| 摘 要: 三维点云语义分割对自动驾驶感知至关重要,但现有方法在面向大规模室外场景时存在局部几何刻画
不足、通道判别性弱、远距离目标易丢失等问题。针对上述问题,提出一种面向自动驾驶场景的三维点云语义分割
方法,通过在编码器末端及跳跃连接处加入通道注意力机制强化边缘细节与远距离目标,抑制冗余。构建多尺度特
征融合 模 块,一 次 性 捕 获 局 部 细 节 与 全 局 上 下 文。加 入 可 学 习 坐 标 位 置 编 码,提 升 几 何 感 知 能 力。在
SemanticKITTI数据集上的实验结果表明,该方法的平均交并比(mIoU)和总体准确率(OA)分别达到57.7%和
91.1%,较基线网络RandLA-Net分别提升1.5和1.1个百分点。 |
| 关键词: 深度学习 点云语义分割 自动驾驶 RandLA-Net SemanticKITTI |
|
中图分类号:
文献标识码: A
|
| 基金项目: 辽宁省大学生创新训练计划项目资助(S202510140035) |
|
| Semantic Segmentation Method of 3D Point Cloud for Autonomous Driving Scene |
|
WANG Ziqian, CHEN Yufan, DUAN Keke
|
(Faculty of Information, Liaoning University, Shenyang 110036, China)
3632569518@qq.com; 3281519295@qq.com; duankeke@lnu.edu.com
|
| Abstract: Three-dimensional point cloud semantic segmentation is very important for autonomous driving
perception. However, the existing methods have some problems such as insufficient local geometric description, weak
channel discrimination and prone to missing long-range objects when facing large-scale outdoor scenes. Aiming at the
above problems, a semantic segmentation method of 3D point cloud for autonomous driving scenes is proposed, which
strengthens the edge details and long-distance targets by adding a channel attention mechanism at the end of the
encoder and the skip connection, and suppresses redundancy. A multi-scale feature fusion module is constructed to
capture local details and global context simultaneously. Learnable coordinate positional encoding is incorporated to
boost the model’s geometric perception capability.Experimental results on the SemanticKITTI dataset show that the
mean Intersection over Union (mIoU) and Overall Accuracy (OA) of this method reach 57.7% and 91.1% , respectively,
which are 1.5 and 1.1 percentage points higher than the baseline network RandLA-Net. |
| Keywords: deep learning point cloud semantic segmentation autonomous driving RandLA-Net SemanticKITTI |
