| Background subtraction is
a common practice in surveillance applications. However, maintaining a
background while the camera is moving is much more involved than that in
a static camera case. One has to warp the previous background to the current
view, detect changes, and update the current background.
I studied the case of a moving camera mounted on a pan-tilt head (a mini-robotic arm). The displacement of the camera center caused by the pan-tilt head is small enough and the parallax it introduced can be safely ignored. That way the warping between the two successive views can be modeled by planar homography. Or in this special case, the warping can be parameterized by 3 parameters, current pan, delta pan and delta tilt. Though the pan/tilt angles can be read from the pan/tilt unit, they are not accurate enough for warping purpose, especially when the camera is on the fly. In this case the exact angles must be computed from the image. I used Sum of Squared Difference (SSD) methods, an iterative gradient descending algorithm to register two successive images. The derivation of the Jacobi is a bit complicated. After two successive frames are registered, the moving target can be detected and the background can be updated. In the following example the first two images are two successive images in a video stream. Detected changes are shown in the third image. The camera is assumed calibrated.    Also please watch the following
videos to see the algorithm in action. The frame rate was 10-15fps on a
P2 200 machine.
|
