Real-Time Multi-Frame Analysis of Dominant Translation

This paper describes an ultra-fast method for estimation of translational motion in video sequences. The method is based on the principle that 2D image translation results in translation of image projections. To obtain superior robustness and accuracy we developed a multi-frame extension, which robustly integrates motion information over time. Our method gives reliable results even when video images are blurred by significant and rapid motion and in the presence of independent local motions. The estimation process is very fast: on a standard desktop computer without using any hardware-or assembler-based optimization it works at a rate of 1400 fps (VGA frame size), exceeding real-time requirements and is typically 500 times faster than some prior art approaches, including 2D phase correlation. Our method also provides a very compact description of a video frame, applicable to motion analysis and other video processing tasks in low-cost DSP systems