Machine Learning Algorithms and Applications for Lidar, Images, and Unstructured Data

Aerial imagery of geographic regions in the form of Lidar and RGB images aids different tasks like survey, urban-planning, mapping, surveillance, navigation, localization and others. Most of the applications, in general, require accurate segmentation and identification of variety of objects. The labeling is mostly done manually which is slow and expensive. This dissertation focuses on roads as the object of interest and aims to develop methods to automatically extract road networks from both aerial Lidar and images. This work investigates deep convolutional architectures that can fuse the two types of data for road segmentation. It presents a design which performs better than the state-of-the-art RGB-only methods. It also describes a simple, disk-packing based algorithm which translates the road segmentation into a OpenStreetMap-like road network graph while giving improved accuracies in terms of connectivity, topology and reduction in outliers. This dissertation also presents a truth finding algorithm based on iterative outlier removal which can be used for reaching a consensus when information sources or ensembles of trained machine learning models are at a conflict. In addition, it introduces a full and published book on Python programming based on the experiences this research provided. The hope is to contribute towards teaching and learning Python.