I'd suggest using JMyron ( http://webcamxtra.sourceforge.net/ ) or OpenCV ( http://opencv.willowgarage.com/wiki/ ) and take advantage of their motion/blob detection algorithms. JMyron is more suited for real time webcam input but you can feed a video image in (I think); I'm not sure how OpenCV works but I've heard good things about it. After you get the positions of various blobs, I'd say to iterate through all of the Bodies in your box2d instance and apply a force to each one, pointing towards the blob.