Active learning for outdoor perception

Many current state-of-the-art outdoor robots have perception systems that are primarily hand-tuned, which makes them difficult to adapt to new tasks and environments. Machine learning offers a practical solution to this problem. Assuming that training data describing the desired output of the system is available, many supervised learning algorithms exist for automatically adjusting the parameters of possibly complex perception systems. This approach has been successfully demonstrated in many areas, and is gaining momentum in the field of robotic perception. An important difficulty in using machine learning techniques for large scale robotics problems comes from the fact that most algorithms require labeled data for training. Large data sets occur naturally in outdoor robotics applications, and labeling is most often an expensive process. This makes the direct application of learning techniques to realistic perception problems in our domain impractical. This thesis proposes to address the data labeling problem by analyzing the unlabeled data and automatically selecting for labeling only those examples that are important for the classification problem of interest. We present solutions for adapting several active learning techniques to the specific constraints that characterize outdoor perception, such as the need to learn from data sets with severely unbalanced class priors. We demonstrate that our solutions result in significant reductions in the amount of data labeling required by presenting results from a large amount of experiments performed using real-world data.