Abstract | ||
---|---|---|
In autonomous vehicle, LiDAR is one of the most importent sensor for measurement range. In particular, LiDAR is widely used in mobile mapping systems (MMS) for building high-definition (HD) map. Not only range but also intensity are useful data for extracting road marking and surface data from sensor. However, LiDAR intensity is affected by surface reflecting characteristic as well as range, incidence angle, atmospheric transmittance and others. Therefore, LiDAR intensity model is difficult to reproduce in simulation environments. In this paper, we proposed realistic intensity rendering method with simulator by generative adversarial network (GAN) trained with unannotated real data. We develop LiDAR projection preprocessing method for operating convolution networks. Our approach have advantage to preserve sensor characteristic and to reduce projection scale loss. We modified CycleGAN architecture to render intensity with unpaired real and synthetic data. Our rendered results show possibility of image-based approach for LiDAR and potential for generating LiDAR dataset by using simulator. Especially, we expect to utilize for military vehicle that has difficulty to reproduce the environments. |
Year | DOI | Venue |
---|---|---|
2021 | 10.1109/BigComp51126.2021.00062 | 2021 IEEE International Conference on Big Data and Smart Computing (BigComp) |
Keywords | DocType | ISSN |
Autonomous Vehicle,LiDAR,Generative Adversarial Network,Data Augmentation,Simulator | Conference | 2375-933X |
ISBN | Citations | PageRank |
978-1-7281-8925-3 | 0 | 0.34 |
References | Authors | |
0 | 2 |
Name | Order | Citations | PageRank |
---|---|---|---|
Seung-chan Mok | 1 | 0 | 0.34 |
Gon-woo Kim | 2 | 0 | 0.34 |