Abstract

Enable GPU optimizations to achieve real time SLAM on the Jetson TX1 embedded computer.

SLAM = Simultaneous Localization And Mapping

TX1 devboard:

Main hotspots

Corner detection on images

Feature extraction on images

What we've done

Reimplement "Fast corner detection" and "ORB feature extraction" with CUDA

Pipeline GPU kernels as much as possible

Overlap unparallelizable CPU work and GPU work as much as possible

Results

Platforms:

PC (Xeon E3 1231 / GTX 760 Kepler)

TX1 devboard (ARM A57 / 256 core Maxwell)

Mean tracking time per frame

Mean and peak fps

Speedups

End