Volker Grabe

As of April 2014, I have left the Max Planck Institute for Biological Cybernetics to join , a vibrant spin-off in Pittsburgh, PA, USA.

During my time as a member of the at the Max Planck Institute for Biological Cybernetics, I have been a Ph.D. student under the co-supervision of (IRISA/INRIA Rennes, France) and (University of Zurich, Switzerland) between October 2010 and March 2014. My doctoral advisor has been (Eberhard Karls Universität Tübingen, Germany).

Between November 2012 and March 2014, I was a visiting student with in his at the University of Zurich in Switzerland.

In the field of vertical take-off and landing vehicles, cameras became the most favored sensors for tracking and localization. However, most systems with a camera as their main sensor usually perform large portions of the computation on a powerful ground station. This clearly limits the flexibility of the system. Additionally, many presented systems lack robust backup strategies to deal with failures of the main tracking system which often relies on a map or given features.

In my work, I was addressing these limitations by relying mainly on optical flow and sensor fusion with inertial sensors. Having developed a robust fall-back behavior which allows the vehicle to hover in place if any more advanced system fails, I was then exploiting this work towards active visual guiding with the aim of autonomous avoidance of static as well as moving obstacles.

Please note that the remainder of this page discusses a project conducted at the end of 2011. For an overview of my work until I left the MPI in early 2014, please consult my publication record. In 2012, I was continuing work on visual velocity estimation. Additionally, I was contributing to sensor fusing techniques for IMU and visual information to recover a metric scale. Since spring 2013, I was mostly addressing the problem of distributed sensing, that is, a swarm of heterogeneous robots collaborates in a mapping task by each contributing with different sensors. So far, this involves cameras, RGBD sensors, IMUs, and laser range scanners.

Recently, in the field of vertical take-off and landing vehicles, cameras became the most favored sensors for tracking and localization. However, most systems with a camera as their main sensor usually perform large portions of the computation on a powerful ground station. This clearly limits the flexibility of the system. Additionally, many presented systems lack robust backup strategies to deal with failures of the main tracking system which often relies on a map or given features.

In this project, we are aiming at developing a robust system which is able to visually estimate its velocity using the optical flow obtained from a monocular camera and thus without the need to maintain a map. All computation will be done on-board the vehicle itself with only the velocity commands being transmitted to the quadrotor. Sensor fusion with an additional IMU (Inertial Measurement Unit) will be done to retrieve the scale and improve the velocity estimation.

Our work will be carried out on quadrotors manufactured from MikroKopter.de equipped with a small Intel Atom 1.6 GHz board. The final setup is shown in Figure 1. Image processing and optical flow extraction are based on the OpenCV library. To estimate the velocity, we use a customized version of the continuous four-point algorithm for planar scenes.

Figure 2 presents a comparison between the linear velocity estimated by our algorithm and the ground truth obtained from a Vicon tracking system during closed-loop control. This comparison demonstrates reliable velocity estimation with a mean error of only 0.028 m/s.

Our results show clearly that it is indeed possible to stabilize a quadrotor in closed-loop control using our algorithm on the limited hardware described. Thus, our system can be used as a robust emergency stopping behavior which does not require any knowledge of the environment (e.g. a map).