Research On Gait Parameter Calculation Method Based On Azure Kinect

At present,there are two main methods for calculating gait kinematic parameters and spatiotemporal parameters of gait,i.e.,visual method and quantitative method.The visual method is commonly used in medical professional examination,that is,the doctor asks the patients to make prescribed actions,and then evaluates the patients according to the posture,time and body movement state of them when performing the action according to the evaluation level standard in the scale,and finally obtains the evaluation of the patients’ gait.The quantitative method mainly collects and detects the walking data of the patient through the sensors,and calculates the gait parameters quantitatively.The quantitative method mainly includes marking method,wearable sensor method and unmarked method.Unmarked method is most extensively researched because of its low cost,convenient testing and short time-consuming.Kinect devices are commonly used in unmarked method,and the latest version is Azure Kinect(hereinafter referred to as AK)which is currently less frequently used in research and applications.The traditional method and machine learning method are mainly used to calculate the gait parameters in unmarked method.The traditional method designs the calculation formula of gait parameters according to the parameter definition,but the generalization of the traditional method is not well due to the individual difference of gait.The machine learning method uses machine learning algorithms such as SVM,fuzzy logic,neural network and other methods to learn the data so as to get the algorithm model to calculate gait parameters.In this manuscript,according to the advantages and disadvantages of traditional method and machine learning method,two kinds of parameter calculation methods are designed to calculate different gait parameters.Based on the relevant definition of gait parameters,the traditional method designs formulas to calculate knee flexion angle,ankle flexion angle,gait speed and gait frequency,which have no obvious individual difference.Based on machine learning method,neural network is used to calculate step size,step length,support phase time and swing phase time,which have obvious individual differences.The traditional method firstly builds the model of lower limb,then puts forward and designs the parameter calculation formula according to the relevant definition of the parameters,next selects the appropriate filtering method according to the shape of the required data,and finally calculates the gait parameters in the MATLAB2015 b environment.In the machine learning method,two neural network models are proposed for different kinds of parameters,in which the multi-layer perceptron network model is used to calculate the step size,and the convolutional-recurrent neural network model is used to calculate the step length,support phase time and swing phase time.The label set comes from the wearable sensor socks.The network model is realized and trained in Python3.8 environment.In order to verify the accuracy and reliability of the calculation method,a consistency test experiment is designed.Specifically,the results obtained by the calculation method proposed in this manuscript and the results obtained by the mature labeling method are used to test the consistency,conducting paired t-test,calculating the correlation coefficient(R value),and drawing the bland-altman diagram and regression-scatter diagram to evaluate and analyze the accuracy and reliability of the calculation results.Finally,the AK equipped with the calculation method proposed in this manuscript is used in the gait analysis of stroke patients to verify its practicability.And ideal results are obtained.