Research On Adaptive Gait Recognition Method Based On Acceleration Signal

Gait recognition technology based on acceleration signal has many problems in practical application not only caused by poor adaptalility of gait feature extraction methods but also by the time limitation of feature stability and under such circumstances the research is still at the stage of experiment.As an intelligent guardian of information security for portable electronic products,such as smart phones and wearable devices,it is necessary to recognize user's identity effectively in a sustainable way for a long time under the minimum constraints.In order to solve the time limitation of gait feature,the method of gait recognition should adjust its adaptability accordingly.Therefore,the gait recognition system influenced by internal and external factors still work well to recognize the user's identity when the changes of short-term adaptive and long-term compensatory happen to gait pattern.Studies show that there is no research concerning its adjusting adaptability.Therefore,this thesis takes adaptive preprocessing method,adaptive outlier detection method and adaptive feature extraction method as the research objects,which are the key steps of gait recognition.The primary researches and contribution are as followed.(1)By using hybrid filtering and adaptive waveform detection,a method of gait recognition is proposed in view of the problem of extracting consistent gait cycle waveform.Firstly,a variety of noise,which is introduced from the process of data acquisition and analog-digital conversion,is removed efficiently via a hybrid filtering method according to characteristic of acceleration signal,low frequency.Then similarity of phase interval waveform of gait cycle is measured adaptively by the standard deviation,which is used to solve the malfunction of gait signal caused by short-term adaptation of gait movement change or rotation displacement of data acquiring equipment.At last,feature templates are calculated via six discrimination functions so that the effective recognition can be reached.The experiment results show that this method is suitable for gait recognition in small sample range,and the recognition rate is 96.67% in 30 subjects × 6 series datasets.(2)By using the multi-features combination,a method of gait recognition is proposed in view of the problem of reducing the dimensions of gait cycle feature.Firstly,the amplitude of gait cycles are normalied by the mahalanobis distance according to quasi-periodic characteristic of gait acceleration signal,and gait signal is filtered by hybrid filtering methods.Then 14 single-features are firstly extracted from the waveform of a cycle signal.After 14 single-features are transformed into 6 feature vectors,the stability and separability of single-feature is analyzed and evaluated.At last,the performance of gait recognition is evaluated by calculating the correct recognition rate with the Euclidean distance.The experiment results show that this method is suitable for gait recognition in small sample range,and the recognition performance is improved comparing with other algorithms of the same kind,and the recognition rate is 94.44% in 30 subjects × 6 series datasets.(3)By using adaptive wavelet denoising and adaptive outlier detection,a method of gait recognition is proposed in view of the problem of selecting the optimal parameters for quality assessment of the wavelet de-noising and extracting stability of gait features,due to the limited number of sampling points and many outlier patterns of gait signals.Firstly,the optimal parameters for quality assessment of the wavelet de-noising are adaptively decided according to the characteristic of wavelet transform,which is suitable for processing non-stationary signal.Abnormal periodic pattern are removed before feature extraction,and the abnormal degree of the gait signal subpattern is discriminated by the value of K-mean distance outlier factor.At last,the mean of gait cycles after the outlier detection is averaged according to Euclidean distance.Thus,gait recognition is improved effectively.The experiment results show that the system recognition performance is improved by taking the adaptive requirements into account and improving stability and effectiveness of gait feature,during which observations are replaced by calculations.(4)By using adaptive wavelet denoising and SIFT descriptor,a method of gait recognition is proposed in view of the problem of extracting gait feature in the extremum region of acceleration signals.This paper uses the quality evaluation method of wavelet de-noising to select the optimal wavelet basis function adaptively,and gait acceleration signals are pretreated.Then it achieves gait feature extraction by using SIFT descriptor based on K-means cluster analysisi and fitting of signature points dataset.At last,gait features are matched by correlation coefficient in order to improve the recognition performance.The experiment results show that this method avoids the division process of gait cycle,and is suitable for gait recognition in small sample range.(5)The conventional gait recognition algorithm based on acceleration signal to extract gait features has low recognition rate in minimal constraint conditions or relatively long time span.To solve this problem,a novel gait recognition algorithm based on SIFT feature sparse representation and template fusion is proposed.Firstly,acceleration signals are collected to construct various DoG(difference of Gaussian)scale-spaces.The location information template of the gait features by sparse representation is built.Then the gait cycle features based on sparse representation is effectively converted with the fusion of gait templates,in order to avoid the changes of short-term adaptive and long-term compensatory happening to gait pattern.The gait features are recognized by the nearest neighbor approach and the voting scheme.The results of experiment demonstrate that the proposed algorithm significantly outperforms other methods.Based on open access datasets of 175 subjects × 6 series,the recognition rate of 98.67% and the verification of 99.89% are obtained.Furthermore,the influence on the recognition effect by different composition of training samples and testing samples is further studied,which indicates the stability and effectiveness of the feature extraction proposed by the method.In addition,the correspondence of gait phase,functional phase and acceleration signal waveform is also discussed in the paper,in order to provide theoretical basis for accelerationbased gait recognition reaserch.