Background Image Estimation Using Minimum Cost Labeling

The patent changes the parameter threshold from 50% (median requirement) to a configurable minimum visibility threshold, allowing the background to be estimated even when visible for less than half the time. This parameter change enables the system to adapt to scenarios where background visibility is limited, thereby improving versatility while maintaining implementation simplicity.

The patent extracts and removes the dependency on depth information and multiple input sequences from the background estimation process. By formulating the problem as a labeling task that works with a single image sequence and pixel intensity values, it eliminates the need for complex depth computation and multiple sequence inputs, thereby reducing device complexity while maintaining adequate background separation accuracy.

Instead of using optical flow to estimate background (which creates circular dependency), the patent inverts the approach by using a labeling-based method that directly estimates background without requiring optical flow computations. This inversion eliminates the circular dependency and process complexity while still achieving accurate background estimation that can subsequently improve optical flow computations.

The patent segments the background estimation problem into independent pixel-level labeling tasks, where each pixel is labeled with the frame number corresponding to its background appearance. This segmentation allows the system to handle camera motion and brightness changes independently at each pixel location, reducing overall process complexity while maintaining reliable background reconstruction accuracy.