A GPU PARALLEL APPROACH TO REAL-TIME DENSE RGB-D VISUAL ODOMETRY

Abstract

This article presents a new approach to implement real-time Dense RGB-D Visual Odometry taking advantage of the computational power of a modern GPU and the kind of parallelization it oers, exploring the quantity of data that needs to be handled. Our objective is to provide a fast and accurate method to estimate the camera motion using photometric error minimization. Assuming constant intensity, the error minimized is modeled in a nonlinear fashion requiring robust iterative methods to solve it. The algorithm developed was made having in mind now everyday hardware GPU present in PCs, phones, and embedded computers. Our results show that the algorithm is robust enough to handle displacements of a camera moving with small velocities and performing real-time odometry. Also, a linear algebra library was created alongside this work to near the gap between parallel programming concepts and linear algebra with high level procedures and data structures.