at first, thank for you to help me about this topic
as such topic, i want to measure distance of robot using optical flow
but i can't find optical flow on the featureless planar
so i calculate dead reckoning using encoder for such topic
but dead reckoning is incorrect
because which include many error ex) slip, structural error of the robot, and etc..
so i wil use a optical flow ( Pyramid Lucas Kanade Algorithm ) for measuring and correcting distance
i have information following
1) can calculate direction of robot using dead reckoning
2) i know height of camera, working distance and FOV(field of view)
3) i know projecting a structured light pattern onto the plane
Hi Kim,
Once you can "projecting a structured light pattern onto the plane", it may calculate relevant distance (CCD to plane) by Dk. See Fig. 1. in reference.
LSW
2016-06-17
SFM / visual odometry maybe a better way.
According to the scenario, that's why to use the light pattern because it can be used to calculate relevant distance from the relevant "move".
1. Ming-Chih, L.; Wei-Yen, W.; Chun-Yen, C. Image-based distance and area measuring systems. Sensors Journal, IEEE 2006, 6, 495-503.
2. Lu, M.C.; Hsu, C.C.; Lu, Y.Y. Image-based system for measuring objects on an oblique plane and its applications in 2-d localization. Ieee Sensors Journal 2012, 12, 2249-2261.
If the light projector moves with the robot (ie you are projecting the light from an LED mounted the robot), you cannot do it: the light pattern will move with the robot and not provide any information about the movement of the robot on the plane.
You need features on the plane.
If you fired a powerful laser at the plane and burned dark spots onto its surface (or sprayed out dots of paint), such that the features remained stationary on the plane, you could use SFM / visual odometry to estimate robot motion.
In order to use optical flow you need to have some features in the image. However, if there are no features in the visual range, looking a little bit towards the ultra violet or infrared side may yield something. Furthermore, enlarging the the image (using suitable lenses) will reveal some features at some enlargement (unless the surface is perfect - check with an optical computer mouse if it works on the surface).
In this case you can put your camera to point forward or any other direction where you can have some features. There is a very interesting work made at the university of Zurich and it's publicly available named (Semi-direct Visual Odometry):
https://github.com/uzh-rpg/rpg_svo
the original website of their projects:
http://rpg.ifi.uzh.ch/software_datasets.html
- Badges
- Science topic
- Similar topics
- Engineering
- Computer Engineering