Good free blob-tracking algorithms?
January 2, 2012 11:04 AM Subscribe
Are there any good open-source blob tracking implementations out there other than OpenCV? Especially anything that's optimized for road traffic.
I've tried the various tracking algorithms in the legacy vidsurv package of openCV, and some of them work moderately well, but they're poorly documented and I find the code hard to understand, so I can't modify it. For my application - following cars in traffic - I feel like I could improve the tracker's performance based on the relative predictability of the cars' movements, but not if I can't understand the algorithms well enough to adapt them.
I've seen cvBlob and cvBlobsLib, but they don't really tackle tracking. There are some simple and clear implementations written for OpenFrameworks and Processing but they don't really work very well.
I'd also welcome general advice on the car-tracking problem. This is for an art project - I don't need industrial-strength solutions.
I've tried the various tracking algorithms in the legacy vidsurv package of openCV, and some of them work moderately well, but they're poorly documented and I find the code hard to understand, so I can't modify it. For my application - following cars in traffic - I feel like I could improve the tracker's performance based on the relative predictability of the cars' movements, but not if I can't understand the algorithms well enough to adapt them.
I've seen cvBlob and cvBlobsLib, but they don't really tackle tracking. There are some simple and clear implementations written for OpenFrameworks and Processing but they don't really work very well.
I'd also welcome general advice on the car-tracking problem. This is for an art project - I don't need industrial-strength solutions.
Response by poster: oceanjesse, are you referring to the section "Finding Regional Orientation or Normal Optical Flow" or something else? (The page numbering in the PDF is a bit funny.) Motion detection is of course generally relevant to the problem, but I can't quite see what's specifically useful there.
posted by moonmilk at 2:17 PM on January 2, 2012
posted by moonmilk at 2:17 PM on January 2, 2012
Hi,
Don't know if this is actually going to be helpful but worth a shot.
I was futzing around with processing and some traffic footage a while ago (don't know where the code is sorry) but if you want to trigger things based on passing cars, use night time footage and you can really easily set things up to detect groups of pixels bigger than and brighter than y.
There are tonne of processing blob trackers, there are three listed on the processing libraries site. Object tracking is a big thing in that community, yes a lot of the work is around building interfaces for installations but not all of it.
Most of the libraries I played with have threshold variables for object size and such, have you tried playing with those - they make a big difference in the output that you get. I did get one of them to easily draw outlines of people walking around a room and that was five years ago, I really think they must be better now.
Sorry if this wasn't helpful you just seem to be in danger of trying to reinvent the wheel.
posted by metsauce at 3:59 PM on January 2, 2012
Don't know if this is actually going to be helpful but worth a shot.
I was futzing around with processing and some traffic footage a while ago (don't know where the code is sorry) but if you want to trigger things based on passing cars, use night time footage and you can really easily set things up to detect groups of pixels bigger than and brighter than y.
There are tonne of processing blob trackers, there are three listed on the processing libraries site. Object tracking is a big thing in that community, yes a lot of the work is around building interfaces for installations but not all of it.
Most of the libraries I played with have threshold variables for object size and such, have you tried playing with those - they make a big difference in the output that you get. I did get one of them to easily draw outlines of people walking around a room and that was five years ago, I really think they must be better now.
Sorry if this wasn't helpful you just seem to be in danger of trying to reinvent the wheel.
posted by metsauce at 3:59 PM on January 2, 2012
Yeah, that's what I meant. I suggested capturing flow gradients because you're trying to track motion. Would that not work? Why exactly not?
posted by oceanjesse at 11:33 AM on January 21, 2012
posted by oceanjesse at 11:33 AM on January 21, 2012
This thread is closed to new comments.
posted by oceanjesse at 11:14 AM on January 2, 2012