Clever Geek Handbook

Optical flow in real time

I use optical flow as a real-time obstacle avoidance and detection system for the visually impaired. I am developing an application in C # and am using Emgu Cv for image processing. I am using Lucas and Kanade's method and I am quite happy with the speed of the algorithm. I use monocular vision, so it is difficult for me to accurately calculate the depth of each tracked feature and alert the user accordingly. I am planning to use an ultrasonic sensor to help with obstacle detection due to the difficulty of calculating depth with a monocular camera. Any suggestions on how I can get an accurate depth estimate using just the camera?

+3


source to share


4 answers


You might want to check this document: Reliable Visual Odometer and Precipitation Detection System Using Consumer Grade Monocrystalline Vision . They use a nice trick to detect obstacles like holes in the field of view.



+2


source


I hate to give such a general answer, but it is best to start with a standard text on structure-from-movement to get an overview of the methods. A good recent book by Richard Seliskey is available on the Internet (chapter 7) and its links. After that, for your application, you can look at recent work at SLAM - Oxford's Active Vision group has excellent work and Andrew Davison's group .



+2


source


comment on RobAu's answer below, "structure from motion" may give better search results than "3d from video"

+1


source


Depth from one drift will only work when the camera is moving. You can learn some 3d approaches from videos . This is a very difficult problem, especially when objects move in the camera's field of view.

0


source






More articles:


All Articles