Footstep Identification Method Based On Gaussian Mixture Model

Footstep identification technology is a very potential biometric identification technology,which uses the sound or vibration signals generated when a person walks to realize the identification of walking person.It has the advantages of concealment,non-contact and not easy to be forged.The footstep signals are easily disturbed by environmental noise because of its small amplitude,which brings great difficulties to the noise reduction preprocessing of the footstep signals,and the footstep identification has problems such as lack of data and low recognition accuracy.Therefore,it is of great significance to carry out research on the preprocessing of footsteps with low signal-to-noise ratio and the footstep identification with large amounts of data.On the basis of summarizing the research status of footstep identification at home and abroad,this paper takes footstep identification system as the research object,and conducts research on the issues related to footstep identification,such as footstep signals acquisition,preprocessing,feature parameters extraction,and model establishment.The details of this paper as follows:First,the footstep signals were acquired using a dynamic signal acquisition and analysis system,and the preprocessing methods of footstep signals such as noise reduction and starting point detection were researched.The footstep signals acquisition program was designed,and a large number of footstep signals were acquired in a low-noise enclosed indoor environment using a dynamic signal acquisition and analysis system,and the footstep signals were analyzed in time-frequency domain.The original footstep signals were denoised using a combination of filtering and spectral subtraction.Then,the starting points of the continuous footstep signals in the natural walking state were detected according to the footstep sound pressure level and the adaptive threshold.Furthermore,the continuous footstep signals were divided into units consisting of only two adjacent footsteps,and each footstep unit was standardized by Z-score to complete the collection and preprocessing of the footstep signals.Secondly,the extraction of feature parameters of footsteps were studied using standardized footstep units,and the testers'Gaussian mixture models(GMM)were built based on feature parameters.The Mel frequency cepstral coefficient(MFCC),perceptual linear prediction coefficient(PLP)and its first-order and second-order dynamic difference parameters of the footstep units were extracted based on the characteristics of human hearing.The Gaussian mixture model of the testers were established based on the feature parameters training data,the K-means algorithm was used to initialize the model parameters,and the expectation maximization(EM)algorithm was further used to complete the Gaussian mixture model parameter estimation.The judgment criteria for the footstep identification were improved,and the Gaussian mixture model was used to realize the identification of testers.Finally,the influencing factors of the feature parameters of footsteps were explored using closed set footsteps recognition system,the recognition accuracy of the closed set footstep recognition system under different training data and test data lengths was further studied,and the equal error rate of the footstep identification confirmation sy stem was studied.The recognition accuracy rate of the closed set footstep recognition system was used as the judgment,and the controlled variable method was adopted to study the influence of frame length,ratio of frame shift to frame length and dimension of MFCC,PLP and their dynamic difference parameters on the recognition results.The recognition accuracy of the closed set footstep recognition system under different training data and test data lengths was further studied.Combined with the research results of gait recognition,the feature parameters were further optimized to reduce the time of feature parameters extraction,Gaussian mixture model training and recognition process.The equal error rate of the footstep identification confirmation system were explored using optimized feature parameters.This paper completed the noise reduction of low signal-to-noise ratio footstep signals,and proposed a starting points detection method using footstep sound pressure level and adaptive thresholds.The Gaussian mixture model was established based on the feature parameters of the footstep signals,and the footstep identification was realized based on the Gaussian mixture model.Research shows that the recognition accuracy rate of the closed set footstep recognition system reached 98.2%,and the equal error rate of the footstep identification confirmation system was 7%.This research has positive significance for the enrichment of biometric recognition technology and the development of multi-biological feature fusion.