The purpose of BEV (Birds-Eye-View) image semantic segmentation is two-fold: first it can endow the UAV (Unmanned Aerial Vehicle) the ability to fly en-route along the river by recognizing navigable water pixels; second it can make UAV discover the potential obstacles in river and inform it to the navigating USV (Unmanned Surface Vehicle) to do better obstacle avoidance. Combined, it can facilitate the application of cooperation and coordination of USV and UAV during fluvial autonomous navigation.
AFID images (2560x1440 and 2720x1536) were taken by waterproof SplashDrone4 during November 2021 on several segments of Wabash River and Wildcat Creek in Indiana, USA. From the original videos, 816 images were manually selected that are temporally apart and have obvious appearance distinctions. These images are semantically (pixel-wise) annotated into 8 classes by human labelers : Drone itself, River, Sky, Forest Vegetation, Dry Sediments, Boat, Bridge and Obstacle in river. The dataset contains both the multi-class masks and binary (water and non-water) masks along with original RGB images, and has both slanted view and nadir view of fluvial scenes aerially. Besides, effects like water reflections and tree shading on water surface exist in this dataset. Most importantly, in-river obstructions like sediments and broken branches are finely labeled for ASV navigation purpose.
After unzipped, "wabash_dataset" folder and "wildcat_dataset" folder contain images collected during en-route navigation of boat along each river and by both forward and downward drone camera perspectives, whereas "bridge1_dataset" to "bridge4_dataset" folders contain images collected by human pilot standing on the river bank and mainly focusing on fluvial scenes with bridges. Among bridges images, "bridge1_dataset" is of images from Wabash River, "bridge2_dataset" to "bridge4_dataset" are images from WildcatCreek. The drone is legally registered, was controlled by human beings and within human sight at all time, and the drone flight operations (safe height, speed and zone) abide by both FAA (Federal Aviation Administrations) regulations and state laws.
- Aerial Fluvial Image Dataset (AFID) for Semantic Segmentation v.1.0 - published 20 Jul 2022 in Datasets