Question about Curb Extraction Results?

[AutopilotTech]

Regarding the current code's processing of KITTI data, from the results, it appears that curbs are not being segmented out. Is this a limitation of the method, or can the results be improved by modifying parameters?

[Nussknacker]

GroundGrid considers curbs as part of the ground class. If you want them marked as non-ground, you can try decreasing the "distance_factor" and "minimum_distance_factor" parameters, as well as the "minimum_point_height_threshold" and "minimum_point_height_obstacle_threshold."

The "minimum_distance_factor" parameters control how "flat" the points in a cell must be to be labeled as ground, while the height threshold parameters determine how high above the ground a point can be to still be considered ground.

However, lowering these parameters may lead to increased over-segmentation, especially in sloped or uneven areas.

[AutopilotTech]

For outdoor autonomous vehicles, it is not reasonable to treat the curb as the ground. The planned route cannot go to the curb. If the curb must be regarded as an obstacle, will the effect of ground segmentation be greatly reduced?

[Nussknacker]

Intuitively, curbs are clearly part of the ground. When using ground segmentation in the perception stack of autonomous cars, it is crucial to clearly segment pedestrians from the ground surface in point clouds, including curbs. Otherwise, tracking pedestrians crossing the road could fail.
While it is true that the planned route usually does not cross curbs, the assumption that this never happens does not hold, as evidenced by the KITTI dataset, which shows many cars parked in spaces where curbs must be crossed. This is also evident in our teaser animation, based on sequence 00 of the KITTI dataset.

If you want to implement a planner based on the output of GroundGrid, I suggest considering more information than just obstacles. GroundGrid also estimates the surrounding terrain and publishes an elevation map along with auxiliary information. The elevation map could be used for traversability estimation by calculating the slope and, in conjunction with the variance map produced by GroundGrid, potentially a road boundary estimation could be implemented.