Viewpoint-Tolerant Semantic Segmentation for Aerial Logistics

Abstract

Semantic segmentation is fundamental for enabling scene understanding in several robotics applications, such as aerial delivery and autonomous driving. While scenarios in autonomous driving mainly comprise roads and small viewpoint changes, imagery acquired from aerial platforms is usually characterized by extreme variations in viewpoint. In this paper, we focus on aerial delivery use cases, in which a drone visits the same places repeatedly from distinct viewpoints. Although such applications are already under investigation (e.g. transport of blood between hospitals), current approaches depend heavily on ground personnel assistance to ensure safe delivery. Aiming at enabling safer and more autonomous operation, in this work, we propose a novel deep-learning-based semantic segmentation approach capable of running on small aerial vehicles, as well as a practical dataset-capturing method and a network-training strategy that enables greater viewpoint tolerance in such scenarios. Our experiments show that the proposed method greatly outperforms a state-of-the-art network for embedded computers while maintaining similar inference speed and memory consumption. In addition, it achieves slightly better accuracy compared to a much larger and slower state-of-the-art network, which is unsuitable for small aerial vehicles, as considered in this work.