PERSON IDENTIFICATION AND IMPOSTER DETECTION USING FOOTFALL GENERATED SEISMIC SIGNALS

Abstract

A smart device, biometric authentication system and a corresponding method thereof for person identification and imposter detection has been disclosed. The method comprises detection and extraction of seismic signals generated from corresponding footfalls, by means of unsupervised learning based detection and extraction module (USLEEM) and detection and identification of imposter and/or registered users respectively by means of an identification module.

Claims

1. A method for person identification and imposter detection, said method comprising steps of: detection and extraction of seismic signals generated from corresponding footfalls, by means of unsupervised learning based detection and extraction module (USLEEM); and detection and identification of imposter and/or registered users respectively by means of an identification module.

2. The method as claimed in claim 1, wherein the said step of detection and extraction of seismic signals generated from corresponding footfalls comprising at least one training phase including: splitting, each of said seismic signals, into N equal segments; extracting, feature vectors (FE-I) corresponding to time domain and frequency domain features from each of said segment of the N segments; clustering each of the said feature vectors into a clustered event so as to form at least one trained model of USLEEM; storing of the said trained model.

3. The method as claimed in claim 1, wherein said step of detection and extraction of seismic signals generated from corresponding footfalls further comprising at least one live phase including: splitting, each of said seismic signals, into N equal segments; extracting, feature vectors (FE-I) corresponding to time domain and frequency domain features from each of said segment of the N segments; and detecting footfalls from the extracted featured vectors (FE-I) thereafter feeding it to the trained model.

4. The method as claimed in claim 3, wherein the step of detection and identification of imposter and/or registered users respectively by means of an identification module, said method steps including: detection of at least one imposter, by means of One—Class machine learning model when feature vectors (FE-II) obtained from the extracted footfalls as claimed in claim 3 are fed to the One-Class model; and identification of registered users, from feature vectors (FE-II) obtained from the extracted footfalls by means of multi-class machine learning model.

5. The method as claimed in claim 1, wherein Gaussian Mixture Model, GMM, clustering facilitates clustering each of the said feature vectors into a clustered event.

6. The method as claimed in claim 5, wherein each cluster is parameterized by a set of p, p, and I and the method further comprises classifying said clustered event into a footfall event and noise by $\mathrm{Class}=\{\begin{array}{c}{C}_1.fwdarw.\mathrm{Event},C_2.fwdarw.\mathrm{Noise}.Math.\text{:}.Math..Math..Math.{\Sigma }_{C_1}.Math.>.Math.{\Sigma }_{C_2}.Math.\\ C_1.fwdarw.\mathrm{Noise},C_2.fwdarw.\mathrm{Event}.Math.\text{:}.Math..Math..Math.{\Sigma }_{C_2}.Math.>.Math.{\Sigma }_{C_1}.Math.\end{array}$ where |Σ.sub.C.sub.k| is the determinant of the co-variance matrix of the kth clusters.

7. The method as claimed in claim 5, wherein during the live phase, each of the N segments of the footfall generated seismic signals are classified by, $\mathrm{Class}=\{\begin{array}{c}{C}_1.Math.\text{:}.Math..Math.p\left(C_1f_{\mathrm{test}}^{w_i}\right)>p\left(C_2f_{\mathrm{test}}^{w_i}\right)\\ C_2.Math.\text{:}.Math..Math.p\left(C_2f_{\mathrm{test}}^{w_i}\right)>p\left(C_1f_{\mathrm{test}}^{w_i}\right)\end{array}$ where f.sub.test.sup.wi is the feature vector of the i.sup.th segment of a test signal Signal.sub.test, p(C.sub.k|f.sub.test.sup.w.sup.i)(=ϕ.sub.k.Math.N(f.sub.test.sup.w.sup.i|μ.sub.C.sub.k,Σ.sub.C.sub.k) for k=1,2) is the probability that f.sub.test.sup.wi belongs to class C.sub.k, test signal Signal.sub.test is the seismic signal generated by an imposter or a registered user during the live phase and it is segmented into equal parts (w.sub.i.sub.test).

8. A smart device for person identification and imposter detection configured to perform the method steps as claimed in claim 1, said smart device comprising: at least one sensing module configured to detect a plurality of seismic signals generated from corresponding footfall of a person to generate a seismic event; an analog-to-digital converter module configured to convert detected analog seismic signals into digital signals, an event extraction module configured to: split a seismic signal, of the plurality of said seismic signals, into N equal segments, and extract, vectors corresponding to time domain and frequency domain features from each said segment of the N segments; cluster each of the said vectors into a clustered event; store said trained model; and an identification module configured to identify either an imposter or a registered person.

9. A biometric authentication system for person identification and imposter detection, said system comprising: an array of smart devices as claimed in claim 8 distributed over a pre-determined zone; and a central controller operatively communicable to said smart devices adapted to perform the method steps as claimed in comprising the steps of detection and extraction of seismic signals generated from corresponding footfalls, by means of unsupervised learning based detection and extraction module (USLEEM); and detection and identification of imposter and/or registered users respectively by means of an identification module.

Description

BRIEF DESCRIPTION OF ACCOMPANYING DRAWINGS

[0023] The above and other aspects, features and advantages of the embodiments of the present disclosure will be more apparent in the following description taken in conjunction with the accompanying drawings, in which:

[0024] FIG. 1 illustrates a system architecture for human identification and imposter according to an embodiment of present invention.

[0025] FIG. 2 illustrates overview of event detection and extraction technique USLEEM according to an embodiment of present invention.

[0026] FIG. 3 illustrates the event extraction process of USLEEM according to an embodiment of present invention.

[0027] FIG. 4 illustrates effect of rectangular and Gaussian window on the frequency content of the extracted footfall event according to exemplary implementations of an embodiment of present invention.

[0028] FIG. 5 illustrates performance comparison of event detection techniques according to the present invention.

[0029] FIG. 6 illustrates performance of SVM-RBF in person identification as different event extraction techniques (USLEEM, UREDT, and Adap-Th) are used according to one of the embodiment of present invention.

[0030] FIG. 7 illustrates learning curve of SVM-RBF obtained using USLEEM and Adap-Th as event extracted techniques according to one of the embodiment of present invention.

[0031] FIG. 8 illustrates performance of imposter detection using OC-SVM and SVDD in three different imposter detection scenarios according to one of the embodiment of present invention.

[0032] FIG. 9 illustrates ROC curve of a) SVDD and b) OC-SVM for imposter detection scenario 2 (refer Table. IV) obtained using dataset10 (number of registered users and imposters are 4 and 3).

[0033] FIGS. 10 and 11 illustrate the training phase and live phase working implementation of a method for person identification and imposter detection using unsupervised learning based detection and extraction module, USLEEM, according to an embodiment of the present invention.

[0034] FIG. 12 illustrate a smart device for person identification and imposter detection, according to an embodiment of the present invention.

[0035] Persons skilled in the art will appreciate that elements in the figures are illustrated for simplicity and clarity and may have not been drawn to scale. For example, the dimensions of some of the elements in the figure may be exaggerated relative to other elements to help to improve understanding of various exemplary embodiments of the present disclosure. Throughout the drawings, it should be noted that like reference numbers are used to depict the same or similar elements, features, and structures.

DETAILED DESCRIPTION

[0036] The following description with reference to the accompanying drawings is provided to assist in a comprehensive understanding of various embodiments of the present disclosure as defined by the claims and their equivalents. It includes various specific details to assist in that understanding but these are to be regarded as merely exemplary. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the various embodiments described herein can be made without departing from the scope and spirit of the present disclosure. In addition, descriptions of well-known functions and constructions may be omitted for clarity and conciseness.

[0037] The terms and words used in the following description and claims are not limited to the bibliographical meanings, but, are merely used by the inventor to enable a clear and consistent understanding of the present disclosure. Accordingly, it should be apparent to those skilled in the art that the following description of various embodiments of the present disclosure is provided for illustration purpose only and not for the purpose of limiting the present disclosure as defined by the appended claims and their equivalents.

[0038] It is to be understood that the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise. Thus, for example, reference to “a component surface” includes reference to one or more of such surfaces.

[0039] All terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which various embodiments belong. Further, the meaning of terms or words used in the specification and the claims should not be limited to the literal or commonly employed sense, but should be construed in accordance with the spirit of the disclosure to most properly describe the present disclosure.

[0040] The terminology used herein is for the purpose of describing particular various embodiments only and is not intended to be limiting of various embodiments. As used herein, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprises” and/or “comprising” used herein specify the presence of stated features, integers, steps, operations, members, components, and/or groups thereof, but do not preclude the presence or addition of one or more other features, integers, steps, operations, members, components, and/or groups thereof. Also, Expressions such as “at least one of,” when preceding a list of elements, modify the entire list of elements and do not modify the individual elements of the list.

[0041] The present disclosure will now be described more fully with reference to the accompanying drawings, in which various embodiments of the present disclosure are shown. The present disclosure may, however, be embodied in many different forms and should not be construed as being limited to the various embodiments set forth herein, rather, these various embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the present disclosure. Furthermore, a detailed description of other parts will not be provided not to make the present disclosure unclear. Like reference numerals in the drawings refer to like elements throughout.

[0042] The subject invention lies in a system for person identification and imposter detection by analyzing footfall generated seismic signals.

[0043] According to an embodiment of the present invention, FIG. 1 illustrates a biometric authentication system for person identification and imposter detection using unsupervised learning based detection and extraction technique, USLEEM, said system comprising an array of smart devices (101) distributed over a zone to record seismic signals corresponding to a footfall of a person, a central controller (102) communicably coupled to the array of smart devices. Each smart device comprises at least one sensing module (1011) configured to detect a plurality of seismic signals generated from corresponding footfall of a person to generate a seismic event, an analog-to-digital converter module (1012) configured to convert detected raw analog seismic signals into digital signals, an event extraction module (1013) configured to split a seismic signal, of the plurality of said seismic signals, into N equal segments and extract, vectors corresponding to time domain and frequency domain features from each said segment of the N segments, cluster each of the said vectors into a clustered event, and store said clustered event. The smart device further comprises an identification module (1014) configured to identify either an imposter or a registered person.

[0044] According to an embodiment of the present invention, all the raw data acquisition, event extraction, person identification and imposter detection/classification happens in the smart device, SD. Final result related to person identification and impostor detection from all the SDs is stored to the database situated in Central controller.

[0045] A footfall is a short-lived event embedded on the noise of a seismic signal. If the entire signal is segmented in smaller chunks, most of the chunks will be void of footfall events. So, processing each chunk of the signal for human identification or imposter detection will lead to wastage of power and resource. So, it is very important to have a robust event detection and extraction technique.

[0046] According to another embodiment of the present invention, FIG. 2 illustrates an event detection and extraction technique to obtain the portion of the signal containing footfall event, the said technique comprising of the phases: training phase to cluster samples into two classes including footfall event and noise; and live phase.

[0047] A seismic signal (containing human footsteps) is split in N equal segments of length 220 ms. Time and frequency domain related features, referring to Table I, are extracted from each segment. The complete feature vector is represented by f.sup.w.sup.i(=[f.sub.i.sup.w.sup.i, . . . , f.sub.5.sup.w.sup.i]), where f.sub.j.sup.w.sup.i is the j.sup.th feature of i.sup.th segment (i=1, . . . , N). Each feature vector f.sup.w.sup.i serves as a single sample for the clustering technique.

TABLE-US-00001 TABLE I Feature Extracted from the i.sup.th segment Segment's Features Statistical Features Spectral Features (Energy bins) f.sub.1.sup.w.sup.i f.sub.2.sup.w.sup.i f.sub.3.sup.w.sup.i f.sub.4.sup.w.sup.i f.sub.5.sup.w.sup.i std kurtosis 40-80 Hz 80-120 Hz 120-160 Hz

[0048] According to an exemplary implementation of the present invention, in the training phase, Gaussian Mixture Model, GMM is used to cluster the samples into two classes, footfall event and noise (absence of an event). Gaussian Mixture Model (GMM) is a clustering method that models the distribution of the data samples as a weighted Gaussian sum. The distribution of a feature vector f.sup.w.sup.i is given as

Σ.sub.k=1.sup.Kϕ.sub.k.Math.N(f.sup.w.sup.i|μ.sub.k,Σ.sub.k)  (1)

[0049] where K is the number of clusters (2 in this scenario) and N is the number of training sample. Parameters ϕ.sub.k, μ.sub.k and Σ.sub.k are the prior probability (weight), mean, and covariance matrix of the k.sup.th clusters. The parameter ϕ.sub.k satisfies 0≤ϕ.sub.k≤1 and Σ.sub.k=1.sup.Kϕ.sub.k.Math.N(f.sup.wi|μ.sub.k,Σ.sub.k) is the multivariate Gaussian distribution of the i.sup.th feature. The log-likelihood of the training samples are given by

ln p(F|Θ)=Σ.sub.i=1.sup.N{Σ.sub.k=1.sup.2ϕ.sub.k.Math.N(f.sup.wi|μ.sub.kΣ.sub.k)}  (2)

[0050] Where F=[f.sup.w.sup.1.sup.T, f.sup.w.sup.2.sup.T, . . . , f.sup.w.sup.N.sup.T].sup.T is the feature matrix and 6={ϕ.sub.1,ϕ.sub.2,μ.sub.1,μ.sub.2,Σ.sub.1,Σ.sub.2}. Θ is obtained by maximizing Equation (2) using Expectation-Maximization (EM) algorithm.

[0051] The two sets of clusters (C.sub.1 and C.sub.2) produced by GMM are unlabeled. Each cluster is parameterized by a set of ϕ, μ and Σ. The following equation is used to assign labels to the clusters.

[00001]$\begin{array}{cc}\mathrm{Class}=\{\begin{array}{c}{C}_1.fwdarw.\mathrm{Event},C_2.fwdarw.\mathrm{Noise}.Math.\text{:}.Math..Math..Math.{\Sigma }_{C_1}.Math.>.Math.{\Sigma }_{C_2}.Math.\\ C_1.fwdarw.\mathrm{Noise},C_2.fwdarw.\mathrm{Event}.Math.\text{:}.Math..Math..Math.{\Sigma }_{C_2}.Math.>.Math.{\Sigma }_{C_1}.Math.\end{array}& \left(3\right)\end{array}$

where |Σ.sub.C.sub.k| is the determinant of the co-variance matrix of the k.sup.th clusters. The covariance of the features matrix of the noise cluster has lower variance as compared to that of the footfall event class.

[0052] Referring to FIG. 2(b), another embodiment of present invention discloses the live phase of the said technique, i.e. it includes the detection and extraction of footfall events from the seismic signal of a user, wherein the user maybe a registered user or an imposter. A test signal Signal.sub.test of predefined length is segmented into equal parts (w.sub.i.sub.test) using a sliding window technique (size of the window and the overlapping ratio are set to 220 ms and 40). To predict the cluster (C.sub.1 or C.sub.2) of each segmented test signal Equation (4) is used.

[00002]$\begin{array}{cc}\mathrm{Class}=\{\begin{array}{c}{C}_1.Math.\text{:}.Math..Math.p\left(C_1f_{\mathrm{test}}^{w_i}\right)>p\left(C_2f_{\mathrm{test}}^{w_i}\right)\\ C_2.Math.\text{:}.Math..Math.p\left(C_2f_{\mathrm{test}}^{w_i}\right)>p\left(C_1f_{\mathrm{test}}^{w_i}\right)\end{array}& \left(4\right)\end{array}$

f.sub.test.sup.w.sup.i is the feature vector of the i.sup.th segment of the test signal and p(C.sub.k|f.sub.test.sup.w.sup.i)(=ϕ.sub.k.Math.N(f.sub.test.sup.w.sup.i|μ.sub.C.sub.k, Σ.sub.C.sub.k) for k=1,2) is the probability that f.sub.test.sup.w.sup.i belongs to class Ck.

[0053] As seen in FIG. 3(a), the results of event detection technique have been obtained and the footfall and noise portions of the seismic signal are labelled (0 as noise and 1 as an event) using the GMM method. The final signal is obtained by multiplying the labelled “1” segment of the signal with a Gaussian window. The center of the Gaussian window is placed at the location where the signal has the maximum amplitude within the segment.

[0054] In an exemplary implementation, the length of the Gaussian windows is considered to be 375 ms and the value of sigma is set to 4. These values are obtained by studying various footfall signals and rectangular window has been used for event extraction.

[0055] FIG. 4 illustrates the effect of the windowing technique on the frequency content of the footfall, and it can be observed that the footfall event obtained using rectangular window starts and ends abruptly. This abruptness in time domain leads to spectral leakages in the frequency domain and introduces unwanted high frequency harmonics. FIG. 4b(i) illustrates spectrogram and FIG. 4c(i) illustrates frequency spectrum of the event and the corresponding harmonics can be observed from these figures. This spectral leakage imparts noises in the features vector affecting the performance of the prediction algorithms. On the contrary, events obtained from Gaussian window, as in FIG. 4a(ii), have smooth edges. These events are free from spectral leakages as illustrated in FIG. 4b(ii) and FIG. 4c(i). This enhances the quality of the feature vector and improves the performance of the classifiers.

[0056] One of the exemplary implementations of present invention discloses a comparative performance analysis among USLEEM and the existing techniques UREDT and Adap-Th in terms of Person identification and imposter Detection. The same can be found in Tables II, III and IV below.

TABLE-US-00002 TABLE II Extracted features from a single footfall event Time Hilbert mean Frequency 0 to Domain std Transform std Energy Bins 250 Hz. skewness skewness (2 Hz.) kurtosis kurtosis event length

[0057] In an exemplary implementation, a dataset consisting labelled footsteps from 8 different individuals (four males and four female) is used to analyse the performance of the technique for person identification and imposter detection. The seismic signals generated by footfalls were recorded, as different volunteers walked around a geophone. Footfall generated seismic signals were recorded with a 16 bit analog to digital converter (ADC), as each volunteer walked around a geophone for five minutes. The sampling frequency of the ADC was 8 kHz. Dataset of an individual was recorded at a time. The entire experiment was repeated 12 times over a month. So, each individual class has almost 1 hour of seismic signal. The sensitivity and gain of the geophone used in this study are 2.88 V/mm/sec and 100. Each individual class has almost 1 hour of seismic signal in the dataset. It was collected over a span of 1 month. The event extraction techniques USLEEM, (UREDT+event extraction), and Adap-Th are used to extract individual footfall events from the entire dataset.

TABLE-US-00003 TABLE III Dataset created using USLEEM, UREDT and Adap-Th technique (j = 2, 3, 5, 7, 10) Footstep Number of footsteps sample Dataset Name per sample USLEEM UREDT Adap-Th Dataset.sub.i 1 47586 46645 46549 Dataset.sub.j j [00003]$\frac{47586}{j}$ [00004]$\frac{46645}{j}$ [00005]$\frac{46549}{j}$

TABLE-US-00004 TABLE IV Details of the scenarios used for performance analysis of imposter detection # # Imposters Registered # detection SN Scenario user Imposters problems 1 1 3 2 56 2 3 56 3 4 56 4 5 56 5 2 4 2 70 6 3 70 7 4 70 8 3 5 2 56 9 3 56

[0058] Table III displays the total number of footsteps extracted by the three event extraction techniques. Each footfall event, (referring to Table II) is treated as a single sample and the final dataset (Dataset.sub.1) is created by extracting features from them. A few more datasets are created by averaging the features from consecutive footfalls e.g. in Dataset.sub.j features from j consecutive footsteps are averaged and are treated as a single sample. In this way, five new datasets are created by assigning j to 2, 3, 4, 7, and 10.

[0059] Feature extraction plays a vital part in a biometric system. Classifiers' accuracy depends on the distinctiveness of features among different individuals. Time and frequency domain features are calculated for each of the footfall events (obtained from the event extraction technique). Mean, standard deviation, skewness, and kurtosis of the event make the time domain features. Length of the footfall event is also considered as an important feature. The length of a footfall is directly related to the shape and structure of the foot of an individual. Feature set also includes spectral energy of frequency bins of size 2 Hz from 0 to 250 Hz. It has been observed that there is no significant information in the footfall signal beyond 250-Hz.

Person Identification and Imposter Detection

[0060] Different multiclass supervised machine learning (ML) algorithms (SVM-Linear (SVM-Lin), SVM Gaussian (SVM-RBF), Logistic Regression (LR), Linear Discriminant Analysis (LDA) and k-Nearest Neighbor (kNN) are used for person identification. Each individual represents a single class in the ML algorithm. The datasets are normalized before training and testing, and 5-fold cross validation is carried out to avoid underfitting and overfitting. The hyper parameters of the classifiers (C for SVM-Lin, and C and γ for SVM-RBF) are obtained from grid search. The performance parameters for the algorithms include accuracy, precision, recall, and F1 score. It has been observed by researchers, combining features of multiple consecutive footsteps or discarding footsteps of low signal to noise ratio (SNR) increases the overall prediction accuracy. The required number of consecutive footsteps and the active area (sensing region) of a seismic sensor determine the total number of sensors required for a specific system. So, there is always a tendency to achieve high accuracy by using optimal number of footsteps. This reduces the overall infrastructure of the system and also decreases the prediction time (which may be very crucial in border areas and high security zones). So, the six datasets (Dataset.sub.1 . . . Dataset.sub.10), referring to Table III, are used to study the performance of the classifiers as number of footsteps per sample are increased.

Detection of Non-Registered User (Imposter)

[0061] The most challenging part in footfall based biometric system is imposter (intruder) detection. The main drawback of the above-mentioned ML classifiers is that they always assign footfalls of an unknown person (whose data were not used to train the model) to one of the predefined classes. The existing system only works when the individuals (classes) remain the same for training and testing scenarios. To overcome this problem, an imposter detection technique based on one-class/unary classifiers capable of predicting the presence or absence of an individual within the pre-registered dataset has been implemented. Using this technique, the system first detects the presence of an individual's footfall in the trained feature space. Classification for person identification is carried out only in case of registered users. Otherwise, it gives an intruder detected warning signal.

[0062] Two, one class machine learning models namely one class support vector machine (OC-SVM) and support vector data description (SVDD) are used for imposter detection. The one class classifiers (OC-SVM and SVDD) generate binary output i.e. +1 if the test sample lies within the decision boundary or −1 if the sample is an outlier. In imposter detection scenario data of each registered user are used to train individual one class models. If there are r number of registered users, there will be r number of trained one class models. The test sample is predicted as imposter (anomaly) if all the trained one class classifiers return −1 i.e. the test sample is anomalous to all the one class classifiers. Eqn. 5 shows the decision rule used for predicting imposters.

[00006]$\begin{array}{cc}\mathrm{Test}.Math.\text{-}.Math.\mathrm{Sample}=\{\begin{array}{c}\mathrm{Imposter},\forall \mathrm{Trained}.Math..Math.\mathrm{Model}=-1\\ \mathrm{Registered}.Math..Math.\mathrm{User},\mathrm{otherwise}\end{array}& \left(5\right)\end{array}$

[0063] Three main scenarios, as in Table III, uses the footfall dataset of 8 individuals to calculate the performance of the techniques for imposter detection. In each scenario, r number of classes is treated as registered users (r={3, 4, 5}). Each scenario is further divided into sub-scenarios. These sub-scenarios are created by fixing the number of registered users and varying the number of imposters from 2 to 8-r. Each sub-scenario is further subdivided into .sup.8Cr combinations of imposter detection problem to compute the robustness of the techniques. It is done to select all sets of r combinations (order of selection does not matter) of classes as registered users. The imposters are selected from the rest. Cases where the number of registered users is less than 3, or the number of imposters is less than 2 were not considered as these scenarios do not resemble realistic situation. The final performance of the techniques corresponding to each scenario is calculated by averaging the results obtained from all sub-scenarios. In the training phase, only the data of registered users are available. The data of the imposters are completely unseen by the trained machines and are only used during the live phase (which resemblance real scenario cases). Fivefold cross validation is carried out in each of the imposter detection problems.

[0064] The performance of human identification and imposter detection techniques, referring to Table V, are calculated using the dataset.

TABLE-US-00005 TABLE V Performance of different classifiers used for person identification with varying footsteps per sample Classifiers Number of Footsteps/sample Accuracy (%) 1 2 3 5 7 10 SVM Lin. 79.00 88.4 91.97 94.97 96.72 97.39 SVM RBF 83.47 90.51 93.06 95.71 96.79 97.54 LR 73.48 84.37 88.23 93.62 96.07 96.82 LDA 67.86 78.75 84.63 89.58 93.24 95.54 kNN 71.86 80.53 84.72 88.23 91.21 94.24

Person Identification

[0065] The first phase of a person identification technique is event detection and extraction. Event detection capability of USLEEM is compared with existing techniques like UREDT, Adap-Th, amplitude threshold (Amp-Th.), and kurtosis based technique. Event extraction technique Adap-Th uses STA-LTA for event detection. FIG. 5 shows the performance of all the techniques in detecting seismic events. An annotated dataset containing 1165 footfall events is used for calculating the performance parameters. USLEEM and UREDT outperform the rest of the techniques. Unlike others, USLEEM and UREDT do not require any tuning parameter or prior knowledge of the data. Both of them use a GMM based clustering technique in the training phase. However, in the final prediction phase, USLEEM uses the trained GMM model obtained from the training phase. On the other hand, UREDT trains an SVM model with 135 features of the footfall event. So, UREDT is computationally more expensive and memory hungry than USLEEM.

[0066] The performance of different classifiers (for human identification) with variation in number of footsteps per sample is presented in Table V. Events are extracted using the USLEEM technique. It is important to achieve high prediction accuracy from few footfalls, as it reduces prediction time, saves energy and is economical. The sensing range of geophones on concretes is 2˜m to 2.5˜m. So, more than one geophone will be required for applications that need very high prediction accuracy. As the system will demand a large number of footsteps per sample. It can be observed, referring to Table V, that in case of 10 footsteps/sample the performance of all the classifiers is almost the same. However, in 1 footstep/sample scenario SVM-RBF outperforms the rest and achieves an accuracy of 83.47% which is 25.61% higher than LDA (which performs worst). The Gaussian kernel of SVM-RBF increases the dimension of the feature space (disjointing the classes) before performing classification. The feature space of the classes is somewhat overlapped in 1-footstep/sample scenario and they slowly become disjoint as number of footsteps per sample increase.

[0067] FIG. 6 exhibits the accuracy (person identification) obtained by SVM-RBF when different event extraction techniques (USLEEM, UREDT, and Adap-Th) are used. It is observed that in all cases USLEEM and UREDT outperform Adap-Th Identification accuracy of 90% is achieved by USLEEM and UREDT in 2-footstep/sample scenario, whereas Adap-Th requires 5 footsteps/sample to attain 90%. Better performance of USLEEM and UREDT is due to the use of a Gaussian window for event extraction. Adap-Th technique uses a rectangular window for extraction of final footfall event. The spectral leakage of the rectangular window affects the frequency domain features of the footfall. The rate of misclassification is higher in Adap-Th technique as most of the features for person identification, referring to Table II, are created from the frequency domain.

[0068] FIG. 7 shows the learning curve of SVM-RBF obtained using USLEEM and Adap-Th technique. Learning curves are used to find the minimum number of training samples required by classifiers to attain a certain accuracy. From an implementation point of view, it's very important to know the exact number of footsteps required per class (individual) to train the model. It is observed, referring to FIG. 7, that USLEEM requires 800 footsteps of training samples per class to achieve an identification accuracy of 91%, and Adap-Th needs 2500 footsteps of training samples per class to attain an accuracy of 88% (with dataset.sub.5). So, for achieving an identification accuracy of approximately 90%, an individual need to walk around the sensor (in the training phase) for 5 minutes when USLEEM is used and for 25 minutes in case of Adap-Th (assuming a normal human being takes 1 min to walk 100 footsteps). A comparison of the learning curves of SVM-RBF with dataset.sub.2 and dataset.sub.5 is also carried out. Samples containing more footsteps trend to achieve higher accuracy with a lower number of training samples. 10% to 15% improvement in accuracy of SVM-RBF is observed when 5 footsteps/sample (dataset.sub.5) are used in place of 2 footsteps/sample (dataset.sub.2). Confusion matrix of SVM-RBF (applied on dataset.sub.2) is shown in Table VI.

TABLE-US-00006 TABLE VI Confusion matrix of person identification obtained using SVM-RBF on Dataset.sub.2. Actual P1 0.91 0.02 0.03 0.04 0.00 0.00 0.01 0.00 Class P2 0.04 0.87 0.01 0.06 0.00 0.00 0.01 0.01 P3 0.04 0.00 0.92 0.02 0.00 0.00 0.01 0.01 P4 0.03 0.03 0.03 0.87 0.00 0.00 0.02 0.02 P5 0.00 0.00 0.00 0.00 0.89 0.09 0.00 0.02 P6 0.00 0.00 0.01 0.00 0.04 0.88 0.00 0.07 P7 0.01 0.00 0.02 0.01 0.00 0.00 0.96 0.00 P8 0.01 0.00 0.00 0.04 0.01 0.04 0.00 0.90 P1 P2 P3 P4 P5 P6 P7 P8 Predicted Class

Imposter Detection Using One Class Classifiers

[0069] A comparative analysis, referring to FIG. 8, of the two one class classifiers (OC-SVM and SVDD) used for imposter detection is illustrated. The hyper-parameters (C=1, γ=0.04 and ν=0.2) for both the classifiers are obtained using grid search. The prediction accuracies of registered and non-registered (imposters) users, referring to FIG. 8, for three different scenarios achieved by OC-SVM and SVDD. It also presents the variation in performance as the number of footsteps per sample is increased from 7 to 10. It can be noticed that SVDD outperforms OC-SVM in all the three scenarios and the performance of the classifiers improves with the increase in the number of footsteps per sample. In all the three scenarios SVDD achieves a prediction accuracy of 76% to 80% with 10 footsteps per samples.

[00007]$\begin{array}{cc}\mathrm{FPR}=\frac{\mathrm{FP}}{\mathrm{FP}+\mathrm{TN}},\mathrm{TPR}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}& \left(6\right)\end{array}$

SVDD encloses the dataset with a spherical boundary and OC-SVM draws a hyperplane between the dataset of the training class and the origin. If r is the number of registered users, SVDD has r spherical boundaries and OC-SVM has r hyperplane on the feature space. The r hyperplanes in the feature space overlap among themselves and results in poor performance of OC-SVM. The hyperspheres of SVDD are more disjoints (non-overlapped) than the hyperplanes of OC-SVM due to its spherical nature. ROC curves of SVDD and OC-SVM are obtained from scenario 2, referring to Table IV and FIG. 9, using 10 footsteps per sample (Dataset.sub.10). SVDD and OC-SVM are discrete classifiers (return only the class label in the output), so each class, irrespective of registered user and imposter, corresponds to a single point (FPR, TPR) in the ROC space. FPR (false positive rate) and TPR (true positive rate) corresponding to a class is calculated using where FP is false positive, TN is true negative, TP is true positive and FN is false negative. The coordinate (0, 1) in the ROC graph represents a perfect classifier. A classifier is better than another, if it lies north-west to the other in the ROC space. So, the ROC curve of SVDD is much better than OC-SVM for both the classes.

[0070] Performance parameters (accuracy, precision, recall, and F1 score) of SVDD for imposter detection are presented in Table VII. The mean and standard deviation are calculated by averaging the performance parameters of all the sub scenarios. Imposter detection accuracy decreases slightly (4% for SVDD and 1% for OC-SVM) from scenario 1 to scenario 3 when dataset.sub.10 is used. An increase in the number of registered user results in large overlapping of different classes in the feature set. However, for both the techniques the overall accuracy and individual F1 score of the classes almost remain the same as number of imposters increases for all the three scenarios. This makes the system robust in imposter detection.

TABLE-US-00007 TABLE VII Performance of SVDD for different imposter detection scenarios obtained using Dataset.sub.10. # Performance Parameters Registered Overall Precision Recall F1 Scenario User Parameters Accuracy Registered Imposter Registered Imposter Registered Imposter 1 3 Mean 80.22 85.88 76.63 72.96 87.47 78.58 81.51 Std 2.48 4.13 1.82 2.35 4.30 2.42 2.62 2 4 Mean 78.31 82.09 75.83 73.46 83.16 77.26 79.14 Std 2.74 4.35 2.08 2.50 4.94 2.53 3.03 3 5 Mean 76.19 78.70 74.57 73.25 79.13 75.62 76.58 Std. 2.77 3.52 2.69 3.21 4.29 2.72 2.97

TABLE-US-00008 TABLE VIII Performance of OC-SVM for different imposter detection scenarios obtained using Dataset.sub.7. # Performance Parameters Registered # Overall Precision Recall F1-Score Scenario User Imposter Accuracy Registered Imposter Registered Imposter Registered Imposter 1 3 2 Mean 75.65 77.14 75.29 75.19 76.11 75.71 75.27 Std. 5.29 7.89 3.19 2.13 10.37 4.12 6.70 3 Mean 75.74 76.52 75.70 75.49 75.99 75.71 75.58 Std. 3.39 4.99 2.29 2.41 6.63 2.78 4.18 4 Mean 75.23 75.74 75.30 74.98 75.47 75.12 75.18 Std. 2.14 3.21 1.49 1.59 4.21 1.77 2.61 5 Mean 75.37 76.13 75.19 74.45 76.30 75.05 75.56 Std. 1.04 1.26 1.35 1.91 1.72 1.15 1.07 Average Mean 75.50 76.38 75.37 75.03 75.97 75.40 75.40 Std. 2.97 4.34 2.08 2.01 5.73 2.46 3.64 2 4 2 Mean 73.11 72.83 74.22 76.21 70.00 74.17 71.66 Std. 4.83 6.36 3.51 2.66 9.15 3.75 6.32 3 Mean 72.53 71.86 73.77 75.51 69.56 73.45 71.38 Mean 2.76 3.59 2.16 1.94 5.25 2.18 3.57 4 Std. 72.68 71.86 73.92 75.40 69.96 73.46 71.75 Mean 1.34 1.52 1.45 1.68 2.24 1.25 1.57 Average Mean 72.77 72.18 73.97 75.71 69.84 73.69 71.60 Std. 2.98 3.83 2.37 2.09 5.55 2.39 3.82 3 5 2 Std. 70.46 69.31 72.32 75.64 65.28 72.14 68.35 Mean 3.88 4.65 3.13 2.37 7.13 2.97 5.15 3 Std. 70.99 69.60 73.01 76.11 65.88 72.58 69.09 Mean 1.21 1.34 1.33 1.50 2.09 1.08 1.49 Average Mean 70.73 69.45 72.66 75.88 65.58 72.36 68.72 Std. 2.54 2.99 2.23 1.93 4.61 2.03 3.32

[0071] Overall, the system records and analyses footfall generated seismic waveforms. The unsupervised learning-based event detection and extraction technique USLEEM outperforms the existing techniques and is computationally cheaper than UREDT. Using SVM-RBF with USLEEM we were able to achieve an identification accuracy of 90.6% (features extracted from two consecutive footsteps), thereby identifying humans as registered users and imposters and making the system suitable for certain applications where the prior information of imposters is not available.

[0072] According to an embodiment of the present invention, a method for person identification and imposter detection comprises steps of detection and extraction (201, 202; 301, 302) of seismic signals generated from corresponding footfalls, by means of unsupervised learning based detection and extraction module (USLEEM); and detection and identification of imposter and/or registered users (304, 305, 306, 307) respectively by means of an identification module. The said step of detection and extraction of seismic signals generated from corresponding footfalls comprising at least one training phase including—splitting (202) each of said seismic signals, into N equal segments, extracting (202) feature vectors (FE-I) corresponding to time domain and frequency domain features from each of said segment of the N segments, clustering (202) each of the said feature vectors into a clustered event so as to form at least one trained model of USLEEM, and storing of the said trained model. The step of detection and extraction of seismic signals generated from corresponding footfalls further comprising at least one live phase including—splitting (301) each of said seismic signals, into N equal segments, extracting (302), feature vectors (FE-I) corresponding to time domain and frequency domain features from each of said segment of the N segments and detecting footfalls from the extracted featured vectors (FE-I) thereafter feeding (303) it to the trained model. The step of detection and identification of imposter and/or registered users respectively is performed by means of an identification module, said method steps including—detection (307) of at least one imposter, by means of One—Class machine learning model when feature vectors (FE-II) obtained from the extracted footfalls of live phase are fed to the One-Class model and identification of registered users, and identification (306) of registered users, from feature vectors (FE-II) obtained from the extracted footfalls of live phase, by means of multi-class machine learning model. Gaussian Mixture Model, GMM, clustering facilitates clustering each of the said feature vectors into a clustered event. Each cluster is parameterized by a set of p, p, and E and the method further comprises classifying said clustered event into a footfall event and noise by

[00008]$\mathrm{Class}=\{\begin{array}{c}{C}_1.fwdarw.\mathrm{Event},C_2.fwdarw.\mathrm{Noise}.Math.\text{:}.Math..Math..Math.{\Sigma }_{C_1}.Math.>.Math.{\Sigma }_{C_2}.Math.\\ C_1.fwdarw.\mathrm{Noise},C_2.fwdarw.\mathrm{Event}.Math.\text{:}.Math..Math..Math.{\Sigma }_{C_2}.Math.>.Math.{\Sigma }_{C_1}.Math.\end{array}$

where |Σ.sub.C.sub.k| is the determinant of the co-variance matrix of the k.sup.th clusters. During the live phase, each of the N segments of the footfall generated seismic signals are classified by,

[00009]$\mathrm{Class}=\{\begin{array}{c}{C}_1.Math.\text{:}.Math..Math.p\left(C_1f_{\mathrm{test}}^{w_i}\right)>p\left(C_2f_{\mathrm{test}}^{w_i}\right)\\ C_2.Math.\text{:}.Math..Math.p\left(C_2f_{\mathrm{test}}^{w_i}\right)>p\left(C_1f_{\mathrm{test}}^{w_i}\right)\end{array}$

where f.sub.test.sup.wi is the feature vector of the i.sup.th segment of a test signal Signal.sub.test, p(C.sub.k|f.sub.test.sup.w.sup.i) (=ϕ.sub.k.Math.N(f.sub.test.sup.w.sup.i|μ.sub.C.sub.k,Σ.sub.C.sub.k) for k=1,2) is the probability that f.sub.test.sup.wi belongs to class C.sub.k, test signal Signal.sub.test is the seismic signal generated by an imposter or a registered user during the live phase and it is segmented into equal parts (w.sub.i.sub.test).
Some of the non-limiting advantages of seismic sensor based biometric system are: [0073] 1. Individuals are not required to orient or position themselves in a special manner. [0074] 2. Easy implementation, sensor data are less affected by environmental parameters, and the sensor is easily camouflageable. [0075] 3. Beneficial for sentries posted in high-security zones (bureaucratic building), military camps, and army check posts.

[0076] Although a method for person identification and imposter detection using unsupervised learning-based detection and extraction technique, USLEEM, a smart device implementing the method and a biometric authentication system thereof has been described in language specific to structural features, it is to be understood that the embodiments disclosed in the above section are not necessarily limited to the specific methods or devices described herein. Rather, the specific features are disclosed as examples of implementations of method for person identification and imposter detection using unsupervised learning-based detection and extraction technique, USLEEM, a smart device and a biometric authentication system thereof.