Scene Recognition for Urban Search and Rescue using Global Description and Semi-Supervised Labelling

Abstract

Autonomous aerial robots for urban search and rescue (USAR) operations require robust perception systems for localization and mapping. Although local feature description is widely used for geometric map construction, global image descriptors leverage scene information to perform semantic localization, allowing topological maps to consider relations between places and elements in the scenario. This paper proposes a scene recognition method for USAR operations using a collaborative human-robot approach. The proposed method uses global image description to train an SVM-based classification model with semi-supervised labeled data. It has been experimentally validated in several indoor scenarios on board a multirotor robot.