Kognitive Systeme

A Generic Monte Carlo Framework for Multi-Sensor Fusion and its Application to Localization

Richard Hanten, Cornelia Schulz

MuSe-SMC is a general purpose Monte Carlo framework, which allows the implementation of localization and optimization processes. While the user defines the state-space representation and the mathematical basics, the Sequential Monte Carlo method in the core of the framework works the same.
Currently, we tested the framework for 2D particle-based localization, supporting several sensor inputs. The general functionality looks as follows.

While the different sensor channels produce inputs to the filter algorithm, the arguments are bound to the model functions which are used for propagation and importance weighting. Finally all measurements prepared in a frontend are pushed to processing queues in the backend. More detailed information about features of the current state of the approach can be found in the related paper [1].

2D Localization

MuSe-MCL2D is an instance of MuSe-SMC in use. Currently, we support 2D laser range finders with standard map representations and sensor models. We also recently developed a sensor model which works fine with stereo camera data based on NDT maps. Additionally, we support sensor models from former papers [2].

The picture above shows the MuSe-MCL2D multi sensor particle filter localization in action, supported by two 2D laser range finders (light green points representing the front laser and the orange ones the rear laser) and a stereo camera (grey). While the laser range finder data is used with a 2D occupancy grid map, the stereo camera data is compared to a 3D NDT map (bluish boxes represent means of joint distributions). The red arrows in the image are the particles and represent possible robot poses.

Getting the Software

As for all our software, MuSe-SMC and MuSe-MCL2D are available via Github.

References

[1] Richard Hanten, Cornelia Schulz, Adrian Zwiener, and Andreas Zell. MuSe: Multi-Sensor Integration Strategies Applied to Sequential Monte Carlo Methods. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. (Under review).
[2] Richard Hanten, Sebastian Buck, Sebastian Otte, and Andreas Zell. Vector-amcl: Vector based adaptive monte carlo localization for indoor maps. In Intelligent Autonomous Systems (IAS), The 14th International Conference on, Shanghai, CN, July 2016. [ details ]