A few months ago a lot of people of the Flash community talked about Marinela, a wonderduf implementation of the Haar detection face done by Ohtsuka Masakazu from the spark project community. In OpenCV, there is another face and hand detection and tracking algorithm called Camshift and described here by Gary Bradski. Camshift uses color information to track object into an image sequence. Not only it can detect object position but also retrieve information about object size and orientation. Furthermore it has the advantages to be easy to implement and computationally efficient, which are really good thing for our Flash Player and my precious time ; -) .So last week I tried to implemented it and to test if it could work in Flash.

The result is better than I thought and makes me really enthusiastic for the following. Of course there is sometimes weird behaviours but it is really a beta version and a lot of optimizations should be done. I worked into RGB color space which maybe is not the best for the sensibility of light changes and I should do image pre-processing on my source image in order to enhance the detection accuracy. I still need to document myself on those points. I also do a lot of 2D loops which are really CPU-killer in Flash!

I needed a good image sequence with a lot of head moves to try it. Gary Brolsma’s video aka “numa guy” was exactly what I wanted . This video is also good for testing the sensibility of the algorithm to the noise, especially when hands and arms, which have the same color than the face, come in front of the head. I also implemented a webcam version which can work pretty well according the lighting and the background color. I notice also a bug sometimes which makes the application fails.

(PS: I also would like to thank my friend Thibault to let me use his head for this demo. Well actually I did not ask him and I do not even have the rights to use it , but I am sure he does not mind, let’s assume it is like a kind of french solidarity )

I will try to explain it in more details and to share source code later, when actually I will have time ! . Feel free to talk about it in comments

This entry was posted on Thursday, May 28th, 2009 at 10:51 am and is filed under computer vision, image processing. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.