Real-time motion classification of HD video sequences on embedded systems

Fast local and global motion estimation (ME) is crucial for a wide variety of different video processing applications, but poses significant requirements on CPU resources. Interestingly, application scenarios shift more and more from traditional PC processing to embedded devices, mostly driven by home automation, machine vision and autonomous navigation. In this paper, we demonstrate how to use DSP capabilities on embedded ARM-based OMAP Systems-On-Chip (SoCs) to provide motion analysis up to HD at speeds of 15-30 fps. We show how to acquire reliable classification of motion types like zoom, rotation or translation, and how to calculate and decompose the so-called homography matrix for further refinement. Our focus lies on navigation sequences from automotive test scenarios, but we also demonstrate the accuracy of the approach on natural video sequences.