Efficient 3D triangulation in hardware for dense structure-from-motion in low-speed automotive scenarios

With the introduction of surround view cameras in modern vehicles and the possibility of calculating dense motion fields in real-time from a moving camera a detailed 3D reconstruction of the static environment is possible (structure-from-motion). Beside the necessity of a motion field between two image frames the task of triangulating those individual 2D point matches to 3D points in the world becomes non real-time on modern CPUs when to be repeated for all image points. In this work we evaluate different approaches to the 3D triangulation optimization problem in a typical structure-from-motion processing chain for an efficient implementation in hardware. An architecture for solving this problem using linear triangulation with an inhomogeneous solution to the equation system is proposed. We evaluate our implementation using FPGAs against a software-implementation with synthetic datasets and from low-speed parking area scenes for numerical accuracy and real-time capabilities. In addition the proposed fixed-point arithmetic implementation is compared against an implementation using floating-point units.