A Factorization Approach for Enabling Structure-from-Motion/SLAM Using Integer Arithmetic

Abstract

SLAM and SfM algorithms typically involve minimization of a cost-function by non-linear least-squares methods. The matrices involved are typically very poorly conditioned, making the procedure sensitive to numerical precision effects. Ensuring accuracy therefore entails the use of high-precision floating-point data-types for representation and compute. In this paper, a square-root filtering approach to EKF-based SfM is presented and is shown to be capable of operating with lower-precision arithmetic than the EKF, while sacrificing only a little in accuracy. Specifically, we demonstrate a prototype that is capable of operating with integer arithmetic rather than floating-point - the first such implementation to the best of our knowledge. This is important given the increasing need to implement advanced vision-based capabilities on low-power embedded and mobile processors, some of which might not even support floating-point arithmetic for reasons of cost and power. Furthermore, an evaluation of the computational complexity shows that the proposed approach typically requires fewer computations than the EKF in practice, resulting in an algorithm that is both numerically more robust and computationally less intensive.