Tracking people by clothing
Jeremy from United States  [12 posts]
8 year
Am experimenting with tracking people by clothing. This seems fairly simple if the person is wearing something with solid bright colors, for example, a red shirt. I can just take a sample of the color and then use a color filter to isolate the shirt from the rest of the background. The shirt will more than likely be the only color of its kind in the picture.
But I am wondering what to do if the clothing has patterns; for instance, a plaid flannel shirt, which may have multiple colors. Is there a way to use the Sample Color module to get a mean color from that shirt which would be unique, and could be isolated from the background? Is there a better way to acomplish this? Or is it even possible to do? Thanks.
Steven Gentner from United States  [1446 posts] 8 year

There are a couple different ways of doing this. Can I assume that the camera that you are using is seeing people in frontal/back/side view as apposed to overhead? What's the application context? I.e. are you using this for a mobile system to follow people or are you just wanting to watch a person as they move in front of a storefront ... or some other senario?

To track people from that viewpoint, its best to incorporate several factors into a probability match. I.e. if you initialize a tracker using a person standing with a colored shirt and black pants the system could look for that colored shirt next to a black color. The more context you can assume the better the tracker would work. In addition, if you can also assume people are moving in front of the camera but the camera is NOT then some background subtraction can help identify those objects that are moving with a colored shirt above black pants. Finally, if you can also assume people are standing and not bending over, the system can also assume a particular aspect ratio from when the tracker is initialized that would only track moving colored objects above a black object that is taller than wider.

So the more you know about the environment and what you will be tracking the more features you can create that will do a better job of eliminating objects that should not be tracked.

If you try the object tracker module and click on a persons middle area (belt area) that would work for a bit ... but eventually it would lose the person as it only assumes color/intensity. Adding additional features would help it keep track of people.


Jeremy from United States  [12 posts] 8 year
I don't use a camera; I'm using footage from old movies in my project, so of course it is all pretty much a frontal view. I just want to see how well I can track an actor  through a scene, and perhaps subsequent scenes, using RR. The idea is to find faces with the  face detection module first, then to sample an area of clothing beneath the face. ) Someone has done a similar project using OpenCV, posted on the following site.)  http://shervinemami.info/shirtDetection.html.

If the article of clothing is solid and bright, such as a lady's dress, I hope to use some probability matches, as you mentioned; RR looks for a skin-colored blob, directly above a specific color sampled previously. This is of course dependent on having film material with good lighting that doesn't change too much. I hadn't thought about using the middle area.

One problem is, a lot of clothing tends to be  multicolored; a dress with polka dots, or a  plaid flannel shirt. Is there a way to sample the colors of those types of garments, such that the same colors can be sampled and later detected in the following frames, where the same article of clothing appears?


This forum thread has been closed due to inactivity (more than 4 months) or number of replies (more than 50 messages). Please start a New Post and enter a new forum thread with the appropriate title.

 New Post   Forum Index