Abstract: It is challenging to detect moving objects from a moving camera as a motion field in the entire image can be induced by the camera motion. A fast and memory-saving detection method is proposed to resolve this problem. First, a fast and memory-saving registration scheme is used to estimate the homography transform between two neighboring frames. Then, neighboring frames is registered with the estimated transform, and frame-to-frame difference is performed to capture the motion cue. Finally, the motion cues are aggregated to construct a constantly updated motion image. After thresholding the motion image, separation of moving objects from the background is achieved. The effectiveness of the proposed method in detecting moving objects from cluttered scenes is validated via experiments on several different video sequences. In addition, this method performs better than previous techniques, while using a fraction of the computation time and a fraction of the memory as well. Specifically, with a memory usage of 825 kByte only, this method runs at 16 frames per second for a sequence with an image resolution of 480360.