Wi-Fi Passive Person Re-Identification based on Channel State Information

11/12/2019 ∙ by Danilo Avola, et al. ∙ Sapienza University of Rome 0

With the increasing need for wireless data transfer, Wi-Fi networks have rapidly grown in recent years providing high throughput and easy deployment. Nowadays, Access Points (APs) can be found easily wherever we go, therefore Wi-Fi sensing applications have caught a great deal of interest from the research community. Since human presence and movement influence the Wi-Fi signals transmitted by APs, it is possible to exploit those signals for person Re-Identification (Re-ID) task. Traditional techniques for Wi-Fi sensing applications are usually based on the Received Signal Strength Indicator (RSSI) measurement. However, recently, due to the RSSI instability, the researchers in this field propose Channel State Information (CSI) measurement based methods. In this paper we explain how changes in Signal Noise Ratio (SNR), obtained from CSI measurements, combined with Neural Networks can be used for person Re-ID achieving remarkable preliminary results. Due to the lack of available public data in the current state-of-the-art to test such type of task, we acquired a dataset that properly fits the aforementioned task.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 3

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

With recent developments in the Internet of Things (IoT) technology, Wi-Fi infrastructures have become common in both public and private areas. Radio signals transmitted by several Access Points (APs), more than just connecting to the Internet, can be exploited for various Wi-Fi sensing tasks [1, 2, 3], including person Identification (ID) [4] and Re-Identification (Re-ID). Person Re-ID is the task aimed to recognize an already known person identity by comparing a probe (i.e., person of interest) to a gallery of candidates and, generally, it is performed by exploiting visual data requiring specific sensors in specific locations. Considering a real-world indoor surveillance scenario, for example, several cameras strategically located at different rooms need to be used. Moreover, visual-based methods have to face the well-known vision challenges related to occlusions, illumination changes and similarity/appearance issues among human subjects. Differently, Wi-Fi based techniques could be new non-invasive and alternative ways to perform the Re-ID task, even more robust to aforementioned vision related challenges. Notice that Wi-Fi signals pass-through obstructions, for example. In addition, whereas the use of cameras may violate human rights or subject’s privacy, radio signals do not suffer from such issues.

Either human presence or movement between Wi-Fi transmitters and receivers influence the radio signals transmitted [5], thus Wi-Fi based sensing applications are feasible by analysing both the Channel State Information (CSI) measurement [6] from the Wi-Fi signal and the features that are possible to extract by processing it (e.g., signal amplitude or phase). Specifically, the CSI describes the characteristics of the wireless communication channel and, in detail, such characteristics represent how the wireless signal propagate from the transmitter to the receiver at specific carrier frequencies along multiple paths. This measure allows inferring many detail about the signal, including scattering or attenuation based on path length. Methods based on wireless signals can be categorised in Active (i.e., Device-Based) [7, 8] and Passive (i.e., Device-Free) [9, 10] approaches. Whereas the former perform Wi-Fi sensing tasks by exploiting a device worn by the human subject, the latter do not. Such techniques, especially the Device-Free ones, add novelty to the state-of-the-art and they found application in a wide range of fields such as surveillance, security, healthcare, monitoring, imaging, human-computer interaction and many others.

In this paper, we propose a Device-Free Wi-Fi method able to re-identify person identities in different indoor environments. Our approach, by using a neural network for re-identification, exploits CSI measurements to estimate the Signal Noise Ratio (SNR) of received wireless signals. Through preliminary experiments, has been verified that human biometrics are embedded even in SNR estimated from such signals. In addition, due to lack of publicly available datasets, we also collected data to validate the proposed method. As far as we know, in current state-of-the-art we are the first in exploiting the CSI for the Re-ID task. Moreover, our work demonstrates that the use of commercial Wi-Fi devices is possible to capture human biometrics for people Re-ID in real-world indoor scenarios.

The paper is structured as follows. In Section 2, a brief overview about the current literature concerning Wi-Fi human-related sensing applications is reported. In Section 3, the proposed method architecture is described in detail, including CSI measurement and SNR estimation. In Section 4, starting with the description of the datasets collected to fit the Wi-Fi person Re-ID task, the obtained experimental results are discussed. Finally, in Section 5 the paper is concluded.

2 State-of-the-art

Tipically, for Wi-Fi sensing systems development the most used measurement is the Received Signal Strength Indicator (RSSI), which expresses the relative signal quality by indicating the power level being received after any possible loss at the antenna and cable levels. The authors of [11] reported that human presence influences the performance of wireless communications, causing significantly RSSI fluctuation in both line-of-sight (LOS) and non-line-of-sight (NLOS) conditions. Indeed, such measurement is widely used for human localization [12, 13, 14], detection [15, 16]

and pose estimation

[17] tasks. In [14], the RSSI has been even used for person identification by exploiting a proximity based algorithm in a Device-Based setting. Wireless sensors have been placed on specific environmental objects, the user wearing the device that features the highest RSSI is identified as the one actually interacting with the surrounding area. However, even if the RSSI is easy to obtain, it is an unstable and noisy measure unable to capture the real changes in the signal, due to multi-path and fading effects, offering limited performance [18].

More and more researchers recently start to use the Channel State Information for its great signal processing capability. Notice that, the CSI allows the use of both signal processing and computer vision algorithms. In

[9], the authors propose an automatic Wi-Fi system for human detection and pose estimation by exploiting CSI measurements to extract the Wi-Fi signal amplitude. In [19], instead, the authors introduce the theoretical analysis of the sensing capability of Wi-Fi signals, by defining a Fresnel zone [20] model for human breathing recognition and pose estimation. Differently, the authors of [21], perform Wi-Fi human localization and detection tasks by using radio images features. Specifically, the CSI measurements from different communication channels are used to obtain radio images from which visual features are extracted. The CSI has also been used for person identification, and existing works are mostly based on biometric signatures. The authors in [22] propose a passive method based on an accurate gait analysis obtained by fusing CSI and footstep sound measurements. Also in [23], gait patterns are captured to recognize people. The authors characterize walking patterns by using spectrograms from CSI. Even in [24]

, the user-authentication is obtained by leveraging on discriminative features extracted from CSI measurements of prevalent WiFi signals captured during daily activities, including walking. Again, in

[25]

, the authors exploit structural biometric features. They perform statistical analysis of the received Channel Frequency Response (i.e., the frequency-domain CSI) amplitude and phase. Differently, in

[26], to achieve user-authentication in secure environments, the presence of a spoofer is detected passively by examining the temporal correlation of CSI measurements. In this case, the authors remark the expected better results compared with existing approaches based on RSSI.

In current literature, Wi-Fi based person Re-ID works exploiting CSI measurements do not exist. Existing methods are mostly based on visual information extracted by processing images or video sequences [27, 28, 29, 30]. The publicly available datasets used by those approaches do not provide wireless signals data, therefore we cannot provide direct comparisons with state-of-the-art. We collected data by ourselves to build an evaluation benchmark for the proposed method and we reported the results.

Figure 1: Proposed method architecture. (A) Desktop PC, (B) 802.11n enabled commercial router.

3 Proposed Method

In this section the architecture of the proposed method is reported in detail beginning with hardware components followed by underlying technologies, CSI measurement, SNR estimation, and learning and re-identification strategies.

3.1 Method Architecture

Our proposed method architecture (Fig. 1) is designed to be simple, in fact, it is comprising of one desktop PC (Fig. 1A) and one 802.11n commercial router (Fig. 1B). The PC, which is the receiver, has one Intel Wi-Fi Link 5300 (IWL5300) Network Interface Card (NIC) with three antennas mounted on it. The router, which is the transmitter, has two antennas and the required 802.11n protocol enabled for transmissions. The architecture exploits the Multiple-Input and Multiple-Output (MIMO) technology, thus both transmitter and receiver make use of all of their available antennas for transmission and reception tasks. When a person is either between or passing through the transmitting and receiving locations, Wi-Fi signals reflection is affected by its presence with respect to the case in which there is an unobstructed path. The CSI measured on received signals has characteristics that could be used to extract features for training a model able to re-identify people. In the proposed method, given a set of person identities, we exploit CSI samples measured from OFDM subcarriers to extract SNR (i.e., the ratio of signal power to noise power) as feature for people Re-ID (Fig. 2

) by using a Multi-Layer Perceptron (MLP) neural network.

3.2 Orthogonal Frequency-Division Multiplexing

The orthogonal frequency-division multiplexing (OFDM) is a method used in modern wireless communications for digital data encoding on multiple carrier frequencies. It provides improvement in terms of communication performance by exploiting frequency diversity of the communication channels. In recent years, this technology is used in popular wireless networks, including IEEE and . Data is divided into multiple streams, each of them coded and modulated on different subcarriers on adjacent frequencies. Generally, overlapping adjacent channels can interfere with one another. To avoid this, in OFDM each subcarrier is orthogonal to each other in order to minimize interference during transmissions, thus the maximum power of each sub-carrier corresponds directly with the minimum power of each adjacent channel. For example, for the OFDM used by physical layer, a 20 or 40 MHz channel is composed of 56 or 114 subcarriers, respectively, such that each subcarrier can be used as a narrowband channel. This is why we chose to use the Channel State Information measured from OFDM subcarriers, it provides a finer granularity of the channel state useful in achieving higher accuracy for Re-ID in practice.

3.3 Channel State Information

In MIMO-based technology systems, multiple transmitting and receiving antennas are used to take advantage from multi-path propagation. Formally, the system can be modeled as:

(1)

where is the received signal vector, is the transmitted signal vector, and are the channel matrix and noise vector, respectively. The channel matrix , with antennas for transmission and antennas for reception, can be defined as:

(2)

where is the gain of each path between the transmitter and the receiver. Basically, there are two types of CSI, i.e., instantaneous and statistical ones. The former is estimated in fast fading systems where channel conditions vary rapidly during transmission. On the contrary, the latter can be only estimated in slow fading systems. Theoretically, if H is known, considering the channel estimation errors the instantaneous CSI can be modeled as:

(3)

where is the channel estimation, is the covariance matrix of the estimation error. Since the conditions of channel H vary, the instantaneous CSI is estimated on short-time basis. Channel matrix H is estimated by combining knowledge of both transmitted and received signals. Given a sequence , where each vector is transmitted over the channel as:

(4)

by combining (i.e., received signals) with transmitted signals and noise matrices, the total signalling becomes:

(5)

and the CSI can be recovered from the knowledge of and .

3.4 Signal Noise Ratio Estimation

In wireless transmissions, radio signals propagation paths and characteristics change accordingly the environment and objects encountered before arriving to the receiver. For CSI measurement, the network interface card used in the proposed method is the aforementioned IWL5300. Since it was designed for commercial use, custom firmware and driver versions of the ones used by [31] were required. Specifically, we measure the channel state in communications between the Access Point (i.e., the router) and the NIC. The measured CSI, consisting of complex-valued and high-dimensional channel matrices for 30 subcarriers, makes difficult achieving the person Re-ID task. Therefore, given a person identity , the channel state of N packets from is measured. Then, obtained the CSI samples, we extract the Signal Noise Ratio per packet as a k-dimensional vector defined as:

(6)

where is the number of subcarriers within each CSI sample, and is the K-lenght all-one SNR vector. Because the wireless devices are equipped with multiple antennas, the SNR is computed from each communication channel state between the transmitting antenna and the receiving antenna . In this way, the high-dimension complex-valued person biometrics within the measured CSI are mapped into the SNR, therefore the feature dimension is reduced. For example, Fig. 2 shows the SNR, from different packets, related to empty path between transmitting and receiving locations (Fig. 2 a) or person identity (Fig. 2 b,c,d).

Figure 2: SNR examples with 3 receiving antennas.

3.5 Multi-Layer Perceptron

Both learning and re-identification steps are based on the use of a Multi-Layer Perceptron (MLP) [32] network consisting of two hidden layers with LeakyReLU [33]

as activation function. The latter has been chosen because the experiments shown that it does not suffer from "dying ReLU"

[34]

problem and speeds up the training. The network loss function is the Cross-Entropy one, and the activation function for the last layer is the Soft-max one. The optimization algorithm used for the network is Adam

[35]

, because empirical results demonstrate achieving of good results faster. Moreover, activations in hidden layers are normalized by using Batch Normalization

[36] to improve accuracy. For each person identity , the input of the network are the SNR vector estimated from packets related to . During the learning stage, for each identity more packets are used to learn its fingerprint, whereas during testing stage even only one packet can be used for the Re-ID. The latter shows that our method is suitable for real-time applications.

4 Experiments

In this section the performed experiments are reported. In detail, the datasets and the obtained results are carefully discussed.

4.1 Datasets

To the best of our knowledge there are no datasets containing Wi-Fi signals concerning person identities, therefore we acquired our own datasets to evaluate the proposed method performance. To collect the data, the following acquisition protocol has been used. A total of 50 person identities have been acquired in 4 different conditions: standing still facing the router, standing still not facing the router, passing through the PC and the router from left to right and, finally, passing through the PC and the router from right to left. These acquisitions have been organized in two datasets, i.e., the standing and the walking ones. Three different acquisitions per subject have been collected for each of aforementioned conditions. This allows the model to capture all changes in both standing and walking patterns for each subject. Moreover, each acquisition lasts 3 seconds with a maximum number of 200 packets transmitted and received. Both the duration of acquisitions and the number of packets to use have been found empirically during preliminary tests. In addition, the acquisition of empty path between transmitting and receiving locations has been performed in order to reduce false positives. Finally, data has been collected in normal rooms, i.e., without using any shielding mechanism for interference introduced by other devices such as smartphones and other Wi-Fi devices. This kind of noisy acquisitions allowed us to test the robustness of the proposed approach for real applications.

4.2 Results

In Table 1, the results obtained by using both the two datasets, considering different packets number per identity, are shown. We used a minimum of 10 and a maximum of 200 packets. This because using less of 10 packets led to bad performance, while using a number of packets higher than 200 does not really increases the model accuracy. Concerning subcarriers, we have used all the 30 subcarriers provided by the network interface. Obviously, better results have been obtain by using the maximum number of packets available, leading to obtain a validation accuracy of for the standing dataset and for the walking one. By using only 10 packets, instead, we obtained a validation accuracy of for the standing dataset and for the walking one.

No. of Packets No. of subcarriers Dataset Training Accuracy Validation Accuracy Rank 1
10 30 Walking 98.30% 82.22% 82%
10 30 Standing 98.13% 93.57% 93%
50 30 Walking 93.42% 88.28% 88%
50 30 Standing 96.56% 96.43% 96%
100 30 Walking 94.71% 88.80% 88%
100 30 Standing 96.98% 96.5% 96%
200 30 Walking 96.79% 91.11% 91%
200 30 Standing 98.07% 97.86% 97%
Table 1: Table showing accuracies obtained with respect to the number of packets and the dataset used.

In general, the experiments performed on the standing dataset gives better results. This may be due to the fact that the moving subjects in the walking dataset could suffer of the Doppler shift. This aspect will be further investigated in order to increase the performance of our method. By looking at our Rank 1 best value, i.e. , we outperform most of the works at the current state-of-the-art which use visual features, remarking the goodness of the proposed approach. In Figure 3, the Cumulative Match Curve (CMC) for the different packets number are depicted. As it is possible to notice, our method achieves rapidly a match score in the first 10 ranks, which is quite impressive considering the proposed approach.

5 Conclusion

In this paper, a Wi-Fi passive re-identification system is proposed. The proposed method uses a commercial Wi-Fi network interface card to extract CSI from the received signals. A Multilayer Perceptron network trained with SNR estimated from CSI measurements is used to perform the person re-identification task. To the best of our knowledge there are no publicly available datasets concerning the re-identification through Wi-Fi signals, therefore we created two datasets comprising of 50 standing and walking human identities between transmitting and receiving locations. In our experiments, we have obtained a rank 1 accuracy of

which is a very promising result.

Figure 3: CMC curves for re-identification by using a) 10 packets, b) 50 packets, c) 100 packets, and d) 200 packets.

References

  • [1] K. Ali, A. X. Liu, W. Wang, and M. Shahzad. Recognizing keystrokes using wifi devices. IEEE Journal on Selected Areas in Communications, 35(5):1175–1190, 2017.
  • [2] X. Zheng, J. Wang, L. Shangguan, Z. Zhou, and Y. Liu. Design and implementation of a csi-based ubiquitous smoking detection system. IEEE/ACM Transactions on Networking, 25(6):3781–3793, 2017.
  • [3] Y. Gu, Y. Zhang, J. Li, Y. Ji, X. An, and F. Ren. Sleepy: Wireless channel data driven sleep monitoring via commodity wifi devices. IEEE Transactions on Big Data, pages 1–1, 2018.
  • [4] Belal Korany, Chitra R. Karanam, Hong Cai, and Yasamin Mostofi. Xmodal-id: Using wifi for through-wall person identification from candidate video footage. In The 25th Annual International Conference on Mobile Computing and Networking, MobiCom ’19, pages 36:1–36:15, 2019.
  • [5] Z. Chen, L. Zhang, C. Jiang, Z. Cao, and W. Cui. Wifi csi based passive human activity recognition using attention based blstm. IEEE Transactions on Mobile Computing, pages 1–1, 2018.
  • [6] Daniel Halperin, Wenjun Hu, Anmol Sheth, and David Wetherall. Tool release: Gathering 802.11n traces with channel state information. SIGCOMM Computer Communication Review, 41(1):53–53, 2011.
  • [7] Samer Mohammed, Allou Samé, Latifa Oukhellou, Kyoungchul Kong, Weiguang Huo, and Yacine Amirat. Recognition of gait cycle phases using wearable sensors. Robotics and Autonomous Systems, 75:50 – 59, 2016.
  • [8] G. Li, T. Liu, and J. Yi. Wearable sensor system for detecting gait parameters of abnormal gaits: A feasibility study. IEEE Sensors Journal, 18(10):4234–4241, 2018.
  • [9] Y. Wang, K. Wu, and L. M. Ni. Wifall: Device-free fall detection by wireless networks. IEEE Transactions on Mobile Computing, 16(2):581–594, 2017.
  • [10] Z. Fu, J. Xu, Z. Zhu, A. X. Liu, and X. Sun. Writing in the air with wifi signals for virtual reality devices. IEEE Transactions on Mobile Computing, 18(2):473–484, 2019.
  • [11] E. Ben Hamida and G. Chelius. Investigating the impact of human activity on the performance of wireless networks — an experimental approach. In IEEE International Symposium on "A World of Wireless, Mobile and Multimedia Networks" (WoWMoM), pages 1–8, 2010.
  • [12] S. Cai, W. Liao, C. Luo, M. Li, X. Huang, and P. Li. Cril: An efficient online adaptive indoor localization system. IEEE Transactions on Vehicular Technology, 66(5):4148–4160, 2017.
  • [13] Y. Fu, P. Chen, S. Yang, and J. Tang.

    An indoor localization algorithm based on continuous feature scaling and outlier deleting.

    IEEE Internet of Things Journal, 5(2):1108–1115, 2018.
  • [14] V. Bianchi, P. Ciampolini, and I. De Munari. Rssi-based indoor localization and identification for zigbee wireless sensor networks in smart homes. IEEE Transactions on Instrumentation and Measurement, 68(2):566–575, 2019.
  • [15] S. Kianoush, S. Savazzi, F. Vicentini, V. Rampa, and M. Giussani. Device-free rf human body fall detection and localization in industrial workplaces. IEEE Internet of Things Journal, 4(2):351–362, 2017.
  • [16] A. Booranawong, N. Jindapetch, and H. Saito. A system for detection and tracking of human movements using rssi signals. IEEE Sensors Journal, 18(6):2531–2544, 2018.
  • [17] H. Abdelnasser, M. Youssef, and K. A. Harras. Wigest: A ubiquitous wifi-based gesture recognition system. In IEEE Conference on Computer Communications (INFOCOM), pages 1472–1480, 2015.
  • [18] Jiang Xiao, Zimu Zhou, Youwen Yi, and Lionel M. Ni. A survey on wireless indoor localization from the device perspective. ACM Computing Survey, 49(2):25:1–25:31, 2016.
  • [19] D. Zhang, H. Wang, and D. Wu. Toward centimeter-scale human activity sensing with wi-fi signals. Computer, 50(1):48–57, 2017.
  • [20] F.A. Jenkins and H.E. White. Fundamentals of Optics. McGraw-Hill Science Engineering, 1957.
  • [21] Q. Gao, J. Wang, X. Ma, X. Feng, and H. Wang. Csi-based device-free wireless localization and activity recognition using radio image features. IEEE Transactions on Vehicular Technology, 66(11):10346–10356, 2017.
  • [22] Yuanying Chen, Wei Dong, Yi Gao, Xue Liu, and Tao Gu. Rapid: A multimodal and device-free approach using noise estimation for robust person identification. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 1(3):41:1–41:27, 2017.
  • [23] Wei Wang, Alex X. Liu, and Muhammad Shahzad. Gait recognition using wifi signals. In Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing, UbiComp ’16, pages 363–373, 2016.
  • [24] Cong Shi, Jian Liu, Hongbo Liu, and Yingying Chen. Smart user authentication through actuation of daily activities leveraging wifi-enabled iot. In Proceedings of the 18th ACM International Symposium on Mobile Ad Hoc Networking and Computing, Mobihoc ’17, pages 5:1–5:10, 2017.
  • [25] Q. Xu, Y. Chen, B. Wang, and K. J. R. Liu. Radio biometrics: Human recognition through a wall. IEEE Transactions on Information Forensics and Security, 12(5):1141–1155, 2017.
  • [26] H. Liu, Y. Wang, J. Liu, J. Yang, Y. Chen, and H. V. Poor. Authenticating users through fine-grained channel information. IEEE Transactions on Mobile Computing, 17(2):251–264, 2018.
  • [27] Sanping Zhou, Jinjun Wang, Deyu Meng, Xiaomeng Xin, Yubing Li, Yihong Gong, and Nanning Zheng. Deep self-paced learning for person re-identification. Pattern Recognition, 76:739 – 751, 2018.
  • [28] S. Zhou, J. Wang, J. Wang, Y. Gong, and N. Zheng.

    Point to set similarity based deep feature learning for person re-identification.

    In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5028–5037, 2017.
  • [29] D. Avola, M. Cascio, L. Cinque, A. Fagioli, G. L. Foresti and C. Massaroni. Master and rookie networks for person re-identification. In Computer Analysis of Images and Patterns (CAIP), pages 470–479, 2019.
  • [30] Xiang Li, Ancong Wu, and Wei-Shi Zheng. Adversarial open-world person re-identification. In Computer Vision – ECCV 2018, pages 287–303, 2018.
  • [31] Daniel Halperin, Wenjun Hu, Anmol Sheth, and David Wetherall. Tool release: Gathering 802.11n traces with channel state information. SIGCOMM Comput. Commun. Rev., 41(1):53–53, 2011.
  • [32] Christopher M. Bishop. Neural Networks for Pattern Recognition. Oxford University Press, Inc., 1995.
  • [33] Andrew L. Maas, Awni Y. Hannun, and Andrew Y. Ng. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, volume 30, page 3, 2013.
  • [34] Lu Lu, Yeonjong Shin, Yanhui Su, and George Em Karniadakis. Dying relu and initialization: Theory and numerical examples, 2019.
  • [35] L. J. Ba D. P. Kingma. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015.
  • [36] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift, 2015.