Automation Technology
Automation Technology
Automation Technology 

Credible Radar Ego-Motion Estimation (creme)

Robust and credible motion estimation is fundamental for all mobile autonomous systems. If we think about motion estimation in this context, various algorithms are available to solve this problem with a variety of sensors. However, an often unstudied use case of such algorithms is the integration of their estimates within an overlaying sensor fusion, where it can be thought of as a virtual odometry sensor -- often referred to as loosely-coupled sensor integration. For this, the estimator's performance regarding its mean error is not the most crucial part, but its credibility. With credibility, we refer to the estimator's ability to evaluate its own estimation error. For the Gaussian case, the returned covariance matrix has to fit to the error distribution of the estimator with regard to the ground truth.

Automotive radar sensors are advantageous in challenging environmental conditions and are widely used for mobile autonomous systems. While these sensors often return a number of targets with corresponding, individual measurement noise, most motion estimation algorithms do not return a correct covariance estimate. To allow an efficient optimization, they use approximate assumptions, which may result in a good estimate of the mean but breaks the propagation of uncertainty. In earlier work, we introduced mixture models for least squares optimization. We leverage this to optimize the motion estimation problem in a feasible time without violating basic probabilistic assumptions. Further details can be found within our publication (see below).

In the following, we provide supplementary material and the source code of our implementation.

Open Source Implementation

The relevant source code is distributed within two projects. In libMix4SAM, the main functionality is implemented and creme contains the relevant code to replicate the publication's experiments.

libmix4sam
An extended version of the libmix4sam (see also the project's landing page) including the automotive radar factor implementations.

creme
Source for creating and running the evaluation examples as well as the Sum-Approximation implementation used within the paper.

Supplementary Material

Supplementary Material is provided as an additional document. It includes extended information on some aspects of the corresponding publication, therefore it is not self-contained.

For redoing the evaluation of the third experiment, based on data of our mobile robot, this dataset file is needed.

Publication