Acceleration of a Dense monocular Localization System using FPGAs

EMANUEL TRABES; GAIA AMOROS, JEREMIAS; SUTTER, GUSTAVO

San Luis

Conferencia; XI Southern Programmable Logic Conference; 2023

Universidad Nacional de San Luis

In this work, we accelerate a dense monocular pose estimation system by leveraging the computation capabilities of FPGAs. In a dense monocular setting, pose estimation is usually defined as an minimization problem, where the function to minimize is defined as the normed difference between the observed image and a corresponding image constructed using the estimated pose and a model of the scene. This minimization problem is solved by using a gradient descent method, for example the levenberg-marquardt algorithm. In a dense setting, this requires to compute for every level of detail, for every iteration and for every pixel, the gradient and the hessian of the model with respect to the unknown pose. Also, high frame rates are required in settings where fast and complex camera movements are present, for example for UAV navigation systems. As as consequence, this problem is computationally complex. We accelerate the pose estimation by using the vitis high level synthesis tool, which allows to describe the FPGAs functionality by using C/C++ programming languages. We apply several optimizations and implement the system using a Zynq Ultrascale MPSoC. We compare our results to a pure CPU solution.