Seeing Through the Grass:
Semantic Pointcloud Filter for Support Surface Learning
Anqiao Li, Chenyu Yang, Jonas Frey, Joonho Lee, Cesar Cadena, Marco Hutter
Published on Robotics and Automation Letters RA-L 2023
Embedded Files
@ARTICLE{10265206,
author={Li, Anqiao and Yang, Chenyu and Frey, Jonas and Lee, Joonho and Cadena, Cesar and Hutter, Marco},
journal={IEEE Robotics and Automation Letters},
title={Seeing Through the Grass: Semantic Pointcloud Filter for Support Surface Learning},
year={2023},
volume={8},
number={11},
pages={7687-7694},
doi={10.1109/LRA.2023.3320016}}
Preview Experiments
Preview Experiments
We recorded six 10-20 minutes trajectories from Perugia, classified into three environments: Grassland, Hillside, and Forest. Each environment contains two trajectories. The training set and testing set each contains three trajectories, one from each of these three environments. Additionally, a trajectory was collected from Hönggerberg, Switzerland, only for evaluating downstream applications.
Comparision of the absolute error of the depth estimation with respect to the distance in the different environments in the testing set. Notably, in all environments, within any distance bins, the filtered pointcloud outperforms the raw pointcloud.
Downstream Applications
Downstream Applications
Comparison of elevation mapping generated from different methods in different environments. On the leftmost side are the onboard camera images. For each method, we provide the generated elevation map and the error map (RMSE) from a top-down perspective with respect to the support surface for this region of the trajectory. We compare our method Ours to the raw pointcloud Raw and the raw pointcloud with additional smoothing Smooth. The colorbars for error maps are placed on the right, with brighter colors indicating higher error. Our method outperforms the baselines in various environments, such as high grass hills (a), forests (b), and Hönggerberg (urban grasslands with impenetrable fences) (c).
Comparison of traversability estimation using the raw, smoothed, and our filtered pointcloud. The traversability is color-coded, where blue indicates traversable and red untraversable. Our method correctly predicts the traversability in the meadow and forest environment. Using the raw or smoothed pointcloud prohibits motion planning in the meadow.
Page updated
Report abuse