Development of Technology to Distinguish Contaminants using LiDAR Near-Field Point Cloud based on Real-Field Data
- Delivery
- Provide download link
- Format
- Price
- Non-members (tax incl.)：￥8,250 Members (tax incl.)：￥6,600
- Paper/Info type
- SAE Paper
No.2026-01-0020
- Pages
- 1-8(Total 8 p)
- Date of publication
- Apr 2026
- Publisher
- SAE International
- Language
- English
- Event
- WCX™ 2026
Detailed Information
| Author(E) | 1) Hunjae Kim |
|---|---|
| Affiliation(E) | 1) Hyundai Motor Co. |
| Abstract(E) | Reliable environmental perception under adverse and contaminated conditions is a critical requirement for autonomous driving systems. Although LiDAR sensors play a central role in such perception, their performance is significantly degraded by surface contamination caused by environmental factors such as rain, snow, dust, anti-icing materials, and bug splatter impacts. However, most existing public datasets and prior studies rely on simulated or laboratory-generated contamination scenarios, which limit their applicability to real-world autonomous driving. To address this gap, we construct a large-scale real-world dataset collected from approximately 22,000 km of on-road driving across diverse regions of the United States, covering a wide range of naturally occurring environmental contamination conditions. The dataset was acquired using a multimodal sensing platform integrating LiDAR, perception RGB cameras, infrared camera sensors, and external monitoring systems, enabling comprehensive observation of sensor behavior under realistic operating environments. Based on this dataset, we propose a scalable contaminant classification framework that focuses on LiDAR surface contamination. A key contribution of this study is the introduction and exploitation of near-field point cloud features, which capture backscattered laser signals caused by surface contamination and exhibit a strong correlation with contamination severity and type. Using raw LiDAR signals, we utilize sixteen feature functions and train supervised learning models to classify seven distinct contaminant categories. Experimental results demonstrate that the proposed approach achieves classification accuracy exceeding 95% under real-world driving conditions, significantly outperforming prior laboratory-based studies. Furthermore, the framework is designed for practical deployment and can be extended to additional contaminant types and geographic regions through incremental data collection and learning. The proposed methodology enables real-time identification of LiDAR contamination sources, providing a critical foundation for adaptive sensor-cleaning strategies. By supporting contamination-aware sensor maintenance, this work contributes to cost- and weight-efficient sensor system design and represents an essential step toward achieving reliable Level 4 autonomous driving. |
