|
Stereo (stereoscopic) Vision object detection
Brenden Geary from United States [5 posts] |
10 year
|
I am fairly new to Roborealm so excuse me if I sound stupid. I have two webcams 2.5" (roughly the distance between my eyes) apart from each other. I can use the 3D_Viewer to create a cool viewing of the 3D image but how can I use it for:
1: Distance Mapping (Depth Mapping)
2: More accurate object detection
I am looking to be able to have RoboRealm or some other software learn all faces of an object (in a constant lighting environment) and be able to detect it from all sides and across the room. The cameras output full 720p HD video so resolution will not be an issue I just have no idea where to start. Thanks to all that can help.
Sincerly,
Brenden Geary
KJ6HVP
www.brendengeary.com
|
|
|
|
Steven Gentner from United States [1446 posts] |
10 year
|
Brenden,
There are a couple of possibilities. You can certainly try using the Stereo module to get a very crude depth map which may be able to help to eliminate background objects. You can also try using the Kinect or Xtion which will provide a better depth map. Again, this would be used to isolate just the object you want to recognize (assuming its the closest thing in the room).
While 720 HD sounds like a lot, in fact its not. Objects even at that resolution "across the room" can easily have only a few pixels in dimension which will not be enough to recognize the object .. but its a good place to start, just don't be surprised if object recognition stops working only a couple feet away from the camera.
The solution to long distance (i.e. room sized) object recognition is to mimic what we do. When we look for an object we look for global properties (like color, shape, location) and when we think we see it we walk closer to investigate. With a mobile system you can do the same, i.e. recognize global features (i.e. train the object recognition of the object from a distance too) and approach to verify that it in fact the object. If you get a false positive (i.e. you think its the object but its not) then approaching it should eventually fail.
Cool project!
STeven.
|
|
|
|
Brenden Geary from United States [5 posts] |
10 year
|
I kind of miss worded my question when I said more accurate object detection. What I meant is being able to detect an object no mater it's orientation or what side you are viewing it from. I have looked into the kinect and it does not work well in close distances (under 2.5ft) which for my first prototype will be the only environment I will be working with. (Looking down at something on a desk etc). How would I be able to produce a depth map with roborealm or opencv?
Thanks,
Brenden
|
|
|
|
Steven Gentner from United States [1446 posts] |
10 year
|
Brenden,
What you'd need to do is to take images from all sides of the object as part of the recognition templates. In this way regardless of the angle/side that the object appears in you will have that view in the database which can be matched to. I'd try that before experimenting with anything stereo based.
Have you tried at least one side with the Object Recognition or AVM modules?
STeven.
|
|
|
|
Brenden Geary from United States [5 posts] |
10 year
|
Ya I have successfully detected objects on only one side. I can create a database of images for the software to reference too but how do I tell it to reference to that database?
Brenden
|
|
|
|
Steven Gentner from United States [1446 posts] |
10 year
|
Which module are you using?
The AVM requires each side to be trained using its interface.
The OR module just requires a folder to be specified that contains all the image files. Its the same way one image would be done.
STeven.
|
|
|
|
Brenden Geary from United States [5 posts] |
10 year
|
|
|
This Thread