LIDAR-BASED LANE DETECTION IN ADVERSE ROAD ENVIRONMENTS

Matthew D. Jabson; Derek Shiboo; Behnam Bahr

Download Paper | Permalink

Conference: 2025 ASEE PSW Conference
Location: California Polytechnic University, California
Publication Date: April 10, 2025
Start Date: April 10, 2025
End Date: April 12, 2025
DOI: 10.18260/1-2--55181
Permanent URL: https://peer.asee.org/55181

Abstract

Lane detection in autonomous vehicles is a critical task that ensures safe and reliable navigation. However, in underdeveloped road environments, lane detection becomes particularly challenging due to poorly defined or worn-out lane markings, unstructured roads, and varying topographies. These are worsened by rain, fog, and snow, which introduce noise, reduce point cloud density, and obscure critical features in both visual and LiDAR-based systems. LiDAR can enable high-resolution 3D spatial data generation, hence motivating its use as a promising alternative to traditional camera-based systems. However, in such cases, successful lane detection requires advanced techniques that could account for the lack of clear lane markings and for the impact of weather-related data degradation. This paper proposes a deep learning-based LiDAR lane detection framework, including cost maps, to make the system more robust in underdeveloped and harsh road conditions. Cost maps serve as a representation of passability and their road features include curbs, surface slopes, elevation variations, and roughness. A cost map in this context assigns lower values to regions that are easier to drive on and higher values to locations representative of impassable space: off-road areas, curbs, or obstacles. The proposed system focuses on the extraction and interpretation of geometric and structural features of the road, instead of exclusively relying on visual lane indicators, through the incorporation of cost maps within the lane detection framework. The system uses some sophisticated pre-processing methods for LiDAR point cloud data to address the challenges brought about by adverse weather conditions. For instance, some employed techniques that aim at reducing noise caused by raindrops, snowflakes, or atmospheric scattering include the statistical removal of outliers and voxel grid filtering. The proposed model uses Gaussian smoothing to alleviate the impacts of sparse or irregular data distribution in foggy situations. Moreover, domain adaptation is applied to enhance model robustness to changes in the quality of LiDAR data due to variation in weather conditions. The methodology for domain adaptation will involve adversarial training where the model learns generalization over multiple domains by emulating weather-induced deterioration during the training process. The deep learning framework employs cutting-edge architectures including PointNet++ and Graph Convolutional Networks (GCNs). PointNet++ is utilized to analyze unprocessed point clouds and derive localized features, whereas GCNs analyze cost map data through the modeling of spatial relationships and geometric constraints present in the roadway context. Taken together, these models consider road surface characteristics along with contextual clues-including curbs and elevation changes-to detect lane boundaries when explicit markings are absent. By integrating cost maps, the system is able to add additional layers of spatial reasoning, where it prioritizes drivable paths according to the traversal metrics encoded in the cost map. The system has been validated using a combination of real-world and synthetic datasets. These datasets represent scenarios with poorly marked or completely unmarked roads, adverse weather conditions, and variable topography. For the evaluation of the proposed framework, Intersection over Union (IoU) is used for lane localization, Root Mean Square Error (RMSE) is considered for cost map precision, and inference time is used for establishing the possibility of real-time applications. Empirical results show that the integration of cost maps with deep learning methods improves accuracy and robustness in lane boundary detection compared to state-of-the-art vision-based and hybrid LiDAR systems. In particular, this system outperforms the baseline approach in scenarios involving faded markings, unstructured roads, and adverse weather. By combining cost maps with advanced preprocessing techniques, this work advances the current state-of-the-art in LiDAR-based lane detection and addresses the dual challenges presented by poorly developed road infrastructure and deterioration of data due to weather conditions. The use of cost maps enhances road features, which in turn enables the system to adapt to complex and disorganized scenarios. This research presents a scalable and resilient approach towards autonomous navigation, improving the safety and reliability of an autonomous vehicle in various complex and demanding international road environments. Keywords: Lane detection, LiDAR, cost maps, deep learning, autonomous vehicles, PointNet++, Graph Convolutional Networks, unstructured roads, adverse weather, point cloud processing, domain adaptation, real-time applications, road geometry analysis.

Citation
Format

Jabson, M. D., & Shiboo, D., & Bahr, B. (2025, April), LIDAR-BASED LANE DETECTION IN ADVERSE ROAD ENVIRONMENTS Paper presented at 2025 ASEE PSW Conference, California Polytechnic University, California. 10.18260/1-2--55181

TY - CPAPER
AB - Lane detection in autonomous vehicles is a critical task that ensures safe and reliable navigation. However, in underdeveloped road environments, lane detection becomes particularly challenging due to poorly defined or worn-out lane markings, unstructured roads, and varying topographies. These are worsened by rain, fog, and snow, which introduce noise, reduce point cloud density, and obscure critical features in both visual and LiDAR-based systems. LiDAR can enable high-resolution 3D spatial data generation, hence motivating its use as a promising alternative to traditional camera-based systems. However, in such cases, successful lane detection requires advanced techniques that could account for the lack of clear lane markings and for the impact of weather-related data degradation.
This paper proposes a deep learning-based LiDAR lane detection framework, including cost maps, to make the system more robust in underdeveloped and harsh road conditions. Cost maps serve as a representation of passability and their road features include curbs, surface slopes, elevation variations, and roughness. A cost map in this context assigns lower values to regions that are easier to drive on and higher values to locations representative of impassable space: off-road areas, curbs, or obstacles. The proposed system focuses on the extraction and interpretation of geometric and structural features of the road, instead of exclusively relying on visual lane indicators, through the incorporation of cost maps within the lane detection framework. The system uses some sophisticated pre-processing methods for LiDAR point cloud data to address the challenges brought about by adverse weather conditions. For instance, some employed techniques that aim at reducing noise caused by raindrops, snowflakes, or atmospheric scattering include the statistical removal of outliers and voxel grid filtering. The proposed model uses Gaussian smoothing to alleviate the impacts of sparse or irregular data distribution in foggy situations. Moreover, domain adaptation is applied to enhance model robustness to changes in the quality of LiDAR data due to variation in weather conditions. The methodology for domain adaptation will involve adversarial training where the model learns generalization over multiple domains by emulating weather-induced deterioration during the training process. The deep learning framework employs cutting-edge architectures including PointNet++ and Graph Convolutional Networks (GCNs).
PointNet++ is utilized to analyze unprocessed point clouds and derive localized features, whereas GCNs analyze cost map data through the modeling of spatial relationships and geometric constraints present in the roadway context. Taken together, these models consider road surface characteristics along with contextual clues-including curbs and elevation changes-to detect lane boundaries when explicit markings are absent. By integrating cost maps, the system is able to add additional layers of spatial reasoning, where it prioritizes drivable paths according to the traversal metrics encoded in the cost map.
The system has been validated using a combination of real-world and synthetic datasets. These datasets represent scenarios with poorly marked or completely unmarked roads, adverse weather conditions, and variable topography. For the evaluation of the proposed framework, Intersection over Union (IoU) is used for lane localization, Root Mean Square Error (RMSE) is considered for cost map precision, and inference time is used for establishing the possibility of real-time applications. Empirical results show that the integration of cost maps with deep learning methods improves accuracy and robustness in lane boundary detection compared to state-of-the-art vision-based and hybrid LiDAR systems. In particular, this system outperforms the baseline approach in scenarios involving faded markings, unstructured roads, and adverse weather. By combining cost maps with advanced preprocessing techniques, this work advances the current state-of-the-art in LiDAR-based lane detection and addresses the dual challenges presented by poorly developed road infrastructure and deterioration of data due to weather conditions. The use of cost maps enhances road features, which in turn enables the system to adapt to complex and disorganized scenarios. This research presents a scalable and resilient approach towards autonomous navigation, improving the safety and reliability of an autonomous vehicle in various complex and demanding international road environments.
Keywords: Lane detection, LiDAR, cost maps, deep learning, autonomous vehicles, PointNet++, Graph Convolutional Networks, unstructured roads, adverse weather, point cloud processing, domain adaptation, real-time applications, road geometry analysis.

AU - Matthew D. Jabson
AU - Derek Shiboo
AU - Behnam Bahr
CY - California Polytechnic University, California
DA - 2025/04/10
PB - ASEE Conferences
TI - LIDAR-BASED LANE DETECTION IN ADVERSE ROAD ENVIRONMENTS
UR - https://peer.asee.org/55181
DO - 10.18260/1-2--55181
ER -

LIDAR-BASED LANE DETECTION IN ADVERSE ROAD ENVIRONMENTS

Paper Authors

Matthew D. Jabson California State Polytechnic University, Pomona

Derek Shiboo California State Polytechnic University, Pomona

Behnam Bahr California State Polytechnic University, Pomona

Abstract

Citation

APA

APA - LaTeX bibitem

MLA

MLA - LaTeX bibitem

Bibtex

EndNote - RIS