A PC-based Real-Time Stereo Vision System

This paper describes a stereo vision system that enables real-time dense depth measurements on a personal computer. The system relies on a very efficient stereo matching engine that, unlike many other approaches which use two distinct matching phases in order to detect unreliable matches, uses a single matching phase. Our matching engine allows for rejecting most unreliable matches by exploiting violations of the uniqueness constraint as well as analysing the behaviour of correlation scores. Real-time capability has been achieved deploying very efficient incremental calculation schemes aimed at avoiding redundant calculations and parallelising the computationally expensive portion of the code with Single Instruction Multiple Data (SIMD) parallel instructions, available nowadays in almost any state-of-the-art general purpose microprocessors. Experimental results on real stereo sequences and preliminary results concerning a 3D people tracking/counting application show the effectiveness of the proposed PC-based stereo vision system for real-time applications.