1

I've started using OpenCV with a fair amount of success with a view to detecting different body parts (hands, arms, face, e.t.c).

The first and most obvious approach to achieve this is using Haar or LBP cascade classifiers, but I would like to know what other options are out there.

I've seen examples of skin detection using HSV images and detecting pixels within a color range. But this does not help determine which bit is which! I've also implemented an MOG2 background subtraction which can help isolate a moving object but this would only work if the desired object is the ONLY one moving.

Using convex hull and defects I've outlined hands and arms but cannot distinguish between the two. Is there a standard way of doing this or is the only reliable way using cascade classifiers?

Improve this question

1 Answer 1

Reset to default
5

You can try HoG descriptors which where developed in 2006. They work very well for faces and pedestrians. They work not that good for single body parts (still better than haar features). They were considered as state of the art until deep learning came up. (around 2012 I think)

The system described in this paper won several competetions when it was presented.

Hog is implemented in OpenCV but the DLib implementation is better.

See also this for a lot of usefull information about computer vision in general.

Improve this answer
Sign up to request clarification or add additional context in comments.

8 Comments

Thanks. At the moment I'm working through the code here which seems pretty good. It takes an average color sample of the hand and then filers based on that. Then it calculates the defects and dismisses those unlikely to belong to a hand.
If you are willing to proivde an starting example to your algorithm like in your posted example, you can also try to use the FLANN algorithm form opencv (docs.opencv.org/2.4/doc/tutorials/features2d/…) Keep in mind, that color based detection is highly depending from lightning.
FLANN looks really interesting, I'm assuming that it compares one image to another and determines the similarity? Also I was hoping that conversion to HSV or HSL would help reduce the lighting interference.
Can you detemine the angle or plane of the image using FLANN?
Yes it compares similarity. Converting to HSV is usefull, but not a complete solution. I am unsure if I understand your last question correct. Do you want to determine the angle (pose) of your tracked object?
|

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.