A millimeter wave radar-based parkinson disease gait feature extraction method, gait recognition method and system

Abstract

The application discloses a Parkinson disease gait feature extraction method, a gait recognition method and a system based on a millimeter wave radar, and the feature extraction method comprises the following steps: collecting a millimeter wave radar signal, pre-processing the signal to obtain a Doppler spectrum and a distance spectrum of the signal; extracting a trunk speed curve from the Doppler spectrum of the signal by using an adjacent column constraint rule, extracting a trunk position change curve from the distance spectrum, and extracting basic gait information from the trunk speed curve and the trunk position change curve; extracting an envelope speed curve corresponding to sample motion from the Doppler spectrum, and screening out key gait features in combination with the basic gait information. The application does not depend on a multi-sensor wearing system, and fully reveals typical motion difference characteristics of Parkinson disease patients by targeted modeling of gait detail characteristics.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of artificial intelligence feature extraction, and more particularly relates to a Parkinson's disease gait feature extraction method, gait recognition method and system based on millimeter wave radar. BACKGROUND

[0002] Parkinson's disease is a common neurodegenerative disease, commonly seen in the elderly population, and the corresponding clinical symptoms are divided into motor symptoms and non-motor symptoms, among which the motor symptoms include bradykinesia, tremor, muscle rigidity and postural instability, etc. Axial symptoms are common symptoms of patients with advanced Parkinson's disease, specifically manifested as abnormal gait posture, unbalanced walking, etc. Accurate measurement and analysis of gait symptoms help to develop subsequent treatment plans for Parkinson's disease patients.

[0003] In the prior art, gait recognition methods mostly rely on wearable sensor systems such as electromyography, inertial measurement units and plantar pressure sensors, etc., and realize motion phase recognition and gait classification by collecting multi-modal sensing signals and combining deep learning models.

[0004] The invention patent with the publication number CN117297589A proposes a gait recognition method based on multiple sensors and BI_LSTM, which uses a lightweight gait collection system to collect commonly used motion signals of the lower limbs of the human body, including surface electromyography of the thigh, plantar pressure signals, and lower limb angle and angular velocity signals; the collected motion signals are preprocessed to obtain smooth data, and then gait division and data calibration are performed, and a gait recognition database is constructed through feature extraction of gait sensing signals; for time series signals, a BI_LSTM submodule is used, and for non-time series signals, a fully connected submodule is used, then the outputs of the two submodules are sent together as new features to the fully connected layer output layer, and the final classification result is obtained. This scheme strongly depends on a multi-sensor wearable system, and the patient has a heavy burden of wearing, the actual deployment is complex, and it mainly focuses on gait phase recognition, but lacks targeted modeling of gait detail features, which cannot fully reveal the typical motion difference features of Parkinson's disease patients, limiting its application potential in disease screening and early warning. SUMMARY

[0005] To overcome the problems in the prior art that gait recognition methods strongly depend on multi-sensor wearable systems and lack targeted modeling of gait detail features, the present application provides a Parkinson's disease gait feature extraction method, gait recognition method and system based on millimeter wave radar.

[0006] The primary object of the present application is to solve the above technical problems, and the technical solution of the present application is as follows:

[0007] The first aspect of the present application provides a Parkinson's disease gait feature extraction method based on millimeter wave radar, comprising the following steps:

[0008] Collecting millimeter wave radar signals, pre-processing the signals to obtain Doppler spectrum and range spectrum of the signals;

[0009] Extracting a torso speed curve from the Doppler spectrum of the signals using an adjacent column constraint rule, extracting a torso position change curve from the range spectrum, and extracting basic gait information from the torso speed curve and the torso position change curve;

[0010] Extracting an envelope speed curve corresponding to sample motion from the Doppler spectrum, and screening out key gait features in combination with the basic gait information.

[0011] Further, the millimeter wave radar signal is pre-processed, including the following steps:

[0012] Using the millimeter wave radar to emit a linear frequency modulation continuous wave signal to the detected human body, and collecting radar echo signals in the human motion process;

[0013] Using the received radar echo signal and the emitted signal to mix in phase / quadrature demodulation to obtain an intermediate frequency signal, and performing analog-to-digital conversion sampling to output an intermediate frequency signal discrete sequence;

[0014] Using fast Fourier transform to convert the intermediate frequency signal discrete sequence into a range spectrum, and using the phase change between multiple chirp signals to perform two-dimensional fast Fourier transform to obtain a Doppler spectrum;

[0015] Performing a difference operation on adjacent frames on the range spectrum, and suppressing spectrum leakage by windowing, and obtaining an enhanced range spectrum and Doppler spectrum by using a non-coherent accumulation method in combination with multiple receiving antenna data.

[0016] Further, the torso speed curve is extracted from the Doppler spectrum of the signals using an adjacent column constraint rule, including the following steps:

[0017] Performing static filtering processing on the Doppler spectrum to obtain a purified Doppler spectrum graph;

[0018] In the purified Doppler spectrum graph, the adjacent column constraint rule is used for torso signal path extraction, the maximum value of each column is used as a path node, the peak value of the next column is detected only within a preset range above and below the maximum value position of the previous column during path searching, and a traversal from the last column to the front is used to ensure that the found path is the strongest energy path;

[0019] Using a five-point cubic smoothing filter algorithm to perform curve smoothing processing on the extracted torso signal path to suppress burrs and abnormal mutations, and obtaining a continuous and smooth torso speed curve.

[0020] Further, the basic gait information is extracted from the torso speed curve and the torso position change curve, including the following steps:

[0021] Performing frequency domain analysis on the torso velocity curve, obtaining the frequency domain secondary peak as the step frequency, and calculating the average step time of each step according to the step frequency;

[0022] Performing peak detection on the torso velocity curve, setting the interval between adjacent peaks according to the average step time of each step, if multiple candidate peaks are detected within the interval, retaining the one with more significant amplitude, and determining the start and end time points of each step;

[0023] Mapping the start and end time of each step with the corresponding torso position change curve, and calculating the basic gait information of each step.

[0024] Further, the envelope velocity curve corresponding to the sample motion is extracted from the Doppler spectrum, including the following steps:

[0025] Arrange the Doppler spectrum data in time frames to form a two-dimensional matrix structure, where each column represents the energy distribution of different speed intervals at a time;

[0026] Set the energy of all rows corresponding to absolute values of speed below the preset speed threshold to zero, and complete the static area energy zeroing;

[0027] Cycle each column of data and perform peak detection, filter peaks below the dynamic energy threshold, and redundant peaks with an interval less than the first preset threshold, the calculation method of the dynamic energy threshold is: extract all positive elements from the region after the static area energy zeroing, calculate the median of the positive elements, multiply the median by the adjustment coefficient to obtain the dynamic energy threshold;

[0028] Select the first point above the threshold from the top as the upper envelope point of the column, and the last point below the threshold from the bottom as the lower envelope point of the column, if no peak in the current column meets the conditions, temporarily mark it as null;

[0029] Perform speed mutation detection on the preliminary constructed envelope point sequence, mark the points with a speed change amplitude exceeding the second preset threshold between adjacent column envelope points as abnormal, and replace them with adjacent legal envelope points, to obtain the preliminary upper envelope curve and lower envelope curve;

[0030] Perform median filtering on the preliminary upper envelope curve and lower envelope curve respectively to remove isolated noise points, and use linear interpolation method to fill in missing envelope points, finally construct smooth and continuous upper envelope curve and lower envelope curve.

[0031] Further, use the envelope velocity curve combined with the basic gait information to filter out key gait features, including the following steps:

[0032] Peak detection and gait segmentation are performed on the envelope velocity curve to obtain an absolute velocity curve reflecting the movement regularity of the arm and foot;

[0033] The envelope velocity curve is differentiated with respect to the trunk velocity curve at the same time to obtain a relative velocity curve of the envelope with respect to the trunk;

[0034] Peak detection is performed on the envelope velocity curve and the relative velocity curve to extract the velocity and distance information in each step, and the time interval between the envelope peak value and the trunk velocity peak value is calculated as a key feature representing gait synchrony;

[0035] Frequency domain analysis is performed on the envelope velocity curve to extract the frequency spectrum peak value as a periodic gait feature;

[0036] Joint analysis of the envelope velocity curve and the relative velocity curve in the time domain and the frequency domain is performed to extract multi-dimensional gait features;

[0037] Statistical processing is performed on the extracted multi-dimensional gait features to screen out key gait features that can significantly distinguish Parkinson's disease patients from healthy people.

[0038] Further, statistical processing is performed on the extracted gait features to screen out key gait features, including the following steps:

[0039] Feature extraction is performed on each original gait feature to calculate its mean, median, first term, maximum value, minimum value, variance, and asynchronous symmetry, respectively, to obtain a statistical feature set;

[0040] Principal component analysis is performed on the statistical feature set to obtain a low-dimensional representation vector;

[0041] Normality test is performed on the low-dimensional representation vector to determine whether the low-dimensional representation vector is subject to normal distribution;

[0042] Independent sample T test is used for features subject to normal distribution, and Mann-Whitney U test is used for features not subject to normal distribution, for significant difference analysis between groups to obtain significance level P value;

[0043] Cohen's d value effect size is calculated for each feature to obtain the performance of each feature in distinguishing Parkinson's disease patients from healthy people;

[0044] The P value and Cohen's d value effect size of each feature are considered comprehensively to screen out key gait features that can significantly distinguish Parkinson's disease patients from healthy people.

[0045] The second aspect of the present application provides a Parkinson disease gait recognition method based on millimeter wave radar, a key gait feature extraction method based on millimeter wave radar is used to obtain a model training set and a test set constructed by the key gait feature, a classifier model is trained by using the training set and the test set, the test set sample is classified and recognized by using the trained classifier model, and a recognition result is obtained.

[0046] Further, the training of the classifier model comprises the following steps:

[0047] The model training set and the test set are constructed based on the key gait feature;

[0048] The training set and the test set are used to train and verify n kinds of classifier models at the same time by using a five-fold cross-validation method;

[0049] One or more key performance indicators, such as cross-validation accuracy, sensitivity, specificity, F1 score and area under the receiver operating characteristic curve, are calculated for each classifier model respectively;

[0050] The key performance indicators are used to select the classifier model with the best classification performance as the final classifier model.

[0051] Further, the classifier model comprises one or more of the following: logistic regression, support vector machine (SVM), random forest, gradient boosting tree, XGBoost, CatBoost and K-nearest neighbor.

[0052] The third aspect of the present application provides a Parkinson disease gait recognition system based on millimeter wave radar, comprising a memory and a processor, the memory comprises a Parkinson disease gait recognition method based on millimeter wave radar program, and the steps of the Parkinson disease gait recognition method based on millimeter wave radar are implemented when the Parkinson disease gait recognition method based on millimeter wave radar program is executed by the processor.

[0053] Compared with the prior art, the beneficial effects of the technical scheme of the present application are:

[0054] The present application does not depend on a multi-sensor wearable system, introduces an adjacent column constraint rule into the trunk speed estimation by improving the trunk path extraction algorithm to extract the trunk speed curve from the preprocessed Doppler spectrum, which has better effect on the speed mutation of Parkinson patients and can more effectively and accurately segment the straight walking stage; the envelope speed curve of the Doppler spectrum is extracted as the arm and foot movement feature, and a multi-dimensional gait feature system containing absolute speed, relative speed, phase time difference and the like is constructed in combination with the trunk speed curve. According to statistical verification, such envelope features have significant effect in distinguishing Parkinson patients from healthy people. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to make the purposes, technical solutions of the present application clearer, the present application provides the following figures and makes the following description:

[0056] Figure 1 The method flowchart provided for the embodiment of the present application is shown in the figure;

[0057] Figure 2 The algorithm flowchart for generating complete range spectrum and Doppler spectrum provided for the embodiment of the present application is shown in the figure;

[0058] Figure 3 The Doppler spectrum diagram provided for the embodiment of the present application is shown in the figure;

[0059] Figure 4 The range spectrum diagram provided for the embodiment of the present application is shown in the figure;

[0060] Figure 5 The Doppler diagram when walking normally provided for the embodiment of the present application is shown in the figure;

[0061] Figure 6 The Doppler diagram when walking without swinging arms provided for the embodiment of the present application is shown in the figure;

[0062] Figure 7 The envelope signal and trunk signal extracted from the Doppler spectrum provided for the embodiment of the present application are shown in the figure;

[0063] Figure 8 The feature economics analysis result diagram provided for the embodiment of the present application is shown in the figure;

[0064] Figure 9 The ROC curve comparison diagram of each model provided for the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0065] In order to make the purposes, technical solutions of the present application clearer, the present application provides the following figures and makes the following description:

[0066] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the protection scope of the present application is not limited by the specific embodiments disclosed below.

[0067] Embodiment 1:

[0068] The present application provides a Parkinson disease gait feature extraction method based on millimeter wave radar, as shown in the figure Figure 1 The Parkinson disease gait feature extraction method flowchart based on millimeter wave radar is shown in the figure, and the specific steps are as follows:

[0069] S1: Acquire millimeter-wave radar signals, preprocess the signals, and obtain the Doppler spectrum and range spectrum of the signals.

[0070] More specifically, preprocessing millimeter-wave radar signals includes the following steps:

[0071] The FMCW millimeter-wave radar transmits linear frequency modulated continuous wave signals to the human body being detected, and collects radar echo signals during the human body's movement.

[0072] The frequency of a linear frequency modulated continuous wave increases linearly with time; the signal frequency... The expression is as follows:

[0073]

[0074] Where Tc is the pulse width of the signal, and B is the bandwidth of the radar waveform. This is the initial frequency.

[0075] Radar transmission signal The expression is as follows:

[0076]

[0077] Where AT represents the amplitude of the transmitted signal, and the signal received by the radar after a time delay caused by distance. The expression is as follows:

[0078]

[0079] in, For time delay, =2d / c, where d is the radial distance between the object and the radar, which changes continuously, and c is the speed of light. This refers to the attenuation of the signal amplitude during propagation.

[0080] The received radar echo signal is coherently mixed with the transmitted signal, and then the intermediate frequency signal is obtained by in-phase / quadrature demodulation. The signal is then sampled by analog-to-digital conversion (ADC) to output a discrete sequence of intermediate frequency signal.

[0081] At the radar receiver, the received signal is coherently mixed with the transmitted signal, and the intermediate frequency signal is obtained after I / Q demodulation. The expression is as follows:

[0082]

[0083] Where Aif represents the intermediate frequency signal amplitude. This is the intermediate frequency. is the phase of the intermediate frequency signal. From the formula, it can be obtained that the distance has a linear relationship with the frequency of the intermediate frequency signal, and the relationship between the subtle change of the distance of the measured target and the radar and the phase is , that is, a linear relationship with the phase difference. Therefore, the distance information of the target can be analyzed by analyzing the frequency component in the intermediate frequency signal, the transmission signal of the radar is composed of a frame of multiple chirps, and the intermediate frequency signal of a single chirp is sampled by ADC to obtain a discrete sequence Sif[n].

[0084] The intermediate frequency signal discrete sequence is converted into a distance spectrum by using fast Fourier transform (FFT), and a two-dimensional fast Fourier transform (FFT) is performed on the phase change between multiple chirps to obtain a Doppler spectrum.

[0085] One-dimensional FFT is performed on the sampling data of each chirp to generate a distance spectrum of each chirp , and the expression is as follows:

[0086]

[0087] Wherein, N is the number of sampling points, k is the frequency index, the distance resolution , and fs is the sampling rate. On the basis of the distance spectrum of the M chirps in a frame of data, the speed information of the target can be analyzed by analyzing the phase change between multiple chirps. The M chirps are continuously transmitted, each chirp corresponds to a column of distance spectrum data, and an N*M matrix is formed. Two-dimensional FFT is performed on the phase change of each column of distance units (the same distance gate) to obtain a Doppler spectrum of each frame, and the expression is as follows:

[0088]

[0089] Wherein, m is the Doppler frequency index, the speed resolution , is the wavelength, and Tframe is the frame time. Therefore, the speed calculation formula is:

[0090]

[0091] According to the speed calculation formula, the Doppler spectrum is multiplied by the speed resolution to obtain a speed spectrum. The distance spectrum is easily affected by the environment, in order to extract the distance information of the human body from the complex environment, it is necessary to remove the static noise, the present application uses a neighboring frame difference calculation method to subtract the value of the previous frame corresponding to the distance spectrum of the next frame, so that the noise caused by the static object can be significantly removed, and the distance information of the dynamic target is maximized. The distance spectrum after the neighboring frame difference calculation is applied to the Hanning window to suppress the spectrum leakage, and then two-dimensional FFT is performed and FFT compensation and Hanning window compensation are added to obtain the final Doppler spectrum.

[0092] In order to improve the strength and accuracy of the signal, the radar antenna parameter setting adopts the Time Division Multiple Access - Multiple Input Multiple Output (TDMA-MIMO) technology, the antenna configuration is 3 transmitting antennas and 4 receiving antennas, 3 transmitting antennas transmit signals in turn, 4 receiving antennas receive signals at the same time, there are 3*4=12 virtual antennas, and the distance spectrum and Doppler spectrum corresponding to these antennas are combined using the non-coherent accumulation method to obtain stable and reliable distance spectrum and Doppler spectrum.

[0093] In order to further extract the information of the human body from the environment, noise suppression and dynamic region extraction need to be performed on the distance spectrum and Doppler spectrum.

[0094] 1) Noise suppression: Because the radar cannot completely collect the full body information of the human body within 1m, the test object will not be allowed to enter the range within 2m of the radar during collection. Therefore, the points within 1m from the radar are deleted, and the part of the Doppler spectrum with a speed of 0 is also deleted to temporarily remove static noise. In the subsequent dynamic region extraction, the 0 speed information within 1m before and after the human body is located is restored after the human body is located by peak value detection.

[0095] 2) Dynamic region extraction: First, determine whether it is the first frame. If it is the first frame, directly perform peak value detection on the Doppler spectrum to obtain the peak value and its distance index. The values outside 1m before and after the distance index of the distance spectrum are set to zero to delete invalid data. The points outside 1m before and after the distance index of the Doppler spectrum are also set to zero. The 0 speed information within 1m before and after the distance index of the Doppler spectrum is restored. Starting from the second frame, do not perform peak value detection on the entire Doppler spectrum, but only perform peak value detection within 1m before and after the distance index of the peak value of the previous frame. The peak value and its distance index are detected, and the points outside 1m before and after the distance index of the distance spectrum and the Doppler spectrum of this frame are set to zero. The 0 speed information within 1m before and after the distance index of the Doppler spectrum is restored.

[0096] In this way, the distance change and Doppler information change caused by human motion can be greatly retained. The distance spectra of all frames are added and spliced according to the time dimension to obtain a complete distance spectrum, and the Doppler spectra of all frames are added according to the distance dimension to obtain a complete Doppler spectrum, as shown in Figure 2 The algorithm flowchart for generating complete distance spectrum and Doppler spectrum is shown.

[0097] The obtained distance spectrum and Doppler spectrum are normalized, and the results are shown in Figure 3 、 Figure 4 The red box in the figure outlines a straight walking phase.

[0098] S2: extracting a trunk speed curve from the Doppler spectrum of the signal using an adjacent column constraint rule, extracting a trunk position change curve from the range spectrum, and extracting basic gait information from the trunk speed curve and the trunk position change curve.

[0099] The trunk speed presents a curve change when walking, reaches a maximum value when the leg is lifted and stepped out to be perpendicular to the ground, reaches a minimum value when the foot is completely landed, and then continues to step out the other leg, so the trunk speed has a certain periodicity when continuously walking, and therefore the peaks of the trunk signal can be detected to segment each step. In a complex environment, due to the influence of multipath effect, other peaks will appear in the Doppler spectrum, affecting the extraction of the trunk signal. In order to solve this problem, the adjacent column constraint rule is added in the extraction process of the trunk signal, and the trunk speed curve is extracted from the Doppler spectrum of the signal using the adjacent column constraint rule, including the following steps:

[0100] Performing static filtering processing on the Doppler spectrum to obtain a purified Doppler spectrum diagram;

[0101] In the purified Doppler spectrum diagram, the trunk signal path is extracted by using the adjacent column constraint rule, taking the maximum value of each column as the path node, detecting the peak value of the next column only in the preset range above and below the maximum value position of the previous column during path searching, and using the traversal from the last column to the front to ensure that the found path is the path with the strongest energy, and in this embodiment, the preset range is about 1.5 m / s above and below;

[0102] The extracted trunk signal path is subjected to curve smoothing processing by using a five-point cubic smoothing filtering algorithm to suppress burrs and abnormal mutations, and a continuous and smooth trunk speed curve is obtained.

[0103] In this embodiment, since the data of each patient walking two laps in a space of 4m*12m is collected, it is necessary to segment each straight walking phase first. In order to make the data accurate and reliable, the walking time of the straight walking phase is required to be greater than 2 seconds, and the speed throughout the journey is not reduced to 0.2 m / s (the average walking speed of healthy elderly people is about 0.7 m / s, and the speed threshold is relatively loose considering that Parkinson's disease patients may have slow walking symptoms). Then, the trunk speed curve of each walking phase is subjected to frequency domain analysis to obtain a frequency domain sub-peak, that is, a step frequency, the average time of each step is calculated through the step frequency, basic gait information is extracted from the trunk speed curve and the trunk position change curve, including the following steps:

[0104] Performing frequency domain analysis on the trunk speed curve to obtain a frequency domain sub-peak as a step frequency, and calculating an average step time of each step according to the step frequency;

[0105] Peak detection is performed on the torso speed curve, and the interval between adjacent peaks is set according to the average step time of each step. In this embodiment, the lower limit of the interval between adjacent peaks is set to 0.7 of the average step time of each step. If multiple candidate peaks are detected within the interval, the peak with a more significant amplitude is retained, and the start and end time points of each step are determined;

[0106] The start and end time of each step is mapped to the corresponding torso position change curve, and the basic gait information of each step is calculated, including one or more of the following: step number, step length, speed change of each step, time consumption of each step, turning time, average speed of each straight walking stage, and step frequency.

[0107] S3: Extracting the envelope speed curve corresponding to the sample motion from the Doppler spectrum, and screening out key gait features that can significantly distinguish sample Parkinson's disease patients from healthy people in combination with the basic gait information.

[0108] More specifically, the envelope speed curve corresponding to the sample motion is extracted from the Doppler spectrum, including the following steps:

[0109] The Doppler spectrum data is arranged in time frames to form a two-dimensional matrix structure, where each column represents the energy distribution of different speed intervals at a time;

[0110] The energy of all rows corresponding to absolute values of speed lower than a preset speed threshold is set to zero to remove background interference from static targets such as the ground, walls, etc., and complete the static area energy zeroing;

[0111] Each column of data is cycled and peak detection is performed, filtering peaks below a dynamic energy threshold and redundant peaks with an interval smaller than a first preset threshold from adjacent peaks. In this embodiment, the first preset threshold is 0.12 m / s, and the dynamic energy threshold is calculated as follows: all positive elements are extracted from the region after the static area energy is zeroed, the median of the positive elements is calculated, the median is robust to outliers, which can avoid interference from outliers, the median is multiplied by an adjustment coefficient, and the threshold value is adjusted by modifying the coefficient to obtain the dynamic energy threshold. In this embodiment, the adjustment coefficient is 1.4;

[0112] The first point above the threshold in the top row is selected as the upper envelope point (maximum positive speed) of the column, and the last point below the threshold in the bottom row is selected as the lower envelope point (minimum negative speed) of the column. If no peak satisfying the condition is detected in the current column, it is temporarily recorded as a null value;

[0113] Detecting the speed mutation of the preliminarily constructed envelope point sequence, marking the points with a speed change amplitude exceeding a second preset threshold between adjacent envelope points as abnormal, and replacing the abnormal points with adjacent legal envelope points to improve the continuity of the upper and lower envelope curves, thereby obtaining the preliminary upper envelope curve and the preliminary lower envelope curve, wherein the second preset threshold is 3 m / s in this embodiment;

[0114] Performing median filtering processing on the preliminary upper envelope curve and the preliminary lower envelope curve respectively to remove isolated noise points, and filling in the missing envelope points by using a linear interpolation method, thereby finally constructing the smooth and continuous upper envelope curve and the smooth and continuous lower envelope curve.

[0115] In order to study which part of the body the envelope information is caused by, the present application makes a comparative experiment: the first group walks normally, and the second group controls not to swing arms while walking, and the experimental results are shown in Figure 5 、 Figure 6 , Figure 5 is a Doppler image when walking normally, Figure 6 is a Doppler image when walking without swinging arms, and a straight walking stage is intercepted and analyzed alone, and it can be obviously found from Figure 6 that: there is basically no information on the positive velocity axis when not swinging arms, so the Doppler effect of the part opposite to the walking speed when walking is caused by swinging arms, and the lower envelope when not swinging arms is jointly affected by multiple body parts such as feet. Similarly, swinging arms will also affect the same speed direction, and by comparing Figure 5 、 Figure 6 the upper envelope, it can be obtained that the intensity of the first half of the single wave of the upper envelope when not swinging arms is obviously reduced, and the second half changes little, and the height of the wave peak is also reduced, so the envelope is jointly affected by the arms and the feet. In order to study the difference in speed between the feet and the arms of the Parkinson's disease patients, the present application uses the envelope of the Doppler spectrum to study the speed law, periodicity and symmetry of the feet and the arms when walking.

[0116] The envelope velocity curve is combined with the basic gait information to screen out key gait features that can significantly distinguish the Parkinson's disease patients from the healthy people, including the following steps:

[0117] Performing peak detection and gait segmentation on the envelope velocity curve to obtain an absolute velocity curve reflecting the motion law of the arms and the feet;

[0118] Taking the trunk velocity curve at the same time as a reference, performing difference on the envelope velocity curve to calculate a relative velocity curve of the envelope relative to the trunk, and the trunk signal and the envelope signal are shown in Figure 7 , and the Doppler spectrum displayed is enlarged in the red box in Figure 3 ;

[0119] Peak detection is performed on the envelope velocity curve and the relative velocity curve respectively, the velocity and distance information in each step is extracted, and the time interval between the envelope peak value and the trunk velocity peak value is calculated as a key feature representing gait synchrony;

[0120] Frequency domain analysis is performed on the envelope velocity curve, and the frequency spectrum peak value is extracted as a periodic gait feature;

[0121] Joint analysis of the envelope velocity curve and the relative velocity curve in time domain and frequency domain is performed, and multi-dimensional gait features are extracted, including one or more of the following: absolute velocity change curve of each step, relative velocity change curve of each step, frequency domain secondary peak value of the envelope curve, time interval between the peak value of the absolute velocity and the relative velocity obtained by the envelope curve and the velocity peak value of the trunk signal respectively;

[0122] Statistical processing is performed on the extracted multi-dimensional gait features, and key gait features that can significantly distinguish Parkinson's disease patients from healthy people are screened out.

[0123] The present application utilizes millimeter wave radar to extract a total of 12 gait parameters, including step number, step length, velocity change of each step, time spent in each step, turning time, average speed of each straight walking stage, step frequency, absolute velocity obtained by the envelope curve, relative velocity obtained by the envelope curve, frequency domain peak value of the envelope curve, and time interval between the peak value of the absolute velocity and the relative velocity obtained by the envelope curve and the velocity peak value of the trunk signal (used to study whether the gait cycle of Parkinson's disease patients is different from that of healthy people). Before further analysis, the present application synchronously collects using the wearable device of GYENNO while using millimeter wave radar to collect data, and by comparing the gait parameters of radar and wearable device, it is concluded that the step length error is 3.19 cm and the step time error is 0.9 s, which is sufficient to prove the reliability of the gait parameters extracted by radar. Due to the different lengths of these parameters, in order to further study what specific features have significant differences in Parkinson's disease, the present application performs averaging, median, asynchronous symmetry, first term, minimum value, maximum value, and variance operations on these parameters to obtain their features, and performs statistical analysis on these features to study whether they have significant effects on the classification of Parkinson's disease.

[0124] More specifically, the extracted gait features are statistically processed, and key gait features that can significantly distinguish Parkinson's disease patients from healthy people are screened out, including the following steps:

[0125] Feature extraction is performed on each original gait feature, and its average, median, first term, maximum value, minimum value, variance and asynchronous symmetry are calculated to obtain a set of statistical features;

[0126] performing principal component analysis (PCA) on the set of statistical features to obtain a low-dimensional representation vector;

[0127] In the embodiment, data of two rounds of walking of the target is collected during data collection, so there are four straight walking stages in total. In order to combine the gait parameters of the four stages, the method of principal component analysis is selected to combine the features of each parameter after feature extraction for subsequent statistical analysis. The present application collects 157 patients with early Parkinson's disease and 89 healthy people as a control group. The control group and the Parkinson's group are matched in gender and age. The statistical analysis results are shown in Figure 8

[0128] performing normality test on the low-dimensional representation vector to determine whether the low-dimensional representation vector obeys normal distribution;

[0129] performing independent sample T test on the features obeying normal distribution, and performing Mann-Whitney U test on the features not obeying normal distribution, to respectively perform significance analysis of differences between groups, and obtain significance level P value;

[0130] calculating Cohen's d effect size of each feature to quantify the difference strength between the two groups of samples, and obtaining the performance of each feature in distinguishing Parkinson's disease patients from healthy people;

[0131] comprehensively considering the P value and Cohen's d effect size of each feature, in the embodiment, the key gait features that can significantly distinguish Parkinson's disease patients from healthy people are screened out according to the feature parameters that satisfy P<0.05 and d>0.5 at the same time. The most significant feature list of each parameter is shown in Table 1.

[0132] Table 1

[0133]

[0134] From the table, it can be concluded that the gait parameters extracted by the radar can significantly distinguish Parkinson's disease patients from healthy people. The walking speed of the Parkinson's disease patients is significantly lower than that of the healthy people, and the foot speed is also significantly lower than that of the healthy people. At the same time, the step length is shorter and the turning time is longer.

[0135] Embodiment 2:

[0136] ​The embodiment provides a Parkinson disease gait recognition method based on a millimeter wave radar.

[0137] More specifically, the training of the classifier model comprises the following steps:

[0138] The model training set and the test set are constructed based on the key gait features, in the embodiment, the key features include step length, step frequency, absolute speed, relative speed, speed peak value, phase interval and other parameter variables which are outstanding in statistical significance and effect strength;

[0139] The training set and the test set are used to train and verify n kinds of classifier models simultaneously in combination with a five-fold cross-validation method, so as to evaluate the stability and generalization ability of the key gait features in different model structures, in the embodiment, n is 7, and the classifier models include: logistic regression, support vector machine (SVM), random forest, gradient boosting tree, XGBoost, CatBoost and K nearest neighbor (KNN);

[0140] The following one or more key performance indicators of each classifier model are calculated respectively: cross-validation accuracy, sensitivity, specificity, F1 score and area under the receiver operating characteristic curve (AUC), which are used to evaluate the classification effect of the model in the Parkinson disease patient recognition task;

[0141] The key performance indicators are used to select the classifier model with the optimal classification performance as the final classifier model, and the comparison results of the performance indicators of each model are shown in Table 2, and the comparison of the ROC curves is shown in Figure 9

[0142] Table 2

[0143]

[0144] From Figure 9 ​The experimental results of the ROC curve comparison chart and the performance index table of Table 2 can see that the classification performance of all models is excellent: the area under the ROC curve (AUC) of logistic regression, support vector machine, random forest, XGBoost, CatBoost and K-Nearest Neighbors is more than 0.92, among which support vector machine (0.983) and CatBoost (0.991) are close to perfect classification, followed by random forest (0.974) and logistic regression (0.969). The cross-validation accuracy is generally higher than 82.5%, and the stability of logistic regression (87.7% ± 0.060) and CatBoost (86.6% ± 0.013) is outstanding; the test set accuracy of random forest (95.9%) and CatBoost (94.6%) is the best, and the accuracy of logistic regression (89.2%) is also close to 90%. In addition, the specificity of random forest is 100%, the specificity (97.9%) and sensitivity (88.9%) of CatBoost are balanced, and the F1 score is as high as 94.1% (random forest). These results fully verify the effectiveness of the gait parameters extracted based on millimeter wave radar in identifying early Parkinson's patients, and the model generalization ability and classification reliability are strongly supported.

[0145] The present application does not rely on a multi-sensor wearable system, and by improving the trunk path extraction algorithm, introduces adjacent column constraint rules in the trunk speed estimation to extract the trunk speed curve from the preprocessed Doppler spectrum. This processing has better effect on the speed mutation of Parkinson's patients, and can more effectively and accurately segment the straight walking stage; by extracting the Doppler spectrum envelope speed curve as the arm and foot movement features, a multi-dimensional gait feature system containing absolute speed, relative speed, phase time difference, etc. is constructed in combination with the trunk speed curve. According to statistical verification, such envelope features have significant effect in distinguishing Parkinson's disease patients from healthy people.

[0146] Embodiment 3:

[0147] The embodiment provides a Parkinson's disease gait recognition system based on millimeter wave radar, including a memory and a processor, the memory includes a Parkinson's disease gait recognition method based on millimeter wave radar program, and the Parkinson's disease gait recognition method based on millimeter wave radar program is executed by the processor to realize the steps of the Parkinson's disease gait recognition method based on millimeter wave radar as described in embodiment 1.

[0148] Obviously, the above embodiments of the present application are merely exemplary but not intended to limit the embodiments of the present application. Based on the above description, any other variations or changes can be made by those skilled in the art without departing from the spirit and principles of the present application. It is not necessary to list all the embodiments here. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application shall fall within the scope of the claims of the present application.

Claims

1. A method for extracting gait features of Parkinson's disease based on millimeter wave radar, characterized in that, The method comprises the following steps: Collecting millimeter wave radar signals, pre-processing the signals to obtain Doppler spectrum and range spectrum of the signals; Extracting a torso speed curve from the Doppler spectrum of the signals by using an adjacent column constraint rule, extracting a torso position change curve from the range spectrum, and extracting basic gait information from the torso speed curve and the torso position change curve; Extracting an envelope speed curve corresponding to sample motion from the Doppler spectrum, and screening key gait features in combination with the basic gait information; The step of extracting the torso speed curve from the Doppler spectrum of the signals by using the adjacent column constraint rule comprises the following steps: Performing static filtering on the Doppler spectrum to obtain a purified Doppler spectrum graph; In the purified Doppler spectrum graph, a torso signal path is extracted by using the adjacent column constraint rule, the maximum value of each column is taken as a path node, the peak value of the next column is detected only within a preset range above and below the maximum value position of the previous column during path searching, and a traversal from the last column to the front is adopted to ensure that the found path is the strongest path; The extracted torso signal path is subjected to curve smoothing processing by using a five-point cubic smoothing filtering algorithm to suppress burrs and abnormal mutations, and a continuous and smooth torso speed curve is obtained; The step of screening the key gait features by using the envelope speed curve in combination with the basic gait information comprises the following steps: Performing peak value detection and gait segmentation on the envelope speed curve to obtain an absolute speed curve reflecting the motion law of arms and feet; Taking the torso speed curve at the same time as a reference, the envelope speed curve is differentiated to calculate a relative speed curve of the envelope relative to the torso; The peak value detection is performed on the envelope speed curve and the relative speed curve respectively, the speed and distance information in each step are extracted, and the time interval between the envelope peak value and the torso speed peak value is calculated as a key feature representing gait synchronism; The envelope speed curve is subjected to frequency domain analysis to extract a frequency spectrum peak value as a periodic gait feature; The envelope speed curve and the relative speed curve are subjected to joint analysis in the time domain and the frequency domain to extract multi-dimensional gait features; The extracted multi-dimensional gait features are subjected to statistical processing to screen key gait features that can significantly distinguish Parkinson's disease patients from healthy people.

2. The millimeter wave radar-based Parkinson's disease gait feature extraction method according to claim 1, characterized in that, The pre-processing of the millimeter wave radar signals comprises the following steps: A linear frequency modulation continuous wave signal is emitted by a millimeter wave radar to a detected human body, and radar echo signals in the motion process of the human body are collected; After the received radar echo signals are mixed with the emitted signals in a phase-locked manner, intermediate frequency signals are obtained through in-phase / quadrature demodulation, and the intermediate frequency signals are sampled through analog-to-digital conversion to output intermediate frequency signal discrete sequences; The intermediate frequency signal discrete sequences are converted into range spectrum by using fast Fourier transform, and two-dimensional fast Fourier transform is performed on the phase changes among multiple chirp signals to obtain Doppler spectrum; The range spectrum is subjected to adjacent frame difference operation, and windowing is used to suppress spectrum leakage, and non-coherent accumulation is used in combination with multi-receiving antenna data to obtain enhanced range spectrum and Doppler spectrum. 3.The millimeter wave radar-based Parkinson disease gait feature extraction method of claim 1, wherein, The step of extracting basic gait information from the torso speed curve and the torso position change curve comprises the following steps: Frequency domain analysis is performed on the torso speed curve to obtain a frequency domain sub-peak as a step frequency, and the average step time of each step is calculated according to the step frequency; Peak detection is performed on the torso velocity curve, and the interval between adjacent peaks is set according to the average step time of each step. If multiple candidate peaks are detected in the interval, the peak with a more significant amplitude is retained, and the start and end time points of each step are determined. The start and end time of each step is mapped to the corresponding torso position change curve to calculate the basic gait information of each step. 4.The method of claim 1, wherein, The envelope velocity curve corresponding to the sample motion is extracted from the Doppler spectrum, including the following steps: The Doppler spectrum data is arranged in time frames to form a two-dimensional matrix structure, where each column represents the energy distribution of different speed intervals at a time. The energy of all rows corresponding to absolute values of speed below the preset speed threshold is set to zero, completing the static region energy zeroing. Each column of data is cycled and peak detection is performed, filtering peaks below the dynamic energy threshold and redundant peaks with an interval less than the first preset threshold. The dynamic energy threshold is calculated by extracting all positive elements from the region after static region energy zeroing, calculating the median of the positive elements, and multiplying the median by an adjustment coefficient to obtain the dynamic energy threshold. The first point above the threshold in the top row is selected as the upper envelope point of the column, and the last point below the threshold in the bottom row is selected as the lower envelope point of the column. If no peak meeting the conditions is detected in the current column, it is temporarily recorded as a null value. Velocity mutation detection is performed on the preliminary envelope point sequence, and points with a velocity change amplitude exceeding the second preset threshold between adjacent column envelope points are marked as abnormal and replaced by adjacent valid envelope points to obtain preliminary upper and lower envelope curves. Median filtering is performed on the preliminary upper and lower envelope curves to remove isolated noise points, and linear interpolation is used to fill in missing envelope points to finally construct smooth and continuous upper and lower envelope curves. 5.The millimeter wave radar-based Parkinson disease gait feature extraction method of claim 1, wherein, Statistical processing is performed on the extracted gait features to select key gait features, including the following steps: Feature extraction is performed on each original gait feature to calculate its mean, median, first term, maximum value, minimum value, variance, and asynchronous symmetry, resulting in a set of statistical features. Principal component analysis is performed on the set of statistical features to obtain a low-dimensional representation vector. Normality test is performed on the low-dimensional representation vector to determine whether it follows a normal distribution. For features that follow a normal distribution, independent sample T-test is used, and for features that do not follow a normal distribution, Mann-Whitney U-test is used for inter-group difference significance analysis to obtain the significance level P value. The Cohen's d value effect size is calculated for each feature to obtain the effectiveness of each feature in distinguishing Parkinson's disease patients from healthy people. The P value and Cohen's d value effect size of each feature are considered to select key gait features that can significantly distinguish Parkinson's disease patients from healthy people. 6.A millimeter wave radar-based Parkinson's disease gait recognition system, characterized by, The system comprises a memory and a processor, the memory comprises a millimeter wave radar-based Parkinson disease gait recognition method program, when the millimeter wave radar-based Parkinson disease gait recognition method program is executed by the processor, the steps of a millimeter wave radar-based Parkinson disease gait recognition method are realized, and the method steps comprise the following steps: constructing a training set and a test set based on key gait features obtained by the millimeter wave radar-based Parkinson disease gait feature extraction method in any one of claims 1-5, training a classifier model by using the training set and the test set, classifying and recognizing test set samples by using the trained classifier model, and obtaining a recognition result.

7. The millimeter wave radar-based Parkinson's disease gait recognition system of claim 6, wherein, The training of the classifier model comprises the following steps: constructing a training set and a test set based on key gait features; training and verifying n kinds of classifier models simultaneously by using the training set and the test set in combination with a five-fold cross-validation method; calculating one or more key performance indicators of each kind of classifier model, such as cross-validation accuracy, sensitivity, specificity, F1 score and area under the receiver operating characteristic curve; selecting a classifier model with the optimal classification performance as the final classifier model by using the key performance indicators.

Images