BOOSTING U-NET WITH FOCAL LOSS FOR ROAD MARKING CLASSIFICATION ON SPARSE MOBILE LIDAR POINT CLOUD DERIVED IMAGES

Lagahit, M. L. R.; Lagahit, M. L. R.; Matsuoka, M.; Matsuoka, M. · 2022 · DOAJ

DOI: 10.5194/isprs-annals-V-5-2022-33-2022

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of accurately classifying road markings from sparse mobile LIDAR point clouds, a common limitation when using lower-cost sensors for High-Definition (HD) map generation. While survey-grade LIDAR sensors produce dense point clouds suitable for deep learning classification, they are expensive. Lower-cost sensors, such as those found in autonomous vehicles, generate sparser data that often results in poor feature representation and reduced classification accuracy. The study aims to improve the performance of the U-Net convolutional neural network (CNN) on these sparse datasets by replacing the standard cross-entropy loss function with a weighted focal loss, which better handles class imbalance. The methodology involved collecting data using a Velodyne Puck (VLP-16) LIDAR sensor mounted on a small autonomous vehicle at the Tokyo Institute of Technology campus. The raw point clouds were geometrically filtered to focus on the roadway and projected into top-down images with a 1 cm x 1 cm resolution, using intensity values for pixel data. An annotated dataset was created through intensity filtering, manual cleaning, and labeling of three classes: road markings, ground, and black (no value). The dataset was augmented via flipping and split into 90% for training and 10% for testing. Six U-Net models were trained, varying the loss function and class weights. Models A–C used cross-entropy loss, while Models D–F used focal loss with increasing weights assigned to the road marking class (up to 99.998%) to prioritize its detection over the background and empty pixels. The results demonstrated that incorporating focal loss significantly improved classification metrics. Recall for road markings increased by up to 94.68% compared to the baseline cross-entropy model, rising from 1.86% to 96.54%. However, this shift caused a decrease in precision due to an increase in "black" pixel misclassifications, which do not correspond to actual point cloud data. By adjusting the precision calculation to exclude these insignificant "black" pixels, the F1-score for road markings improved substantially, reaching a maximum of 86.39%. This performance is comparable to or better than previous studies using dense point clouds. Additionally, the study extended the method to generate classified 3D point clouds by combining the classified images with depth maps derived from elevation values. The significance of this work lies in its potential to enable the use of affordable, low-level LIDAR sensors for HD map creation and autonomous vehicle navigation. By successfully classifying road markings on sparse data with high accuracy, the proposed method reduces the reliance on expensive survey-grade equipment. The findings suggest that weighted focal loss is an effective strategy for handling class imbalance in sparse LIDAR-derived imagery, offering a viable pathway for cost-effective, automated road marking extraction and digital map updating.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• signage environment