针对无序、采样不均匀以及存在相互遮挡的工件点云分割效果不佳的问题,提出一种多尺度自适应通道维度注意力点云分割网络(PointECA).该算法中的多尺度特征提取模块能够较好地融合不同尺度的局部邻域特征,得到较为丰富的全局特征信息;自适应性通道注意力模块能够对不同尺度局部特征的通道维度交互学习,实现较好的语义分割效果.此外,制作了用于语义分割实验的Workpieces数据集.大量实验数据表明:PointECA在无序且有相互遮挡场景下,对工件部件分割的平均交并比达到了95.42%,能够为无序工件的快速分拣提供较好的条件.
To address the problems of disorder,uneven sampling,and the poor segmentation of workpiece point clouds with mutual occlusion,a multiscale adaptive channel attention point cloud segmentation network(PointECA)was proposed.In this algorithm,multi-scale feature extraction module was used to better fuse the local neighborhood features of different scales and richer global feature infor-mation was obtained;the adaptive channel attention module was used to interactively learn the channel dimensions of local features at different scales to achieve a better semantic segmentation effect.In addition,the Workpieces dataset for semantic segmentation experi-ments was produced.A large amount of experimental data shows that PointECA achieves 95.42%mean intersection over union for work-piece part segmentation in disordered and mutually occluded scenes,which can provide better conditions for the fast sorting disordered workpieces.