Affiliation: Project performed in Prof. Brian Williams MERs lab in MIT CSAIL. I was sponsored by the Research Science Institute 2012.
Project Title: Active Detection of Drivable Surfaces in Support of Robotic Disaster Relief Missions
Project Description: Over the past few decades, the usage of unmanned vehicles has grown exponentially, expanding into applications such as the automation of industrial processes and automobiles. However, their utility has often been limited by operational concerns. Fully controlled unmanned vehicles require multiple human operators, while their fully autonomous counterparts lack the ability to handle the complex maneuvers necessary in natural disaster relief and/or search and rescue situations. Semi-autonomous UAVs offer a feasible compromise between the two extremes. In this scenario, an unmanned aerial vehicle (UAV) sends birds-eye images of the terrain beneath it to a computing cluster, which will identify easily traversable terrain and generate a path of least risk to an unmanned ground vehicle (UGV). If the path's risk is below a certain threshold, then the UGV will be permitted to proceed on its own. Otherwise, a human operator will be notified, so that he or she may control the UGV directly until it exits the most dangerous terrain. This paradigm allows a single operator to manage several UAVs simultaneously.
This paper focuses on developing an algorithm to identify easily traversable terrain. We focus mainly on roads since they are one of the simplest types of terrain for a UGV to cross. We first generate an approximate analysis of a terrain's roadmap by taking in aerial image data, examining each pixel's color, and completing edge detection. The color analysis in this work employs color normalization, "Histogram of Oriented Gradients'' (HOG) features, and a Support Vector Machine (SVM) learning algorithm. By itself, this approach can effectively group objects of similar appearance in the aerial picture, but it does not distinguish roads from similar objects, such as rooftops. Nonetheless, the color analysis provides a preliminary map of the potential locations of most of the roads in an image.
The guiding principle behind this paper is that a single pixel may contain a high amount of noise. By analyzing multiple pixels that we believe to be part of a road, we reduce the collective noise. In short, we recognize that roads go to a destination and are therefore continuous structures. To take advantage of this design philosophy, we improved our road analysis by implementing the problem as a Markov Random Field (MRF), where each pixel is not only dependent on its own characteristics, but on those of its surroundings. Gaussian Process Machine Learning (GPML), a form of regression, further refines our results and also provides a confidence level in our road identification. We found that this method enabled us to remove most of the non-road objects. Our risk analysis provides a reliable framework for passing down navigation instructions to unmanned ground vehicles.
Please click here to download the paper.