Authors: Thomas M. Howard, Arin Morfopoulos, Jack Morrison, Yoshiaki Kuwata, Carlos Villalpando, Larry Matthies, and Michael McHenry
Year: 2012
- MER = Mars Exploration Rover ie Opportunity and Spirit
- Perception algorithms on FPGA instead of CPU => low power
- Grid-based Estimation of Surface Traversability Applied to Local Terrain (GESTALT)
Parallelization of tasks for FPGA implementation:
- Image aquisition: capturing, downsampling, transmission and storage
- Stereo: image rectification, filtering, disparity
- Visual Odometry: identification and tracking of features
- Traversability: heatmap for path planning
- Path planning
Describes acceleration of stereo and visual odometry
- Stereo: use a rectification table and return a complete disparity map (ie range info for each pixel)
- 4 steps of VO: features detection, feature matching with SAD (cache loading of features for fast matching), RANSAC to get a large set of inliers and reprojection error minimization
Acceleration of GESTALT
- Input: robot pose (VO), point cloud (stereo)
- Convert the map in a square grid, uses statistics of pc in the grid cells (mean, variance) to get slope and roughness and deduce traversability
- Select the best path to follow after generating a set of candidates
IDEA: GESTALT on a global map is more efficient at each step
Pas tout compris à la partie hardware, mais donne un bon exemple de la marche à suivre