Abstract: Current 3D object detection models predominantly utilize data-driven deep learning techniques, in which dataset quality plays a pivotal role in model performance. To mitigate the scarcity of coal dust datasets and the labor-intensive process of real-world data collection, a point cloud data augmentation method is proposed based on the scattering and absorption effects of coal dust on LiDAR electromagnetic waves. By incorporating the optical characteristics of coal dust particles, we develop a propagation simulation model to characterize LiDAR signal attenuation and scattering phenomena within coal dust-laden atmospheres. Real point cloud data acquired under clear weather conditions are subsequently transformed via the simulation model, adjusting 3D coordinates and reflection intensities to synthesize coal dust-adaptive point clouds. Five state-of-the-art models (PV-RCNN++, PV-RCNN, PointRCNN, PointPillars, Voxel-RCNN_Car) are trained on the augmented dataset, demonstrating consistent accuracy improvements in coal dust scenarios. The most architecturally complex model, PV-RCNN, improved performance by 1.88, 1.74, and 0.84 percentage points in the car, pedestrian, and cyclist categories, respectively . Results indicate that using the augmented point cloud data to train object detection models significantly improves detection accuracy in coal dust environments in comparison with models trained in clear conditions.The proposed methodology enables robust perception in open-pit mining operations, thereby advancing reliable data support for autonomous mining vehicle deployment.

Key words: autonomous mining truck, LiDAR, three-dimensional point cloud, data augmentation, physics simulation, object detection

摘要： 当前三维目标检测模型大都基于数据驱动的深度学习技术，数据集的质量对模型的性能至关重要。本文针对煤尘环境数据集缺失、建立真实煤尘环境数据集费时费力的问题，提出了一种基于煤尘对激光雷达电磁波散射和吸收效应的点云数据增强方法。该方法针对煤尘粒子的光学特性，构建了雷达波在煤尘中的传播仿真模型，模拟激光雷达信号在煤尘环境中的衰减与散射。然后，在晴朗环境下采集的真实点云数据基础上，基于仿真模型对点云的三维坐标和反射强度进行修正，生成符合煤尘环境感知特性的仿真点云数据。在增强后的仿真数据集上训练并测试了五种主流三维目标检测模型（PV-RCNN++、PV-RCNN、PointRCNN、PointPillars、Voxel_RCNN_Car），这五种检测模型在煤尘环境下的检测精度均有所提升，其中模型复杂度最高的PV-RCNN模型在汽车、行人和骑行者类别上的表现分别提高了1.88、1.74、0.84个百分点。实验结果表明：在煤尘环境中，使用增强后的点云数据训练的目标检测模型，相较于在晴朗条件下训练的模型，其检测精度显著提升，能更可靠地感知露天矿复杂环境，为无人驾驶矿车的稳定运行提供了数据支撑。

关键词: 无人驾驶矿车, 激光雷达, 三维点云, 数据增强, 物理仿真, 目标检测