Regularized Optical Flow for Detecting Moving Foreground from Videos Taken by Steady-Moving Cameras

Optical flow (OF) is widely used for detecting the foreground from videos. However, conventional approaches often require complex post-processing steps to achieve satisfactory detection results, leading to unsatisfactory generalization. To address this issue, this paper presents a regularization-based algorithm that simultaneously detects moving foreground and moving background, and estimates their OFs according to their specific spatial-temporal characteristics. This simultaneous approach enables foreground detection from only two consecutive video frames captured from the steady-moving camera, creating the foundation for potential online processing capability. The advantages of the proposed algorithm are demonstrated through a series of real-world case studies, highlighting its effectiveness for foreground detection and OF estimation tasks.