| 摘 要: 为了实现三维点云在室内和工业环境中的实际应用,文章改进了传统的目标检测转换器(Detection Transformer, DeTR)神经网络,并提出了一种基于分层抽象的多层点云特征提取方法;同时,设计了曲面特征提取模块对三维点云进行预处理,增强了点云的附加特征。在公开数据集ScanNet V2和工业室内数据集上对本文方法进行实验验证和评估,该方法在ScanNet V2上的mAP@0.5准确率超过最先进的模型(State-of-the-Art, SOTA) CAGroup3d,达到76.0%;在ScanNet V2上的mAP@0.25准确率超过最先进的模型CAGroup3d,达到62.2%,消融实验进一步验证了所述方法的准确性和高效性。 |
| 关键词: 三维点云;目标检测;工业环境;Transformer |
|
中图分类号: TP389.1
文献标识码: A
|
| 基金项目: 浙江省自然科学基金项目(LY22F020021);浙江省重点研发计划“领雁”项目(2023C01145);国家自然科学基金项目(61802095,61572162) |
|
| Indoor 3D Point Cloud Object Detection with Multi-scale Features and Attention Mechanism |
|
GU Fangyu, HU Haiyang
|
(School of Computer Science and Technology, Hangzhou Dianzi University, Hangzhou 310018, China)
gfy2345@163.com; huhaiyang@hdu.edu.cn
|
| Abstract: To facilitate the practical application of 3D point cloud in indoor and industrial environments, this paper proposes a multi-layer point cloud feature extraction method based on hierarchical abstraction with the improvement of traditional object Detection Transformer (DeTR) neural network. Additionally, a surface representation module for preprocessing 3D point cloud is designed to enhance the additional features of the 3D point cloud. Experimental validation and assessment of the proposed method are conducted on the public dataset ScanNet V2 and an industrial indoor dataset. Experiment results show that the mAP@ 0.5 accuracy of the proposed method exceeds the State-of-the-Art (SOTA) model CAGroup3d, reaching 76.0% ; the mAP@ 0. 25 accuracy exceeds the SOTA model CAGroup3d, reaching 62.2% . The ablation experiment further validates the accuracy and efficiency of the method. |
| Keywords: 3D point cloud; object detection; industrial environment; Transformer |
