A Real-Time Background Subtraction Algorithm for Freely Moving Cameras

To extract moving foreground from a video captured by a moving camera, dense optical flow or point trajectories are often introduced to handle the camera motion, but they make the moving foreground extraction very slow. To solve this problem, a simple and effective real-time algorithm for moving camera background subtraction is proposed. Firstly, a superpixel-based region growing algorithm is proposed to preprocess the input image frames. Then, the camera motion is estimated in a block-based fashion with a background feature-point filtering method based on the relative flow. Finally, the background subtraction result is obtained via a verification process based on the accordance of the optical flow and the camera motion. Experimental results show that the proposed algorithm can process a 640×480 video in real-time. In addition, the foreground detection accuracy of the proposed algorithm outperforms other real-time moving foreground extraction methods.