For Safer Navigation: Pedestrian-View Intersection Classification

Abstract

Intersection classification is one of the key components of autonomous navigation. Several related research works have been conducted as a prior task to solve problems such as autonomous driving, aircraft hovering, and navigating in mines. However, to the best of our knowledge, none of these studies support pedestrian-view navigation to guide the small and slow robots for which it is too dangerous to be operated on normal roads along with normal vehicles. To address this problem, we propose: 1) a pedestrian-view-level intersection classification image dataset, 2) ResNet-based architecture fine-tuned on the proposed dataset, and 3) thorough experimentation to explore the capabilities of our proposed architecture. The detailed analysis reported in this paper enabled us to find the network configuration that is neither underfit nor overfit to our data and achieves 80% test accuracy.