Using Drones as Reference Sensors for Neural-Networks-Based Modeling of Automotive Perception Errors

This repository holds additional figures to the mentioned IEEE IV 2020 publication, which is available at IEEEXplore: https://ieeexplore.ieee.org/document/9304615

To cite the following graphics, please cite the paper itself. It includes a link to this repository. Bibtex:

@InProceedings{Krajewski2020UsingDrones,
  author={R. {Krajewski} and M. {Hoss} and A. {Meister} and F. {Thomsen} and J. {Bock} and L. {Eckstein}},
  booktitle={2020 IEEE Intelligent Vehicles Symposium (IV)}, 
  title={Using Drones as Reference Sensors for Neural-Networks-Based Modeling of Automotive Perception Errors}, 
  year={2020},
  pages={708-715},
  doi={10.1109/IV47402.2020.9304615}}
}

Short Background

Vehicles are perceived by both a lidar-based system under test (SUT) and a UAV-based reference measurement system (ref). For both systems, the vehicle state vector consists of positions, velocities, width, length, and orientation angle. The following graphics show the errors of the lidar-based objects with respect to the reference. These errors can be

• observed in actually measured data
• predicted by the neural-networks-based Gaussian error model. The inputs to this model can be:
  • reference data that actually got measured
  • artificially generated reference data.

Please read the paper for more information.

Figures

Errors observed and predicted on measured reference data

Position x

Position y

Velocity vx

Velocity vy

Width

Length

Orientation angle

Errors predicted on artificial reference data

Position x

Position y

Velocity vx

Velocity vy

Width

Length

Orientation angle