Description
Autonomous navigation of Unmanned Aerial Vehicle (UAV) along and Autonomous Surface Vehicle (ASV) within rivers and creeks has been a popular research area in recent years, where semantic segmentation neural networks have been implemented to recognize the navigable space. Currently, it is still difficult to release the power of deep semantic segmentation learning for ASV to make long-range navigation plans, and to swiftly and safely do obstacle avoidance while navigating in narrow, rapidly flowing and obstacle intensive creeks/rivers due to lack of aerial fluvial scene data for supervised training. To tackle this problem, and to enrich the aerial fluvial semantic segmentation training data, we collected aerial BEV (birds-eye-view) images, with multiple camera perspectives, of fluvial scenes with drone that flew above inland waterways. Images have been manually selected and semantically labeled with emphasis on extruded obstacles from river and riverbank, to form the novel dataset with 8 classes (Water, Boat, Bridge, Sky, Forest vegetation, Dry sediment, Drone itself and in-river Obstacles) and 816 high-resolution (2K and 2.7K) images in total.
Cite this work
Researchers should cite this work as follows:
- Wang, Z.; Wu, L.; Mahmoudian, N. (2022). Aerial Fluvial Image Dataset (AFID) for Semantic Segmentation. Purdue University Research Repository. doi:10.4231/B129-XD47
Tags
Notes
Initial release.