Semantic Grid Mapping based on Surface Classification with Supervised Learning |
|---|
| Torsten Engler, Felix Ebert, Hans-Joachim Wuensche |
- Abstract:
- LiDAR-based occupancy grid mapping can lead to overly conservative detection of obstacles in non-urban autonomous driving scenarios, e.g. grass in the middle of the lane is often interpreted as obstacle although it is actually driveable. We therefore aim to augment our current grid-based environment representation with additional information derived from pixel-level semantic segmentation in camera images. We project the resulting segmentation map onto an additional semantic layer in the environment grid representation by utilizing LiDAR data for pixel-to-cell association to improve our driveability analysis. We apply supervised machine learning techniques for pix- elwise prediction of class labels. Datasets for non-urban en- vironments are rare. Therefore, we created a custom dataset. Due to the huge effort necessary to create such a dataset, its size is relatively small and hence neural networks might not be able to train effectively. Thus, low numbers of training samples require a careful choice of the classifier and/or data augmentation techniques. We therefore compare classification performance of neural networks with random forest classifiers.
- Download:
- IMEKO-TC17-2018-008.pdf
- DOI:
- -