Development of technology to distinguish contaminants using LiDAR near-field point cloud based on real-field data

Sensing technologies are essential to autonomous driving, enabling environmental perception and supporting navigation and its decision-making. However, most publicly available datasets are collected under perfect weather conditions and fail to capture the challenges of real-world environments. To address this gap, we built a large-scale dataset from extensive driving campaigns across diverse regions of the United States, covering conditions such as rain, snow, dust, clear weather and anti-ice material affects. Data were acquired through a multimodal sensing devices - including perception cameras, LiDAR, infrared sensors, and external monitoring systems - providing a comprehensive representation of the vehicle's operational environment. Building on this dataset, our study focuses on LiDAR data, emphasizing contaminant classification as a critical requirement for sensor-cleaning strategies essential to reliable autonomous driving. Using raw LiDAR signals collected over 22,000km of driving, we developed a supervised learning-based framework that defines twelve feature functions to classify eight types of contaminants. Although currently limited to 8 classes, the framework is scalable and capable of accommodating additional contaminant types as environmental variability increases. This advancement enables the optimization of sensor-cleaning systems, a prerequisite for achieving Lv.4 autonomous driving. Moreover, its integration into vehicle development could support manufacturers in minimizing cost and weight while maximizing efficiency and profitability.