Obstacle detection for intelligent vehicles using semi-supervised active learning

Reliable environment perception system is critical to path planning and autonomous navigation of intelligent vehicles. One feasible way to percept environment is obstacle detection by classifying image patches as obstacle or non-obstacle. Accurate classification system depends on appropriate training data. For intelligent vehicles, a large number of images can be easily obtained while labeling them is tedious. Additionally, the accuracy is limited for the scene diversity. In this paper, we propose a semi-supervised active learning algorithm which can exploit the most certain unlabeled examples and query the most informative examples to enhance the performance of classifiers. In view of the scene diversity, we present a two-level classification system which first distinguishes the scene category using level-I classifier before calling the suitable level-II classifier to detect obstacles. The experimental results demonstrate the efficiency of our algorithm and two-level classification system.