VLSI implementation of belief-propagation-based stereo matching with linear-model message update

This paper presents both parallel and sequential implementations of linear models for the computation of message update, a critical operation in belief propagation (BP)-based stereo matching that computes the depth information from two images captured at different positions. An improved parallel implementation of the message update is presented that can execute the forward and backward pass concurrently. We also present two versions of sequential designs to reduce area cost of message-update computation. Finally, a complete BP-based stereo matching processor is designed.