【摘要】 针对激光雷达在煤棚环境下测量煤堆时感知数据噪声大、物体不易区分等问题，提出了一种基于自监督AFF-GrowSP语义分割网络的多煤堆体积自动化测量方法。首先采用基于优化思想的混合滤波算法，对感知数据进行预处理，再将其通过改进的GrowSP分割网络执行分割操作，引入特征融合模块AFF动态调节不同尺度特征的融合权重;最后，使用基于MLS的贪婪三角化算法对分割后的煤堆点云进行三维重建，并计算其体积。结果表明:在公开数据集S3DIS上，AFF-GrowSP分割网络的平均交并比、整体准确率分别达到45.2%和79.4%，相较于GrowSP提升了0.4%和0.5%;在真实矿井煤棚场景下的多煤堆建模和体积测量应用中，该方法对多个不规则煤堆体积测量的平均相对误差为3.09%，实现了多个不规则煤堆体积的高精度测量，解决了煤棚环境下煤堆点云分割的难题，为煤炭储量的自动化管理提供参考。更多还原

【Abstract】 To address issues in coal shed environments,such as noisy light detection and ranging(LiDAR) perception data and indistinct object separation during coal pile measurement,an automated multiple coal pile volume measurement method was proposed.This method is based on a self-supervised semantic segmentation network called AFF-GrowSP.First,perception data were preprocessed using a hybrid filtering algorithm based on optimization principles.Then,the preprocessed data were segmented by an improved GrowSP segmentation network.This network incorporated an Adaptive Feature Fusion(AFF) module to dynamically adjust the fusion weights of features at different scales.Finally,a greedy triangulation algorithm based on Moving Least Squares(MLS) was employed to reconstruct the 3D structure of the segmented coal pile point cloud and calculate its volume.The results indicate that: On the public dataset S3 DIS,the AFF-GrowSP segmentation network achieves a mean Intersection over Union(m IoU) of 45.2% and an Overall Accuracy(OA) of 79.4%.This represents an improvement of 0.4%in m IoU and 0.5% in OA compared with the original GrowSP.In the application of multiple coal pile modeling and volume measurement in real-world mine coal shed scenarios,this method achieves an average relative error of 3.09% for measuring the volumes of multiple irregular coal piles.This method solves the challenging problem of coal pile point cloud segmentation in coal shed environments.It enables high-precision measurement of multiple irregular coal pile volumes and provides a valuable reference for the automated management of coal reserves.更多还原