Vision-based detection of micro unmanned aerial vehicles indoors and outdoors, no tracking yet.. Although the results are presented for a single quadrotor, the approach is generalizable. Not present in the video, but a distance estimator is also integrated to the system and tested indoors. Its giving a median error around 20 cm. Details are at http://www.mdpi.com/1424-8220/15/9/23805
You need to be a member of diydrones to add comments!
Comments
Cluster detection in it's most basic form is a very simple and elegant solution.
First you build a reference database that contains many small sections taken from the reference object at various angles. Then later when you want to find the object in a scene, you start to look for clusters of data that is similar to those reference sections. Meaning the area with the biggest concentrations of similar data is most likely to be where the object is.
The HOG (Histogram of oriented gradients) part is just a more efficient way of representing the section data that tries to compensate for varying conditions between reference and real world data.
Ok, thanks for the explanation. So it's a learned pattern then rather than something more hard-coded. Really great.
@Mustafa Thanks for the nice comments! There are lots of to do before achieving an onboard system with large field of view, high speed and accuracy to be used in control algorithms. Vision seems to be an important candidate to develop such a system. However, miniature radar technologies are also very promising such as this one: http://www.uav-alaska.com/#!research/c1lvo , http://journals.cambridge.org/action/displayAbstract?fromPage=onlin...
Results are perfect, Such a great study. When the tracking is also applied, this will be the future of UAV swarm applications.
@John Yes, that is the most probable reason.
I noticed you lose the solution every time you bank over hard. Is this just because of a limited training set?
Hi,
@Randy Color is not used as a cue in this study. Although the outdoor videos are given in color, their grayscale versions are used in detection (Training is also done with grayscale images obtained from different videos.). The method can learn different types of uavs at the same time. Testing of this scenario will be one of the future directions actually. But, if you want to use this approach to detect only your quadrotors in a swarm application, it is again very useful.
@John You are right, the detection window is encapsulating all of the quadrotor, since the method is utilizing only the shape and texture of the whole quadrotor. I should indicate that cascaded classifiers of three different methods are compared in the study: Haar-like features, Local Binary Patterns (LBP) and Histogram of Oriented Gradients (LBP). Use of LBP seems to be a better choice.
I should also note that, the results are for pure detection, no tracking is applied yet.
The paper abstract says it is using HOG clusters, so while the color helps making it more unique it should be tracking the visual shape of the copters and not color alone (as you can see by the tracking box fitting the quad outline as it moves, and not just the colored parts).
And regardless, this approach combined with cheaper next gen sensors (lidar etc) is the future for sure.
I guess that colourful pilar on the top of the copter is to make recognising it easier? I guess that's ok as long as you're attempting to track only your own drone that you control the colour of.