Stefan Karlsson/PersonalPage/MotionAnalysis

Jump to: navigation, search

Motion estimation by Optical Flow holds great promise for human machine interaction, gesture recognition and much more. Achieving high quality motion estimation in real-time is a topic of great interest. Current work has lead to the formulation of the interactive "squiggle". The below video clip is an implementation in matlab. It requires reasonably fast frame-rates (say 15 frames per second for good results, the video is at 30 frames per sec). The speed of computations enables it to run at over 100 frames per second with no real demands on the hardware(the thumbnail is quite bad, please click the video for it to start):

Meet the squiggle

600px and an uncostrained version (perhaps more fun to play with): 600px

The algorithm is easily parallizeable, and is stable for larger motions. Currently it runs on a pyramid of two level to somewhat deal with larger motions.

Implementation on handheld devices is a promising future application.

The speed of computations can handle over 100 frames per seconds(at resolution 128 x 128) in the current matlab implementation on a mid-range laptop (using the built in cam of the dell latitude). The frame rate does not need to be nearly as fast as that though, but 15 frames per second is recommendable.