Vehicle Recognition Based on Carrier-Free UWB Radars Using Contrastive Multi-View Learning
Vehicle Recognition Based on Carrier-Free UWB Radars Using Contrastive Multi-View Learning
复制标题
DOI：
10.1109/lmwc.2022.3216048
复制
发表时间：
2023-03
期刊：
影响因子：
--
通讯作者：
Yuying Zhu;Shuning Zhang;Si Chen
中科院分区：
文献类型：
--
作者：
Yuying Zhu;Shuning Zhang;Si Chen
提出了一种基于雷达的新型无载波超宽带车辆自动识别系统。我们提供了完整的收发单元，设计了厘米级尺寸、带宽为0.89 ~ 5.02 GHz的超宽带Vivaldi天线，并提出了基于多视点的对比学习算法用于目标识别。将天线的仿真结果与实测结果进行了比较，结果表明两者吻合较好。虽然目标识别受益于大量的、精心策划的标记数据，但它在有限的注释数据问题中的应用仍然是一个挑战。此外，目标方面敏感问题也会影响模型的性能。为此，我们将对比学习与未标记的多视图数据相结合，以学习目标方面不变表示。在不同数据集上的实验证明了该方法的有效性和泛化性。
A novel carrier-free ultrawideband (UWB) radar-based automatic vehicle recognition system is reported. We provide complete transceiver units, design a UWB Vivaldi antenna with centimeter-scale dimensions and bandwidth from 0.89 to 5.02 GHz, and propose the multi-view-based contrastive learning algorithm for target recognition. Simulated results of the antenna are compared with measured results and are shown to be in good agreement. Though target recognition benefits from large, curated labeled data, its application to problems with limited annotated data remains a challenge. In addition, the target-aspect sensitive issue also impacts the performance of models. For that, we combine contrastive learning with unlabeled multi-view data to learn target-aspect-invariant representations. The experiments on the different datasets demonstrate the effectiveness and generalization of our method.