Davide Scaramuzza and pals in the Robotics and Perception Group at University of Zurich have devised a new method called Semi-Direct Monocular Visual Odometry for autonomous gps denied micro areal vehicle navigation. Using only a single downward-looking camera in combination with an Inertial Measurement Unit the team was able to achieve robust autonomous navigation with fast vehicle state update frequency and low cpu cost. The high frame-rate tracking allows the filter to converge even in scenes of repetitive and high-frequency texture (e.g., asphalt, grass), as it is best demonstrated in the video.
A video of their work in action:
https://www.youtube.com/watch?v=2YnIMfw6bJY
The paper can be read here http://rpg.ifi.uzh.ch/docs/ICRA14_Forster.pdf
Open source code https://github.com/uzh-rpg/rpg_svo
Comments
Now the problem is to put object avoidance on the same platform.
Impressive!! Great job guys.
Like any sensor, what can disturb this one, direct light maybe?
The Parallella has great potential (I have 2 of them from the kickstarter), and the multi core design is very well suited for visual computing. But the 16 core DSP is not generic. You would most likely have to redesign the SVO algorithm in question for the hardware and better parallelism.
Hmm I wonder if the 16gpu core 2 arm core parallela board from adapteva can work here...?
hzl
@John Arne Birkeland: ok i missed that. Thanks!
They reference the ODroid-U2 at page 6 in the paper.
@John Arne Birkeland: How do you know it's 1.7 Ghz ARM?
Nice variation on more traditional SLAM. Having just quickly browsed trough the presentation, it seems that have been able to greatly optimize the computations needed by assuming that the camera will always be looking downwards, which makes much sense for copter navigation.
But before anyone starts asking for a PX4 implementation, it is worth nothing that even with the optimizations they still need a 1.7Ghz quad core ARM to get the 55fps tracking rate. In comparison the PX4 and PX4Flow both have a 168mhz single core CPU. So you will need a dedicated board for this.
Looks great! How hard can it be to make it work on anything else than Kmel robotics fc?