Abstract
Recently, person identification has been a prerequisite in many applications of the Internet of Things. As a new biometric identification technology, gait recognition has a wide application prospect with the advantages of long-distance recognition and difficulty to forge. However, the existing gait recognition methods have some problems, such as complex algorithm calculation, high user participation, and large equipment overhead. In this paper, we propose RF-Gait, a method that identifies a person through unobtrusive gait perception with COTS RFID. The key insight is that wireless signal fluctuation can be exploited to distinguish each person’s unique gait behavior. To this end, we first collect and preprocess the gait-induced data composed of multiple RFID tags. Furthermore, multivariate variational mode decomposition is utilized to extract the intrinsic features in the spatial multichannels cooperatively. By developing a support vector machine model, we identify a person via the intrinsic walking pattern. Finally, extensive experiments show that our method can identify a person with an average accuracy of 96.3% from a group of twenty persons in a complex indoor environment.
1. Introduction
In the era of the Internet of Things, human-computer interaction has become crucial to integrating the physical world and the information world. Person identification provides a guarantee for the security of human-computer interaction. Gait, as an emerging biometric feature, refers to the human characteristic activity of moving the body through the interactive action of two feet. For human identification, gait has the following two advantages: on the one hand, from the perspective of medicine, the length of leg bones, the strength of muscles, the height of center of gravity, and the sensitivity of motor nerves determine the uniqueness and stability of gait, so it is difficult to be imitated by others in a short time; on the other hand, gait recognition is not strict on distance and protects privacy.
Some kinds of methods have been developed to promote the recognition performance of gait, and among them, video-based and radio-based methods have contributed significantly. Typical video image methods extract the contour image of user motion for gait feature extraction and recognition. However, it is worth noting that the video image methods need complex algorithm, large amount of calculation, and require expensive computing equipment for feature extraction and comparison [1]. Therefore, researchers consider to use wireless signal for gait recognition. As we know, when a person walks in the sensing area, the wireless signal will be reflected and diffracted [2], and furthermore, the mapping relationship between each person’s gait and the signal fluctuation is unique. WiFiU [3] is the first method that uses COTS WiFi devices to fulfill gait recognition. Channel State Information (CSI) is transformed to the spectrogram in the time-frequency domain, and the spectrogram signature is extracted for gait recognition. GaitSense [4] proposes a WiFi CSI-based novel method that can extract characterized gait patterns and is termed as gait body-coordinate velocity profile. In [5], the spatial gain provided by a RFID tag array is used to resist the different effects of the same person when the person is walking. However, indoor ambient noise may limit the recognition performance of these methods. Due to the fact that the subcarriers of WiFi CSI all use the same propagation paths, there are no discernible differences among them. Additionally, RFID tags’ spatial gain is not fully utilized for gait recognition. In some case, FMCW radar has been used for extract fine-grained gait features in the noisy wireless signal. However, this requires the use of customized equipment, which is costly and does not meet actual needs.
To realize accurate human gait recognition via commercial devices, we need to overcome two critical challenges. The first challenge is the RF signal composition between human gait and other indoor static objects which is highly nonlinear. In order to extract fine-grained gait signals for person identification, we adopt MVMD to decompose the received RF signal and acquire the fine-grained gait signal. The second challenge is that there are differences between each person’s different walking signals. In order to eliminate the difference and accurately identify the person’s identity, we construct the feature set of human gait from the time domain and frequency domain of RF signals and train the SVM machine learning model for human identification. Moreover, we propose RF-Gait, which use COTS RFID for gait recognition and the cheap RFID tags can provide rich spatial diversity for fine-grained gait feature extraction. The key insight of this paper is to adopt multivariate variational mode decomposition (MVMD) [6] to extract intrinsic features from noisy RFID measurements so as to further improve the accuracy of gait recognition. Our main contributions can be summarized as the following aspects: (1)We propose a new fine-grained gait feature extraction method with COTS RFID. Owing to device defects, the constant change of indoor environment, and diversity in personnel walking, these noises are nonstationary and nonlinearly mixed with the gait based signal fluctuation. As a result, the proposed method decomposes the signal and aligns the decomposed signals with common frequencies across modes(2)We implement RF-Gait and comprehensively evaluate the performance in a complex office building under various conditions. Extensive experiments show that our method can identify a person with an average accuracy of from a group of ten persons in a real world scenario where people can freely walk and the surrounding environment will change
The reminder of this paper is organized as follows. We first explain some background knowledge and present the system overview in Section 2. The details of our method design are described in Section 3. The implementation and evaluation of the proposed method are presented in Section 4. We review the related works in Section 5. Finally, Section 6 concludes this paper and provides some suggestions for future.
2. Preliminaries and System Overview
2.1. Working Principle of RFID
UHF RFID usually consists of three parts, including reader, antenna, and tag. Figure 1 shows the schematic diagram of RFID backscatter communication. The reader transmits radio frequency continuous wave signal through the antenna. After the tag is activated by the signal sent by the reader, it will return its own information to the reader through the backscatter link. General commercial readers, such as Impinj Speedwayr420, can not only obtain the ID of the tag but also give the indicators of the reflected signal of the tag: signal strength (RSSI), signal phase, and signal Doppler frequency shift. Since the Doppler frequency shift provided by the reader is very noisy [7], researchers usually use two indicators, RSSI and phase for RFID sensing.
In free space, the signal needs to go through two transmissions between the reader and the tag: the forward link from the reader to the tag and the backscatter link from the tag to the reader. When the distance between the reader antenna and the tag is , the tag RSSI and phase received by the reader can be expressed as [8, 9] where and is determined by environment and equipment deployment conditions. is the signal wavelength, and and represents the phase offset brought by the reader and the tag itself, respectively.
2.2. Feasible Study
The leading attraction of the human gait as a biological feature is contactless, and more importantly, gait is a unique feature of individuals. Murray et al. discovered that the standing, swinging, and gait-related body movements of the same participant are surprisingly similar, and walking has a strong periodicity [10]. What is more, the above mentioned gait characteristics of different participants present obvious individual differences. Specifically, as shown in Figure 2, when the reader antenna and a tag array (In the experiment part, we discussed the setting of array size and finally set it to 4 rows and 2 columns.) are placed on both sides of the personnel walking route, the human body can be modeled as a conducting cylinder [11]. Meanwhile, the wavelength of UHF RFID is about , which is roughly close to the width of the human body. Therefore, the person is able to reflect and diffract the RF signal when he passes between the reader antenna and the tags, and there is a one-to-one mapping relationship between the impact of each person’s gait on the RF signal. As a result, on the far right of Figure 2, the red sine curve represents the influence of personnel walking on RF signal which is the component we are trying to extract, the green line represents the influence of indoor static environment on RF signal, and the actually received RF signal which is red line can be ideally expressed as [12] where , , and represent the measured RSSI and phase of the tag, corresponding to the polar angle and polar radius in the I-Q plane, respectively.
Moreover when a person walks in the RFID sensing area, the received signal can be decomposed into where the received signal is divided into static and dynamic components in Equation (4). represents the superposition of all static propagation paths, and represents the superposition of all dynamic propagation paths. is the number of dynamic propagation paths, and is the length of the dynamic propagation path. As shown in Figure 3, we find that the periodic dynamic component of the received signal, which including both the phase and RSSI measurements, can be utilized to exploit the unique gait pattern of different persons.
(a) RSSI profiles when a person crosses the reader-tag area
(b) Phase profiles when a person crosses the reader-tag area
2.3. System Overview
The framework of RF-Gait is shown in Figure 4. To perform person identification based on gait patterns, we deploy several RFID tags to form a tag array for comprehensive gait perception. For the received RF signals from the tag array, the main processing steps consist of three main components: (1) Data preprocessing: with the RF signals received from the tag array, RF-Gait first filters data with interpolation and smoothing and then analyzes for walking detection and direction. (2) Feature extraction: based on the processed signal series, RF-Gait extracts feature vectors via MVMD algorithm and intrinsic mode sifting criterion. (3) Person identification: our method finally identifies the person who is walking in sensing area according to the gait features stored in the trained model.
3. Method Design
In this section, we detail the method design of RF-Gait including data collection, data preprocessing, feature extraction, and person identification.
3.1. Data Collection
The first step of our method is to collect the RFID signal measurements. We deploy the reader antenna and multiple tags on both sides of the walking path. When a person passes between the reader antenna and tags in different ways, we collect the raw RSSI measurements and phase measurements .
3.2. Preprocessing
3.2.1. Data Sanitizing
The commercial reader uses the slot-ALOHA mechanism to read the channel information, resulting in the nonuniform distribution of the received data in the time domain. In order to facilitate the subsequent processing, we carry on the linear interpolation with sampling points per second to the measurements. Furthermore, in order to solve equipment noise and measurement errors, we consider using the Savitzky-Golay filter [13] for data smoothing. The key insight of the filter is to perform weighted filtering in the predefined window. Moreover, it can more effectively retain the change information of the measurements while smoothing the data. The raw RSSI and phase measurements and the data after sanitizing are shown in Figure 5, and the sanitized signal composed of sanitized RSSI and phase can be denoted as
(a) Raw RSSI
(b) Sanitized RSSI
(c) Raw phase
(d) Sanitized phase
3.2.2. Walking Detection
To better extract gait features from walking patterns, we need to perform motion detection and segment the time series of the sanitized measurements. Therefore, in this section, we first analyze the short cumulative energy of signal series as follows: where denotes the amplitude value, and denotes the length of the sliding window.
As a result, we calculate the signal energy in each window and compare it with a predefined moving detection threshold. As shown in Figure 6, we utilize this scheme to judge the starting and ending points of the movement.
The next step is aimed at distinguishing the walking direction. For this problem, because different directions have different effects on labels, we can consider using the RSSI difference of two columns of labels to solve it. The difference can be represented as where and mean the tags’ RSSI of two different columns in the same row. When , this means that the person walks from the tag in the right column to the tag in the left column, and vice versa.
3.3. Feature Extraction
3.3.1. Multivariate Variational Mode Decomposition
Since the dynamic component of the signal is nonstationary and superimposes with the static component in a nonlinear way, thus we resort to an adaptive method to extract the intrinsic modes from the signals consisting of multichannel tag measurements automatically. The basic idea of MVMD is to decompose the original multichannel signals into several simple and high quality signals (i.e., intrinsic mode functions which are termed as IMFs). As a result, the preprocessing multichannel signals are decomposed into predefined number IMFs and the residual : where .
The goal is to extract simple oscillatory mode IMFs involved in the signals , which are the informative information for realizing person identification. The result cost function of our problem Equation (8) is given by
In the above proposed variational model of decomposition task, we aim to minimize the sum of bandwidths of all IMFs. In order to resolve the multiobjective optimization problem (9), we adopt the Alternating Direction Method of Multiplier- (ADMM-) based optimization strategy [6]. First of all, we transform the original constrained optimization problem into the form of unconstrained optimization by Lagrange multiplier. In the following, two alternate update steps are given for the mode and the center frequency .
Step 1. Updating mode : The other variables are fixed, and can be updated by solving the problem: where and are the parameters of Lagrange multiplier method.
Step 2. Updating center frequency : In this step, is updated by solving the following problem:
The detailed optimization process is given in Algorithm 1.
| ||||||||||||||||
After MVMD processing, we get IMFs, but there is only a small number involving rich gait information. Hereinafter, we rank the IMFs’ quality of each tag via the correlation coefficients, and for the tag, the correlation of the IMF is defined as
According to the above Equation (12), we sift IMFs of each tag with the highest correlation .
3.3.2. Feature Construction
Considering the real-time requirements for person identification, we choose a lightweight and efficient scheme to extract features based on statistics theory. Specifically, for the above selected set of IMFs , we leverage features to portray the time and frequency profile of the walking pattern. The features are the (1) mean, (2) standard deviation, (3) skewness, (4) kurtosis, (5) form factor, and (6) crest factor.
In this way, we can finally form the feature matrix as where contains time domain features and frequency features.
To visualize the learned features, we use t-SNE [14] to project them into a two-dimensional feature space. Figure 7 illustrates the gait features of 10 subjects, which have excellent discrimination capability.
3.4. Person Identification
In this step, we introduce how to perform person identification based on extracted features. Due to the multiclassification problem and the need to ensure higher recognition accuracy in a short period of time, we choose support vector machine (SVM) [15] for identification. Specifically, RF-Gait feeds the extracted IMFs’ feature vectors into an SVM model with a radial basis function kernel function in the model training step. Then, when a person is walking in the sensing area, RF-Gait determines the identity of this person via the trained model.
4. Performance Evaluation
4.1. Experiment Settings
In this paper, we perform extended experiments to verify the proposed method RF-Gait in a typical indoor environment. The experiment contains an Impinj Speedway R420 reader with a circular polarization antenna ( gain) and Impinj-H47 tags that form a tag array. The reader operates at the stable frequency of , and the tags are deployed to form a uniform linear tag array which is shown in Figure 8. The tag array is away from the reader antenna, and the height of the reader antenna is . The plan of the experimental scene is shown in Figure 9. As shown in Figure 9, multiple places in the scene are used to collect walking data. We recruit 10-20 volunteers, everyone with walking times on each place. From the data set, of the data is randomly selected for model training, and the remaining is used for testing our proposed method RF-Gait.
4.2. System Performance Analysis
4.2.1. Walking Detection Accuracy
In this section, we first evaluate the accuracy of the walking detection scheme. To verify the effectiveness of our scheme, we asked volunteers to walk normally, intermittently, and pause in the sensing area and we used the camera to record the results as the ground truth. As shown in Table 1, our scheme achieves a good result.
4.2.2. Time Required for Verification
Many gait recognition methods based on video image processing need to run complex image processing algorithms. The method proposed in this paper is very lightweight, and the time required for authentication is very short. As shown in Table 2, the average time required for signal processing, feature extraction, and comparison to authenticate a person is only 0. 357 s.
4.2.3. Person Identification Accuracy
To evaluate the accuracy of RF-Gait, we utilize a confusion matrix to describe the overall performance of RF-Gait as shown in Figure 10. The diagonal elements of the matrix represent the probability of correct recognition using the extracted feature matrix, and other elements represent the probability of the wrong classification into others. Figure 10 shows that our system achieves an average accuracy of in the typical complex indoor environment.
Furthermore, we compare RF-Gait with variational mode decomposition and two state-of-art gait-based person identification methods [16, 17] which are empirical mode decomposition-based methods in Figure 11. The accuracy of our proposed method is higher than the other three methods. Our method is higher than the traditional VMD scheme, higher than the literature [16], and higher than the literature [17], respectively.
4.2.4. The Impact of the Horizontal Distance between the Reader Antenna and Tag Array
In our proposed method, the width of an entrance determines the distance between the reader and the tag array. We evaluate the effect of entrance width on the system performance. The distance varies from to at intervals of . As shown in Figure 12, we find that the recognition accuracy tends to get better as the distance from the entrance becomes larger. Through literature research, we believe that the dynamic path, that is, personnel walking, has a more significant influence than the static RF signal component [18].
4.2.5. The Impact of Walking Difference
In this part, we mainly evaluate the influence of walking speed and user lateral positions on the recognition effect. First, we evaluate the effects of different walking speed by letting the same 10 subjects walk through the sensing area at different velocities. As shown in Figure 13, except for the fast walking which achieves a decreased accuracy of , RF-Gait maintains high recognition accuracy when walking slowly and normally. Second, we set three kinds of lateral positions, one is close to the tags, one is close to the reader, and the other is walking in the middle. Then, we evaluate the influence of this condition on the recognition accuracy. As shown in Figure 14, when people walk in the middle, the recognition effect is the best. We speculate that when close to the reader and tags, the normal reading of the RF signal will be affected [19], so the recognition effect will be reduced.
4.2.6. The Impact of Different Clothes
In the actual scene, the same person will wear different clothes, so we need to evaluate the impact of different clothes on RF gait. As shown in Figure 15, there is no significant impact on the recognition effect when the clothes of walkers change. We believe that this is because normal clothing materials have little effect on the RF signal of UHF band [20].
4.2.7. The Impact of Different Classifiers
In order to analyze the recognition accuracy of different classifiers on selected features, we select four commonly used multiobjective classifiers, support vector machine with RBF (SVM), decision tree (DT), sparse representation classification (SRC), and naive Bayes (NB). As shown in Figure 16, we find that the result of SVM is the best, so we choose it as the classifier in our paper.
4.2.8. The Impact of Tags’ Number
As more tags can acquire richer spatial gain for person identification while increasing the person identification delay, we study the impact of the number of tags on the performance. In the scenario shown as Figure 17, we change the number of tags from to and show the results in Figure 17. We find that with the increase of the number of tags, the recognition accuracy grows significantly, but the growth is not obvious after the number is . Therefore, we make a trade-off between recognition accuracy and system delay and select eight tags to collect gait information for person identification.
4.2.9. The Impact of Training Set Size
In practical application, the less the number of training samples, the less the cost of actual deployment. In order to investigate the influence of the number of training samples on the results, this paper changes the number of training samples from to and uses the remaining samples as the test. The certification results are shown in Figure 18.
5. Related Work
This section reviews the related literature in RFID-based sensing and gait recognition techniques.
5.1. RFID-Based Sensing
As a mature automatic identification technology, RFID technology is widely used in industrial manufacturing [21], warehousing [22], and logistics [23]. RFID-based sensing technology utilizes signal characteristics reflected back from RFID tags for indoor contextual sensing. The emerging RFID-based sensing applications are widely applied to user authentication [12, 24, 25], indoor localization [26, 27], activity recognition [7, 28], and so on [29]. RF-Mehndi [25] is an RFID-based user authentication system. The authors find that the coupling effect of a tag array is distinctive when different users touch the tags so as to achieve biometric acquisition. 3DLRA [27] leverages RFID tgas to develop a 3D indoor localization system and analyzes the variation characteristic of signal indicators using deep learning. In [28], the authors quantify the correlation between RF phase values and human activities by modeling the signal reflection of RFID tags in contact-free scenarios. Furthermore, RFID is also be used for material sensing, vibration sensing, and so on. In [29], the impedance-related phase change is utilized for material sensing and finally the authors can detect the category of the material.
5.2. Gait Recognition
Existing gait recognition methods mainly focus on three categories: methods based on video image processing [30], methods based on sensors [31, 32], and methods based on radios [33–35]. Among them, the method based on video image processing performs gait feature extraction and identification by extracting the contour image of the user’s movement. In [31], the authors use Gait Energy Image (GEI) as a template for human identification by gait. The sensor-based method extracts the user’s acceleration and other information when walking through various sensors carried by the user and analyzes the user’s gait behavior. In [36], inertial measurement units (IMU) are consider for recognizing gait and the authors design a deep convolutional neural network to extract discriminative gait features. The radio-based methods identify human gait by analyzing the wireless signal and establishing the mapping relationship with human walking. In [35], the authors apply weighted multidimensional dynamic time warping to compute the similarity of two walking profiles which are collected from the RFID tags.
6. Summary and Future Work
This paper introduces a walking-induced method RF-Gait which is capable of identifying persons using COTS RFID. RF-Gait employs multiple spatially distributed tags to obtain fine-grained profiles of the human gait and enables human identification via the MVMD algorithm and SVM-based identification model. Experimental results show that the average identification rate of this method can reach , and it has good robustness and stability. However, this paper only considers the gait perception of a single person, which is not well performing in multiperson situations. An in-depth study is needed to identify the gaits of more than one person at a time.
Data Availability
The data used to support the findings of this study are available from the corresponding author upon request.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
This work was financially supported by the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant no. XDC02040300).