 |
Using Blob Tracking module for image match?
from United States [214 posts] |
15 year
|
Hello,
I'm not sure if this is possible but it seems it might be. Suppose a cat moves in front of a stationary camera and then stops. The new Blob Tracking module can be used to track the cat based on the cat's blob characteristics after suitable filtering of the image while the cat is moving. My question is: when the cat has stopped, is there a way to use the blob statistics captured while the cat was moving to now locate the cat even when it is *not* moving? In other words, is there a way to turn the Blob Tracking statistics gathered while tracking into an image template? Sort of like combining the Blob Tracking module with the Train function in the Image Match module?
There seems to be a related option in the Blob Tracking documentation that reads "8. Absent After Frames" but that option does not appear in the version of the module in RR 2.2.9. Also, I have in mind a template that lasts longer than a few frames--something that could even be saved to disk and used as an Image Match template at a later time.
Thanks!
patrick
|
|
|
|
Anonymous |
15 year
|
Patrick,
Good thoughts ... some of which we were thinking too. The issue is a bit of a chicken and egg problem. The Blob tracking assumes that you have identified the blob based on some known qualities and just joins these blobs over time. The stats it gathers are much less than what the Blob Filter module does ... and the reason is that once a blob is defined it doesn't take too much to track it. BUT it doesn't meant that the stats used to track would be good ones to *find* the object in a static image, they are just too basic to guarantee a singular and correct match.
Probably the best stat that might be used for this is the histogram colors which we have experimented around in other non-released modules in order to track objects (and save their signatures to disk for later use). The histogram is a nice and quick way to do this but really requires the object to be somewhat distinct in color (i.e. more red than the rest of the image). So we are looking into improving this and finding the right balance of accuracy and speed before releasing that module. This is not unlike the Mean Shift tracking that you may have heard off ... it does get interesting to see how many disparate algorithms can actually be very similar in basic principals but used in very different ways.
A generic object tracking module is in the works too but would not work for a black cat ... it might work for a textured cat but unless there is some inner texture it would probably not track well. Probably a module that specializes in snakes (see google for more about snakes in image processing) might work for a black cat situation ...
I think the "Absent After Frames" change slipped in without an official comment since it was just renamed from another obscure field in that interface (either the trail or sample setting).
STeven.
|
|
|
|
from United States [214 posts] |
15 year
|
Hi STeven,
Great to hear that you guys are already working on some of these things. The reason I was thinking along these lines is this: every time I play with the Movement Module, I am struck by how a moving object literally "jumps out" of the image. So if I watch my cat (orange with white stripes) through the Movement Module as it moves across an otherwise stationary room, the figure of the cat is nicely carved out of the black background in RoboRealm's view screen. Combine this with the Erode and Close modules, then perform an And operation with the Source image using the Math module, and you get a nice full color image of the cat completely isolated from the background.
So like you say in your message, suppose we could now get a histogram or texture signature of this isolated part of the image (the cat), then use it as a template across the image to look for the cat even when it is not moving.
Sounds like you guys are already well on your way to figuring it out. If you need a beta tester for some of the modules, I'd be happy to oblige!
--patrick
|
|
This Thread
