A VLSI architecture for VGA 30 fps video segmentation with affine motion model estimation

This paper proposes a VLSI architecture for VGA 30 fps video segmentation with affine motion model estimation. The adopted algorithm is formulated as a contextual statistical labeling problem exploiting multiscale Markov random field (MRF) models. The algorithm optimization for VLSI implementation is characterized by image division method, ICM labeling limited to region boundary, and omission of motion models estimation for new regions. The optimization reduces the computational costs by 82%, the amount of memory by 95%, and the amount of data traffic by 99% without accuracy degradation. The VLSI architecture is characterized by pipeline processing of the divided images, concurrent motion models estimation for multiple regions, and a common processing element of update and detection labeling. The architecture enables VGA 30 fps video segmentation with 167 MHz frequency. The estimated core area using 0.18 ¿m technology is 30 mm². This processor is applicable to the video recognition applications such as vehicle safety, robot, and surveillance systems under the restriction of energy consumption.