基于CSI的零样本原型网络跨域活动识别
信号处理
2024-09-06 v2
摘要
无线传感的跨域能力目前是基于无线信号信道状态信息（CSI）的人类活动识别（HAR）面临的主要挑战之一。来自新域样本标注的困难促使了少样本和零样本策略的使用。在此背景下，原型网络因其合理的跨域可迁移性而备受关注。本文提出了一种新颖的零样本原型循环卷积网络，通过CSI实现HAR的零样本学习策略。该方法从可用的源域中提取原型，以对目标域中未见且未标记的数据进行相同或相似类别的分类。实验使用三个具有真实测量值的数据集进行，结果包括跨数据集评估。总体而言，结果改善了现有技术水平，使其成为跨域HAR的一种有前景的解决方案。
引用
@article{arxiv.2312.07076,
title = {CSI-Based Cross-Domain Activity Recognition via Zero-Shot Prototypical Networks},
author = {Guillermo Diaz and Iker Sobron and Inaki Eizmendi and Iratxe Landa and Manuel Velez},
journal= {arXiv preprint arXiv:2312.07076},
year = {2024}
}
备注
The authors have identified a significant error in the neural network configuration, specifically related to the addition of the LSTM layer after the CNN blocks and the method used to input data into the network. As a result, we have verified that the outcomes are inconsistent with what would be expected from a correctly configured neural network