Sparse Convolutional 3D Neural Networks for the Assessment of Environment Traversability
Antonio Santo, Arturo Gil, David Valiente, Álvaro Martínez and Enrique Heredia
Lecture Notes in Electrical Engineering  (2025)
Ed. Springer  ISBN:eBook ISBN 978-3-031-94989-0 Print ISBN 978-3-031-94988-3  ISSN:ISSN 1876-1100 ISSN 1876-1119 (electronic)  DOI:https://doi.org/10.1007/978-3-031-94989-0  - Vol. 1436. págs. 199-218

Resumen:

 Ensuring accurate assessment of the surrounding environment is crucial


for the efficient operation of autonomous mobile robots, especially when faced with


the complexities of unfamiliar and natural terrain that lacks a predefined structure. In


this context, traversability assessment is presented as a fundamental component of


the autonomous navigation. This research introduces a systematic methodology that


employs a LiDAR sensor to capture detailed 3D point clouds, thus facilitating the


analysis of traversability regions on both conventional roads and natural scenarios.


The proposed approach integrates a well-structured sparse encoder-decoder config-


uration with rotation invariant features. This configuration is meticulously designed


to replicate the input data by associating the acquired traversability features to each


individual point in the 3D point cloud. Experimental results confirm the robustness


and effectiveness of our method, especially in outdoor environments. Notably, our


approach outperforms existing methodologies, making a significant contribution to


the ongoing progress in the field of autonomous robotic navigation.