Which methods can be used to detect hand in a complex background

I have to detect hand regions in a video without using skin color as a hint. Do you have any idea on which algorithms I can make use of to detect hand. I am searching for a simple algorithm which is quite easy to implement.

edit retag close merge delete

At what distance (range) will be the hand be from the camera?

What does "complex background" mean? Any background and lighting conditions are possible?

What about your system? What kind of resources do you have available?

Will it work real time?

( 2016-02-03 10:24:24 -0500 )edit

The hand will be quite close to the camera. By complex background i mean that the lighting conditions may vary, the background may consists of people walking around but the system should detect the hand of the person in the front of the camera. And yes, it is real time.

( 2016-02-03 10:31:49 -0500 )edit
1

Hands are deformable objects you cannot rely on rigid object detection. So this is not doable using a "simple algorithm which is quite easy to implement".

I've seen gesture recognition systems that worked with 3D sensors (such as kinect) that worked quite well in indoor environments. These sensors are good at these kind of tasks bc they provide a distance measure and ignore varying lighting conditions

They prompted the user to stand out his hand to the sensor so the system could detect it, and after that a tracker (such as Kalman Filter) would keep track of the hand.

The gesture recognition was based on machine learning algorithm, meaning, the system was fed a large number of labelled training data with different gestures, and performed quite well on recognizing them on live data.

( 2016-02-03 10:55:13 -0500 )edit
1

If you want to do this solely using RGB data, I'd start by trying to apply a people detector. After detecting a person you can make a set of assumptions to make your life easier to detect the hand.

I've never seen a robust hand detector and gesture recognition system that worked in real time with RGB data only, so I really don't think this is possible using state-of-the-art algorithms.. But i am no authority on the matter.

( 2016-02-03 11:00:23 -0500 )edit

The deformation on these hands is so harsh, that it will be very challenging. This kind of stuff is being actively researched so be prepared to come up with complete new algorithms ... So basically I am searching for a simple algorithm which is quite easy to implement. --> Is a no go!

( 2016-02-04 05:09:24 -0500 )edit