| 赵宇轩,贾克斌,李志坚.基于编解码结构的激光雷达动态物体点云分割网络[J].高技术通讯(中文),2024,34(10):1091~1097 |
| 基于编解码结构的激光雷达动态物体点云分割网络 |
| Dynamic object point cloud segmentation network for LiDARbased on encoder-decoder structure |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2024. 10. 008 |
| 中文关键词: 激光雷达; 深度学习; 点云分割; 自动驾驶 |
| 英文关键词: LiDAR,deep learning,point cloud segmentation,autonomous driving |
| 基金项目: |
| 作者 | 单位 | | 赵宇轩 | (北京工业大学信息学部 北京 100124)
(北京工业大学计算智能与智能系统北京市重点实验室 北京 100124)
(先进信息网络北京实验室 北京 100124) | | 贾克斌 | | | 李志坚 | |
|
| 摘要点击次数: 229 |
| 全文下载次数: 197 |
| 中文摘要: |
| 随着科技的不断发展,自动驾驶技术越来越多地进入到人们的生活中。 动态物体点云分割是其中十分关键的一项任务,它可以为地图构建、路径规划等任务提供前置帮助。 本文提出一种基于编码器-解码器结构的激光雷达动态物体点云分割网络,使用自校正卷积替换上下文特征提取模块中的普通卷积,提升神经网络的特征学习能力;并在网络解码阶段加入通道注意力机制,提升网络对重要特征通道的关注学习程度,从而达成更好的分割效果。 本文在 SemanticKITTI MOS 数据集上进行实验,实验结果表明,本文所提出的动态物体点云分割网络相比原有方法取得更优表现,交并比(IoU)达到 72. 1% 。 |
| 英文摘要: |
| With the continuous development of science and technology, automatic driving technology has increasingly en-
tered people’s lives. Dynamic object point cloud segmentation is one of the key tasks, which can provide pre-help
for map construction, path planning and other tasks. In this paper, a LiDAR dynamic object point cloud segmenta-
tion network based on encoder-decoder structure is proposed, which uses self-correcting convolution to replace com-
mon convolution in context feature extraction module to improve the feature learning ability of neural network. The
channel attention mechanism is added in the decoding stage of the network to improve the learning degree of the net-
work’s attention to important feature channels, so as to achieve better segmentation effect. In this paper, experiments
are carried out on the SemanticKITTI MOS dataset, and the experimental results show that the proposed dynamic ob-
ject point cloud segmentation network achieves better performance than the original method, with an intersection over
union (IoU) of 72. 1% . |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
