XModal-ID: Using WiFi for Through-Wall Person Identification from Candidate Video Footage
XModal-ID: Using WiFi for Through-Wall Person Identification from Candidate Video Footage
复制标题
DOI：
10.1145/3300061.3345437
复制
发表时间：
2019-08
期刊：
影响因子：
--
通讯作者：
Belal Korany;Chitra R. Karanam;H. Cai;Yasamin Mostofi
中科院分区：
文献类型：
--
作者：
Belal Korany;Chitra R. Karanam;H. Cai;Yasamin Mostofi
在本文中，我们提出了XModal-ID，一种新的WiFi视频跨模态步态的身份识别系统。给定当一个未知的人在一个未知的区域行走时测量到的WiFi信号和一个行走的人在另一个区域的视频片段，XModal-ID可以确定这两种情况下是否是同一个人。XModal-ID仅使用一对现成WiFi收发器的信道状态信息（CSI）幅度测量。它不需要对要识别的人进行任何预先的无线或视频测量。同样，它不需要任何关于操作区域或人员轨迹的知识。最后，它可以通过墙壁识别人。XModal-ID利用视频片段来模拟如果视频中的人走近一对WiFi收发器时会产生的WiFi信号。然后，它使用一种新的处理方法来鲁棒地提取关键步态特征从两个真实的WiFi信号和基于视频的模拟信号，并将它们进行比较，以确定是否在WiFi区域的人是同一个人在视频中。我们通过构建一个大型测试集来广泛评估XModal-ID，该测试集包含8 $主题，2 $视频区域和5 $ WiFi区域，包括3个穿墙区域以及复杂的步行路径，所有这些都在训练阶段看不到。总的来说，我们总共有2，256个WiFi视频测试对。然后，XModal-ID在预测一对WiFi和视频样本是否属于同一个人方面达到了85%的准确率。此外，在XModal-ID将WiFi样本与$8$候选视频样本进行比较的排名场景中，它获得了$75%$、$90%$和$97%$的前1、前2和前3准确度。这些结果表明，XModal-ID可以在各种实际场景中鲁棒地识别在新环境中行走的新人。
In this paper, we propose XModal-ID, a novel WiFi-video cross-modal gait-based person identification system. Given the WiFi signal measured when an unknown person walks in an unknown area and a video footage of a walking person in another area, XModal-ID can determine whether it is the same person in both cases or not. XModal-ID only uses the Channel State Information (CSI) magnitude measurements of a pair of off-the-shelf WiFi transceivers. It does not need any prior wireless or video measurement of the person to be identified. Similarly, it does not need any knowledge of the operation area or person's track. Finally, it can identify people through walls. XModal-ID utilizes the video footage to simulate the WiFi signal that would be generated if the person in the video walked near a pair of WiFi transceivers. It then uses a new processing approach to robustly extract key gait features from both the real WiFi signal and the video-based simulated one, and compares them to determine if the person in the WiFi area is the same person in the video. We extensively evaluate XModal-ID by building a large test set with $8$ subjects, $2$ video areas, and $5$ WiFi areas, including 3 through-wall areas as well as complex walking paths, all of which are not seen during the training phase. Overall, we have a total of 2,256 WiFi-video test pairs. XModal-ID then achieves an $85%$ accuracy in predicting whether a pair of WiFi and video samples belong to the same person or not. Furthermore, in a ranking scenario where XModal-ID compares a WiFi sample to $8$ candidate video samples, it obtains top-1, top-2, and top-3 accuracies of $75%$, $90%$, and $97%$. These results show that XModal-ID can robustly identify new people walking in new environments, in various practical scenarios.