BROWSE
Detail
Conference Paper
A Photo-Realistic Synthetic Dataset for Analyzing the Effects of Moving Objects on Visual Localization Algorithms for Drones
Cited 0 time in 
Share 
Share
- Authors
- Jeonggi Yang, Soojeon Lee, Byoung-Sun Lee
- Issue Date
- 2019-10
- Citation
- International Symposium on Robotics Research (ISRR) 2019, pp.1-15
- Language
- English
- Type
- Conference Paper
- Abstract
- We compare the performance of various visual localization algorithms for drones, especially when moving objects exist. For that purpose, we construct a photo-realistic synthetic dataset using the Unreal Engine with the AirSim plug-in. The dataset acquired from two different virtual 3D spaces (i.e., a warehouse and a house), contains a virtual drone’s stereo images, IMU data and semantic segmentation images. Our dataset format is compatible with that of the EuRoC MAV dataset, thus can easily be utilized by the research community. By using the dataset, we analyze how much the existence of moving objects affects the performance of the open-source visual localization algorithms. Evaluation results show that the visual localization algorithms are hindered by the image features located on moving objects in some conditions, but the use of tightly coupled stereo camera and IMU is able to reduce the hindering effect. The proposed dataset is available on https://github.com/etri/Octopus-MAV-dataset.
- KSP Keywords
- IMU data, Image Features, Semantic segmentation, Synthetic Datasets, Tightly coupled, Unreal Engine, localization algorithm, moving objects, open source, photo-realistic, plug-in