基于改进PointNet 模型的毫米波点云语义分割

Semantic Segmentation of Millimeter Wave Point Clouds Based on Improved PointNet

摘要: 毫米波成像技术在安检领域得到了普遍应用，研究基于毫米波点云图像的语义分割技术具有重要的意义。PointNet 采用了与任务无关的最远点采样(FPS)来逐步下采样点云，导致毫米波点云中为数不多的前景点信息丢失。因此，本文提出了一种基于自注意力机制的实例感知下采样点云语义分割网络。具体来说，本文结合自注意力机制实现面向任务的下采样策略来保留前景点，防止前景点信息的丢失。最后，由于毫米波点云图像中人体点云数量与前景点云数量极不平衡，改进使用Focal Loss作为语义分割损失函数以提升性能。实验结果表明，本文提出的语义分割模型相对于基准模型PointNet 在平均交并比(mIoU)方面有6.19%的提升，同时准确率有5.57%的提升。

Abstract: Millimeter-wave imaging technology has been widely applied in the security screening field, making the study of semantic segmentation techniques based on millimeter-wave point cloud images of paramount importance. PointNet employs a task-agnostic farthest point sampling (FPS) method for progressively downsampling the point cloud, resulting in the loss of sparse foreground information in the millimeter-wave point clouds. Consequently, this paper introduces an instance-aware downsampling point cloud semantic segmentation network based on the self-attention mechanism. Specifically, this work integrates the self-attention mechanism to implement a task-oriented downsampling strategy to preserve foreground points and prevent the loss of foreground information. Lastly, due to the significant imbalance between the number of human body points and foreground points in millimeter-wave point cloud images, an improved focal loss is utilized as the semantic segmentation loss function to enhance performance. Experimental results demonstrate that the semantic segmentation model proposed in this paper achieves a 6.19% improvement in mean Intersection over Union (mIoU) and a 5.57% increase in accuracy compared to the baseline model PointNet .