Graphics

3-Jul-09 3:34

Hi,all.

Currently I’ve been developing a Mosaic system of Two Calibrated Cameras using VS2008 and OpenCV , but my algrithm works badly

Here is my program routine, Any advice or help would be greatly appreciated!

1. I calibrated two cameras watching a same area, and undistort the input images,

2. then I use cvFindChessBoardCornerGuesses and cvFindCornerSubPix to get corresponding points of two undistorted input images,

3. and I process the corresponding points using cvFindHomography to get the homography of the two cameras

4. finally use cvWarpPerspective to warp the right camera image plane to the left camera image plane

But the warped image is extremely different from the ideal one

What's wrong with it?
Best Wishes!

Tim Craig4-Jul-09 8:41

4-Jul-09 8:41

How far apart are your cameras? And when you say that when the right image is warped to the left image plane is far from "ideal", do you know what the ideal image should be from theory or what you just expect it should be?

Have you tried the OpenCV group on Yahoo Groups? You might be more likely to get a good answer there than here.

You measure democracy by the freedom it gives its dissidents, not the freedom it gives its assimilated conformists.

I'm a proud denizen of the Real Soapbox[^]
ACCEPT NO SUBSTITUTES!!!

Bob Yiu4-Jul-09 22:58

4-Jul-09 22:58

Hey,Tim,Thanks for your reply,

My right camera is only 15 degree right rotation from the left camera.
I know the ideal warp result

I've post a message on OpenCV group,
http://tech.groups.yahoo.com/group/OpenCV/message/64370
welcome to join discussion

Do you know how to get the homography of two cameras by the camera matrix or extrinsic matrix?

Thanks!

Tim Craig5-Jul-09 21:12

5-Jul-09 21:12

Do you have the book "Learning OpenCV"[^]? I noticed a section of homography in it. In my exploration of OpenCV I haven't gotten quite that far. I've been spending a lot of time getting a framework in place so I can quickly and easily write test applications. That and fighting the lack of documentation and weirdness in OpenCV.

You measure democracy by the freedom it gives its dissidents, not the freedom it gives its assimilated conformists.

I'm a proud denizen of the Real Soapbox[^]
ACCEPT NO SUBSTITUTES!!!

Bob Yiu6-Jul-09 23:43

6-Jul-09 23:43

I've got the book "learning Opencv",but it doesn't solved my problem

So how is it going of your framework?

Tim Craig7-Jul-09 17:59

7-Jul-09 17:59

The one thing I got out of the homography section was that he suggests using at least 10 images when using the chessboard technique to get a good matrix. That's more than I expected.

My framework has been one step forward then two steps back. Sniff | :^)

I had to get a new computer last year and to get it quickly, I took one with Vista preinstalled. OpenCV and Vista don't get along well. I've finally started gaining traction. After a false start, I settled on wxWidgets for the GUI and have a lot of the basics of OpenCV wrapped in C++ to hide some of the ugliness (like the relationship between CvMatrix and IplImage Dead | X|

). Right now I'm working through object detection and obstacle avoidance for mobile robots.

You measure democracy by the freedom it gives its dissidents, not the freedom it gives its assimilated conformists.

Bob Yiu7-Jul-09 20:14

7-Jul-09 20:14

object detection and obstacle avoidance for mobile robots?
That's really an interesting thing.What developing lib are you using? OpenCV or other robot vision lib in sorceForge.net

Tim Craig7-Jul-09 20:54

7-Jul-09 20:54

Right now I'm using OpenCV, particularly for acquiring the image. There's a lot there for testing ideas but I sometimes wonder if it was really with it. I suspect that for any production system I'll want to recode the algorithms to remove the generality that OpenCV brings and do specialized versions to work with just the format of the camera I'll be using. I've looked at a few other libraries but so far OpenCV has the best support and that's minimal. I do have a friend from Homebrew Robotics here who's interning at Willow Garage so I could get access to Gary Bradski if I really needed to. The last few days I've been looking at what it will take to visually detect the target cones for a RoboMagellan robot. How are you planning to use your stereo rig? Another friend has a two wheel balancing robot that has a stereo vision setup running into a Beagleboard.

You measure democracy by the freedom it gives its dissidents, not the freedom it gives its assimilated conformists.

Bob Yiu8-Jul-09 16:27

8-Jul-09 16:27

My two cameras are relatively fixed,so I want to program a reliable algorithm based on some stable thing just like camera parameter.

In the past,I've done mosaic algorithm based on SIFT,SURF,KLT,Harri Corner,But the mosaic result of these kind of algorithm based of feature points matching is very unstabile,sometimes good,sometimes bad

I've seen some projects of robot stereo vision before,They are mostly based on feature points matching.

Tim Craig8-Jul-09 18:40

8-Jul-09 18:40

Getting something robust enough to reliably work over a wide range of situations seems to be one of the big stumbling blocks in machine vision. Laser scanners seem to be more widely successful for mobile robots today since they provide easy range information but we're talking $$$. While a number of the DARPA Urban Challenge teams investigated vision systems, I don't think any actually used them for navigation. Willow Garage is using vision but they still have laser scanners on their robot.

You measure democracy by the freedom it gives its dissidents, not the freedom it gives its assimilated conformists.

Bob Yiu11-Jul-09 20:37

11-Jul-09 20:37

Hey Tim,
Today I got another project from my boss;
Car Safety System based on cameras
That means fix some cameras(two or more) on the car, protecting the car from some objects getting too close.
It reminded me of you.I think this project is similar with your project of robot avoiding obstacle.
But my hardware system is DM6467 or DM6446 from TI(Texas Instruments)
Have you developed the software system of robot avoiding obstacle?

Tim Craig12-Jul-09 9:03

12-Jul-09 9:03

That sounds similar to the OMAP35x they use in the Beagle Board except the OMAP has an ARM8. I haven't gotten that far along yet. I have a level with a laser line generator I want to try to detect with the camera and use the parallax shift to detect objects and get a range estimate similar to this article[^] although I envisioned the laser and camera positions reversed so the line would always be in view on the floor. Wouldn't work too well outside probably. I've been able to detect the spot from a laser pointer fairly reliably with a webcam but haven't gotten around to trying the line yet. Might have to spring for an optical bandpass filter. I'll try some red plastic first. Laugh | :laugh:

Are you planning on trying to extract depth with a single image or are you going to use your stereo rig?

You measure democracy by the freedom it gives its dissidents, not the freedom it gives its assimilated conformists.

Bob Yiu2-Aug-09 22:23