. For so many sensors in this platform, there are so many massive data would be collected. The optical images and laser point cloud data contains the attributes of objects and both of them are complementary for each other.To improve the usability of these data, the optical images and laser point cloud data should be registered. Besides, coupled with the complexity of the scene, point cloud data processing is a time-consuming and computationally intensive job. The level of auto-classification of objects is also relatively low in the processing. All of the problems mentioned above limits the further application of Vehicle-borne Mobile Measurement System.To resolve the problems above, this paper focus on the methods of images and point cloud data registration, point cloud data auto-classification. This paper mainly covers the following work:1. This paper summarizes the current research of data processing in Vehicle-borne Mobile Measurement System, include the registration of optical images and point cloud data, point cloud classification etc. This paper also analyzes the advantages and disadvantages of relevant technique and determines the research objectives of the paper.2. The author introduces the fundamental principles of Mobile Measurement System, including the composition and working mechanism, and analyzes the relationship among all the sensor data. This paper also describes the optical principle and the imaging process of fisheye lens, and analyses deformation characteristics and correction algorithm of the fisheye images. The principle of laser scanning and the coordinate system of scanner are also mentioned in the chapter;3. This paper presents a new method to match the panoramic images with laser point cloud data based on projection regression. According reversible principle of optical transmission, the imaging process of fisheye lens is interpreted as a light emission from the projection center,then the light source of fisheye images and laser scanner would be process to the same point to realize the registration of fisheye images and point cloud data. The correctness of this method was theoretically proved in this paper. Combined with specific experimental data, the author describes the whole process of creating global panoramic image, including correcting,registering,stitching,splitting and texture mapping. The author also describes the whole process of how to deal with the point cloud data to realize the registration,including data registration,de-noising,raster processing and mapping into the surface the virtual global. The method would not be influenced by gray and geometric characteristics of data;4. Validation of registration accuracy. This paper proposed a angle approximation approach to measure the distance in the panoramic image based on the registration between panorama and point cloud data. The correctness of this method was also theoretically proved in the chapters. Some experiments were done for different angular resolution and results show that the errors are very small and this method has great practicality;5. Point cloud classification experiments based on knowledge and feature images. A series of feature images, including spatial feature image,reflected intensity feature image and RGB feature image, are generated by using horizontal cylindrical projection and orthographic projection. Classification methods of street tree point cloud are brought by using layered projection and overlaying analysis,and the classification experiments are done with specific data. The results show that this method is effective on the extraction of tree point cloud.6. Point cloud classification experiments based on machine learning methods. Object recognition is an important method of vehicle-borne laser scanning point clouds processing technique. The traditional method is splitting them before classification. This paper proposes a new method to realize object recognition according to context semantic environment and analysis of vehicle-borne laser scanning point clouds original features and creates a feature vector with17features. Taking roadside trees recognition in Wuhan City as an example, the article calculates the feature vector comprised of17features and uses SVM(Support Vector Machine) to classify. Particle swarm optimization algorithm and genetic algorithm are used separately to optimize the model's parameter in the classification process. The influence on the classification accuracy of different features and different numbers of training samples is studied through a series of experiments.The article also uses ANN(Artificial Neural Networks) to classify the point clouds of roadside trees. Experiments show that the machine learning method provides a promising solution for classifying objects from vehicle-borne laser scanning point clouds.