Eelis Peltola

Determining the drivable area (and more broadly, traversability) is a vital part of simultaneous localization and mapping (SLAM) when a mobile robot operates in a terrain with different factors of drivability. They can range from prioritizing asphalt roads over dirt roads in on-road scenarios to segmenting parts of the forest floor as hard-to-drive or non-drivable. Harsh and unstructured environments, such as ground surfaces featuring snow, vegetation, or gravel, all need some form of drivability analysis when using SLAM.

My research focuses mainly on exploring deep learning methodologies for the perception sensors commonly utilized on automated mobile machinery: camera and LiDAR. Because training data from harsh environments and unstructured terrain scenarios is scarce, special focus is given to expanding the capability of drivability estimation models in scenarios where little training data is available. This is done by utilizing simulated and augmented data together with real-world data. Well-generalizable models which are fast to tune for new scenarios are more reliable to perturbations and cheaper to start using.

I focus on mapping the drivable area as part of a SLAM approach. End-to-end autonomous driving models often do not have any independent drivability model, but then also lack clarity on what they consider drivable. Inside the drivable area research space, different approaches to the problem of drivability include methods using pre-tuned semantic segmentation and others using terrain unevenness.