Accelerated Stereo Vision Using Nvidia Jetson and Intel AVX

Abstract

Stereo vision is a low cost and passive mechanism to perceive the environment for robotic applications. The huge compute requirements of stereo vision algorithms have been a major challenge for their usage in real world applications on small robots. Standard stereo depth estimation algorithms Sum of Absolute Differences (SAD), census transform and an advanced algorithm Semi-Global Matching (SGM) are discussed in this work. This paper presents novel real time implementation of these three stereo vision algorithms on two different compute platforms i) Intel AVX (Advanced Vector Extension) and ii) Nvidia Jetson GPU (Graphical Processing Unit). The Intel CPU implementation of stereo algorithms is optimized by using OpenMP (Open Multi-Processing) for multi-threading, AVX registers for vectorization and several other novel ideas for real time processing. Nvidia Jetson implementation is efficiently designed for maximum speed-up on a low end GPU such as Jetson TK1. Post processing steps such as local extrema detection, left-right consistency and median filter are used to improve the final disparity image. We have achieved speedup of the order of 30x when compared with naïve CPU implementation.