Design space exploration of a stereo vision system using high-level synthesis

Abstract

Stereoscopic vision is an essential building block of modern assisted driving and surveillance applications. Semi-Global Matching (SGM) is a very efficient approach, which outperforms most local algorithms and can deliver real-time performance if properly implemented in hardware. In this paper we describe the design space exploration of the SGM algorithm for automotive applications. The paper also highlights the methodology that we used for the transformation of the high-level code from a reference software implementation, which was unsuitable as a starting point for high-level synthesis, to the hardware implementation. Stream-based processing of the SGM algorithm, despite its complex data dependencies, is achieved by focusing on the inner most loops of the algorithms. Changing the choices of the loop implementation and type of the targeted memory implementation yield different RTL code with a broad range of area vs performance trade-offs.