In this thesis, we have considered the development of traversability estimation for a vehicle (or robot) in order for the robot to autonomously navigate in unrevealed environments. We take a self-supervised learning approach to estimate traversable region for reducing human supervision. Due to the fact that the characteristic of terrain is unrevealed, we aim to automatically build the traversability estimation model which can maintain and adapt to unknown and changing environments in any circumstance.
We first develop online positive learning framework for detecting traversable region in unstructured outdoor environment. To achieve this goal, we investigate the characteristics of environment by using two different types of sensors: camera and 3D laser scanner. We define robust and effective features as histograms that describe the color and geometric characteristics. In addition, we describe a novel approach to combine two different types of sensor data to reconstruct 3D scenes. This multi-sensor-based information is used for detecting traversable region in online positive learning framework.
We design a self-supervised learning framework for traversability estimation using proposed clustering method called incremental nonparametric Bayesian clustering (INBC). In the self-supervised learning framework, training samples are automatically labeled based on the outcome of the vehicle's interactions while moving certain terrain. These labeled training data represented by color and texture histograms are used as the input data in the online learning. With the labeled training data, INBC allows traversability assessment in real time and determines the number of clusters without any prior knowledge. This method effectively groups unknown regions with similar properties while a vehicle is in motion. The vehicle can be deployed to new environments by automatically adapting to changing environmental conditions.
Lastly, we extend the self-supervised learning framework to conditional random field (CRF) framework. Since the characteristic of self-supervised learning framework makes a robot explore hazardous areas for acquiring labeling data. To prevent a robot from failing, we develop a new approach of obstacle probability map construction. Based on traversable probability map from INBC, we combine two probability maps in augmented Multi-Scale CRF framework.