Waviz Background Models
The main contribution of this work is an algorithm that explicitly harnesses the scene dynamics to improve segmentation. Our algorithm detects new objects based solely on the dynamics of the pixels in a scene, rather than their appearance. Thus we can distinguish a swaying tree from a moving vehicle.
Background & Objective: Background subtraction is the most common approach for discriminating a moving object in a relatively static scene. Basically, a reference model (background) for the stationary part of the scene is estimated and the current image is compared with the reference to determine the changed regions (foreground) in the image.
A major shortcoming of all the above background methods is that they neglect the temporal correlation among the previous values of a pixel. This prevents them detecting a structured or periodic change, which is often the case since real-world physics induces near-periodic phenomenon in the environment: the motion of plants driven by wind, the action of waves on a beach, and the appearance of rotating objects.
Technical Discussion: We directly estimate models of cyclo-stationary processes to explain the observed dynamics of the scene and then comparing new observations against those models.
We generate a representation of the background using the frequency decompositions of the pixel's history. For a given frame, we compute the frequency transform coefficients (either FFT or DCT) and compare them to the background coefficients to obtain a distance map for the frame. Then, we fuse the distance maps in the same temporal window of the transform coefficients to improve the robustness against the noise and to remove the trail artifacts. Finally, we apply a threshold to the distance maps to determine the foreground pixels.
Publications:
Technology Area: Computer Vision
Modification Date: November 18, 2008