Automated Registration of Multi-source, Multi-sensor Data
The automatic registration of multi-source airborne and space-borne imaging and ranging data has generated much research interest in remote sensing and digital photogrammetry. This is driven by the increasing availability of large volumes of Earth observation data, and the need for automated integration of multi-sensor, multi-resolution data to generate redundant and complementary spatial information products for many applications, especially in feature extraction and building reconstruction. Conventional registration methods rely on physical correspondences and invariably fail in the registration of data acquired from different types of sensors. This research aims to develop a robust technology for registration of multi-source, multi-sensor data to facilitate improved data integration. The research explores statistical dependence between data. It investigates the joint probability density function of data sets to measure the predictability of one from the other, thus facilitating registration of remote sensing data collected using either the same or different types of sensors. Therefore, the concepts of statistical similarity such as Mutual Information (MI) are investigated as a foundation for the registration process. The inherently registered intensity and height information of Light Detection And Ranging (LiDAR) data has been exploited to improve the robustness of the similarity measures through the particular multivariable mutual information definition called Normalised Combined MI. It utilises both the elevation and intensity information of LiDAR data simultaneously for registering imagery to 3D LiDAR point clouds and improves the robustness and performance of statistical similarity to find the correct transformation of data sets. In addition, the local similarity measure, which relies on determining the similarity of only small parts of the data sets, called templates, enables the registration process to support a more complex transformation between the data sets. The computation cost is thus decreased, and the reliability of registration is improved through an overdetermined solution of the transformation parameters. In order to improve the relationship between the similarity value and the transformation, the appropriate parameters of the joint PDF, such as bin-size and smoothing kernel, have to be determined. This increases the robustness of registration through provision of a convergence surface of the similarity measure with less local maxima, which speeds up the optimisation process. The proposed registration approach has been applied to the registration of satellite and aerial imagery data, with different resolutions from 50cm to 5 cm, to LiDAR data with the densities ranging from 0.5-35 pts/m2. The experimental results obtained demonstrate that the alignment of optical imagery to 3D LiDAR point clouds via the improved intensity-based method can yield greater accuracy than that produced by conventional feature-based registration approaches.