Aging gait detection method and system based on millimeter wave radar

Abstract

The invention discloses an aging gait detection method and system based on a millimeter wave radar, and the method comprises the following steps: S10: employing the millimeter wave radar to collect the data of the activities of a human body going up and down stairs; s20, performing Fourier transform on the activity data after sampling to obtain signal frequency spectrum and target distance information; wherein the continuous Chirp signals of which the intervals are fixed values are stored as row vectors, and receiving frequency responses of targets at different distances can be obtained by performing 1D-FFT on each row; performing fast Fourier transform on a plurality of Chirp signals of the same distance unit to obtain target speed information, namely a time-Doppler graph; and S30, constructing a YOLOv11 model, constructing a data set by the time-Doppler maps, carrying out classification training by using the YOLOv11 model, carrying out batch classification prediction on the new time-Doppler maps, and finally outputting the prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal detection, in particular to a method and system for detecting aging gait based on millimeter wave radar. BACKGROUND

[0002] Aging will significantly impair human motor function, and the change in leg movement is most obvious when climbing stairs, specifically manifested as abnormal lower limb gait, which has become a key information for detecting the aging process. Based on this core cognition, the present application determines to collect gait data of users when climbing stairs as the detection basis.

[0003] In the field of gait detection, traditional methods are mainly divided into two categories: one is gait detection technology based on wearable devices, and the other is gait detection technology based on visual perception. However, both of these two technologies have obvious limitations: the former requires users to wear specific sensors, which not only may bring inconvenience to users in daily life, but also may interfere with their normal activities; the latter is greatly affected by environmental factors, and the recognition accuracy will be greatly reduced in strong light, shadow, occlusion and other scenes, and also has the problem of insufficient privacy protection. Therefore, there is an urgent need for a new type of gait detection method that is non-contact, has good privacy and strong environmental adaptability to meet the actual needs of gait detection in the aging process. SUMMARY

[0004] The present application proposes a method and system for detecting aging gait based on millimeter wave radar to solve the above problems and meet the needs, which can achieve the above technical purposes and bring other technical effects due to the following technical features.

[0005] One object of the present application is to propose a method for detecting aging gait based on millimeter wave radar, which comprises the following steps: S10: using millimeter wave radar to collect human activity data when climbing stairs; S20: Fourier transform is performed on the sampled activity data to obtain signal frequency spectrum and target distance information; wherein, continuous Chirp signals with a fixed interval are stored as row vectors, and each row can obtain the receiving frequency response of different distance targets through 1D-FFT; the fast Fourier transform is performed on multiple Chirp signals in the same distance unit to obtain target speed information, i.e. time-Doppler graph; S30: building YOLOv11 model, constructing dataset from time-Doppler graph, using YOLOv11 model for classification training, and batch classification prediction on new time-Doppler graph, and finally outputting the accuracy of prediction.

[0006] In addition, the method for detecting aging gait based on millimeter wave radar according to the present application can also have the following technical features: In an example of the present application, after the step S20, further comprising: using the MTI method to process the target speed information to suppress environmental static interference, improve signal-to-noise ratio, distinguish gait key actions from non-target motion, and enhance low-speed gait signals to avoid feature loss.

[0007] In an example of the present application, the MTI method uses a two-pulse canceller for MTI filtering, and its expression is: In the formula, is the output signal of the two-pulse canceller, is the input signal; is the delay time.

[0008] In an example of the present application, in the step S30, the YOLOv11 model specifically comprises: The backbone network is based on the C2f module, efficiently extracts multi-scale features in the input image through residual connection and multi-branch design, covers low-order details and high-order semantics, the network enhances the receptive field through the SPPF module, and introduces a spatial attention module to further improve the spatial context expression ability of the feature map; The neck network adopts a multi-scale feature fusion structure PAFPN, realizes the transmission of high-level semantics to low-level through a top-down feature pyramid path, and fuses the detailed information from the bottom up, so that the model has both detail perception and semantic understanding ability; The detection head is composed of 1-2 convolution layers, a global average pooling layer and a full connection layer, the convolution layer performs weighted calculation on the feature map through a sliding convolution kernel, while retaining the strong energy patches of the aging gait and the stable energy band of the healthy gait, the high-dimensional feature map is compressed to low-dimensional; the global average pooling layer calculates the average value of the convolution feature map channel by channel, and aggregates the spatially distributed feature information into a feature vector; the full connection layer maps the aggregated feature vector to a class score through weight connection, and each output node corresponds to a gait class.

[0009] In an example of the present application, in the step S30, the YOLOv11 model is used for classification training, specifically comprising the following steps: S31: dividing the data set into a training set and a validation set according to a proportion; S32: inputting the training set into the YOLOv11 model for model training, and after the model training is completed, inputting the validation set into the trained YOLOv11 model, and evaluating the model performance on the validation set; S33: using the trained YOLOv11 model to perform batch classification prediction operation on new image data, and outputting the accuracy rate of batch prediction.

[0010] In one example of the present application, the evaluation of the model performance on the validation set includes: a training loss indicator, a validation loss indicator and an accuracy rate.

[0011] Another object of the present application is to provide a millimeter wave radar-based aging gait detection system, comprising: A data acquisition module configured to use a millimeter wave radar to collect human stair climbing and descending activity data; A data processing module configured to perform Fourier transform on the sampled activity data to obtain signal spectrum and target distance information; wherein the continuous Chirp signals with a fixed interval are stored as a row vector, and each row can obtain the receiving frequency response of different distance targets through 1D-FFT; the fast Fourier transform is performed on multiple Chirp signals in the same distance unit to obtain target speed information, i.e. time-Doppler graph; A data training module configured to build a YOLOv11 model, construct a data set from the time-Doppler graph, use the YOLOv11 model for classification training, and perform batch classification prediction on the new time-Doppler graph, and finally output the prediction accuracy.

[0012] In one example of the present application, it further comprises a noise suppression module configured to use the MTI method to process target speed information to suppress environmental static interference to improve signal-to-noise ratio, distinguish gait key actions from non-target motion, and enhance low-speed gait signals to avoid feature loss.

[0013] In one example of the present application, the YOLOv11 model comprises: A backbone network comprising a C2f module, an SPPF module and a spatial attention module, wherein the C2f module is configured to efficiently extract multi-scale features in the input image through residual connection and multi-branch design; the SPPF module is configured to enhance the receptive field, and the spatial attention module is configured to further improve the spatial context expression ability of the feature map; A neck network comprising a multi-scale feature fusion structure PAFPN, which is configured to realize the transmission of high-level semantics to low-level through a top-down feature pyramid path, and fuse the detailed information from the bottom up, so that the model has both detail perception and semantic understanding ability; The detection head comprises a convolution layer, a global average pooling layer and a full connection layer, wherein the convolution layer is configured to perform weighted calculation on the feature map through a sliding convolution kernel, and compresses the high-dimensional feature map to a low-dimensional feature map while retaining the strong energy patches of the aging gait and the smooth energy band of the healthy gait; the global average pooling layer is configured to perform average value calculation on the convolution feature map channel by channel, and aggregate the spatially distributed feature information into a feature vector; and the full connection layer is configured to map the aggregated feature vector to a class score through weight connection, and each output node corresponds to a gait class.

[0014] In an example of the present application, the data training module comprises the following steps: A data division unit configured to divide the data set into a training set and a validation set according to a proportion; A data training unit configured to input the training set into a YOLOv11 model for model training, and after the model training is completed, input the validation set into the trained YOLOv11 model to evaluate the performance of the model on the validation set; A data prediction unit configured to use the trained YOLOv11 model to perform batch classification prediction operation on new image data, and output the accuracy rate of batch prediction.

[0015] Compared with the prior art, the present application has the following advantages: The present application collects the gait data of going up and down the stairs by cooperating the millimeter wave radar and the DCA1000EVM, analyzes the distance and speed information through 1D-FFT and 2D-FFT, filters out the static interference and strengthens the effective gait signal by using the MTI technology, and finally constructs the data set based on the time-Doppler graph, and classifies and trains by using the YOLOv11, which effectively solves the problems of low classification accuracy caused by the interference of the contact type device to the natural gait, the influence of the visual device on the light shielding, and the noise residue after signal processing in the traditional gait detection. The method has high environmental adaptability, and shows stronger robustness to light conditions and shielding conditions, can maintain high detection accuracy in complex environment, and reduces misjudgment. At the same time, the non-contact detection method avoids the invasion of user privacy, reduces the psychological burden of the user, and improves the user acceptance. Moreover, the YOLOv11 has high processing speed, can efficiently process the gait data and quickly feedback the detection result, and lays a solid foundation for subsequent realization of dynamic monitoring of gait changes in the aging process.

[0016] The optimal embodiments for implementing the present application will be described in more detail below with reference to the accompanying drawings, so that the features and advantages of the present application can be easily understood. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings of the embodiments of the present application will be briefly introduced below. Among them, the drawings are only used to show some embodiments of the present application, and not to limit all embodiments of the present application to this.

[0018] Figure 1 Flow chart of a millimeter wave radar-based aging gait detection method according to an embodiment of the present application; Figure 2 Structure diagram of a YOLOv11 model according to an embodiment of the present application; Figure 3 Gait heat map obtained by 2D-FFT according to an embodiment of the present application ((a) does not contain filtering, (b) contains filtering); Figure 4 Structure diagram of a double-pulse cancellation used in MTI filtering according to an embodiment of the present application; Figure 5 Algorithm experimental result diagram of a YOLOv11 model according to an embodiment of the present application; Figure 6 Test effect diagram after model training according to an embodiment of the present application. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the technical solutions of the present application more clear, the technical solutions of the embodiments of the present application will be clearly and completely described below in combination with the drawings of the embodiments of the present application. The same reference signs in the drawings represent the same parts. It should be noted that the described embodiments are part of the embodiments of the present application, not all embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0020] Unless otherwise defined, technical terms or scientific terms used herein shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terms "first", "second", and similar terms, used herein do not necessarily have any sequence or order, unless otherwise defined, but are used for the purpose of nomenclature. Similarly, the terms "one", "another", and similar terms, do not necessarily refer to the same article, unless otherwise defined. The terms "comprising", "comprise", and similar terms, mean including the elements or objects listed thereafter and not excluding other elements or objects. The terms "connected", "coupled", and similar terms, do not necessarily mean physically or mechanically connected or coupled, but can include electrical connection or coupling, whether direct or indirect. The terms "upper", "lower", "left", "right", and similar terms, are used only to indicate relative positions, and when the absolute positions of the described objects are changed, the relative positions may also be changed accordingly.

[0021] According to a millimeter wave radar-based aging gait detection method according to a first aspect of the application, as shown in the accompanying drawings, the method comprises the following steps: Figure 1 S10: Using a millimeter wave radar to collect human activity data when going up and down stairs; for example, selecting an IWR1843 millimeter wave radar, configuring radar parameters, and matching a DCA1000EVM; S20: After sampling the activity data, performing Fourier transform (1D-FFT) to obtain signal spectrum and target distance information; wherein, continuous Chirp signals with a fixed interval are stored as a row vector, and each row is subjected to 1D-FFT to obtain the receiving frequency response of different distance targets; a plurality of Chirp signals in the same distance unit are subjected to fast Fourier transform (2D-FFT) to obtain target speed information; S30: Building a YOLOv11 model, constructing a data set from a time-Doppler diagram, using the YOLOv11 model for classification training, and performing batch classification prediction on a new time-Doppler diagram, and finally outputting the prediction accuracy.

[0022] ​The detection method cooperates with the IWR1843 millimeter wave radar and the DCA1000EVM to collect gait data on and off the stairs, analyzes distance and speed information through 1D-FFT and 2D-FFT, filters out static interference and strengthens effective gait signals through the MTI technology, finally constructs a dataset based on a time-Doppler diagram, and classifies and trains through YOLOv11, effectively solving the problems of interference of natural gait by contact type equipment, influence of visual equipment on light and shielding, and low classification accuracy caused by noise residue after signal processing in the traditional gait detection. The method has high environmental adaptability, shows stronger robustness to light conditions and shielding conditions, can maintain high detection accuracy in complex environments, and reduces misjudgment. At the same time, the non-contact detection method avoids the invasion of user privacy, reduces the psychological burden of the user, and improves the user acceptance. Moreover, the processing speed of YOLOv11 is fast, which can efficiently process gait data and quickly feedback the detection result, laying a solid foundation for subsequent realization of dynamic monitoring of gait changes in the aging process.

[0023] In an example of the present application, after the step S20, further comprising: using an MTI method to process target speed information to suppress environmental static interference to improve signal-to-noise ratio, distinguish gait key actions from non-target motion, and enhance low-speed gait signals to avoid feature loss.

[0024] In an example of the present application, the MTI method uses a two-pulse canceller for MTI filtering, as shown in Figure 4 The expression is: In the formula, is the output signal of the double-pulse canceller, is the input signal; is the delay time.

[0025] The output can also be expressed as: When the input is a unit pulse, the corresponding is a unit step function, the system impulse response is: The Fourier transform of is the system frequency response: .

[0026] In an example of the present application, as shown in Figure 2 in the step S30, the YOLOv11 model is built, specifically comprising: The backbone network is based on the C2f module, and through residual connection and multi-branch design, it efficiently extracts multi-scale features in the input image, covering low-order details (such as edges, local textures) and high-order semantics (such as overall patterns, energy trends). The network enhances the receptive field through the SPPF module and introduces a spatial attention module to further improve the spatial context expression ability of the feature map. In the stage of extracting features by the backbone network, the C2f module used by YOLOv11 will extract "from shallow to deep" features from the time-Doppler graph. The first few layers of the network, such as convolutional layers and pooling layers, mainly extract low-order features, which correspond to the detailed energy distribution in the graph, such as the shape of the strong energy patch in the 0-1s of the aging gait graph and the initial energy peak shape near 0s of the healthy gait graph. They belong to "edge, texture level" details, similar to local motion features of human gait. The last few layers of the network, the C2f module extracts high-order semantic features through residual connection and multi-branch structure, which correspond to the overall energy trend of the graph, such as the trend of the aging gait graph showing "strong energy in the early stage and rapid decay in the later stage", and the trend of the healthy gait graph showing "stable energy and long duration". They are "abstract gait pattern level" features and are also the core basis for distinguishing between aging and healthy gait. In addition, the backbone network outputs three feature maps with different resolutions: 56x56, 28x28, and 14x14, which correspond to "detail features" (used to capture the fine shape of energy patches), "medium semantic features" (used to capture the extension law of energy bands), and "global features" (used to capture the energy distribution trend in the entire time range), respectively, preparing for subsequent feature fusion.

[0027] The neck network adopts a multi-scale feature fusion structure PAFPN (Path Aggregation Feature Pyramid Network), which realizes the transmission of high-level semantics to low-level through the top-down feature pyramid path, and fuses the detailed information from the bottom up, so that the model has both detail perception and semantic understanding ability; the process of feature fusion in the neck network (Neck), the Neck part of YOLOv11 adopts PAFPN (Path Aggregation Feature Pyramid Network) structure, and its core function is to integrate features of different levels. The shallow feature has the characteristics of high resolution, rich details but weak semantic information, such as being able to clearly see the edge of each strong energy patch in the aging gait figure; the deep feature has low resolution and strong semantic information but blurred details, such as being able to grasp the trend of "overall energy stability" of the healthy gait figure, but not being able to see the local small patches. Through the "top-down" and "bottom-up" feature fusion path of PAFPN, the detailed features of the shallow layer are injected into the deep layer, and at the same time the semantic features of the deep layer are transmitted to the shallow layer, finally outputting the fusion feature map with "details + semantics". In this way, both the "0-1s strong energy patch shape details" in the aging gait figure and the "early strong and late weak" semantic trend are retained; at the same time, both the "energy band stability details" and the "overall sustained stability" semantic trend in the healthy gait figure are retained, so that the model can pay attention to both "local gait action details" (energy pulse generated by foot swing) and "global gait pattern" (stability of overall movement).

[0028] The detection head is composed of 1-2 convolutional layers, a global average pooling layer, and a fully connected layer. The convolutional layer performs weighted calculation on the feature map through a sliding convolution kernel, compresses the high-dimensional feature map to a low-dimensional feature map while retaining the strong energy patches of the aging gait and the stable energy band of the healthy gait, and avoids subsequent calculation redundancy. The global average pooling layer calculates the average value of the convolutional feature map channel by channel, aggregates the spatially distributed feature information (such as energy changes at different time points) into a feature vector, and the feature vector reflects the overall trend characteristics (such as the energy attenuation rate of the whole period). The fully connected layer maps the aggregated feature vector to a class score through weight connection, and each output node corresponds to a gait class. The fusion feature map output by the detection head is converted into a class probability through the classification head (Head). The classification head is composed of 1-2 convolutional layers, a global average pooling layer, and a fully connected layer. The convolutional layer performs weighted calculation on the feature map through a sliding convolution kernel, compresses the high-dimensional feature map to a low-dimensional feature map while retaining the key features such as the strong energy patches of the aging gait and the stable energy band of the healthy gait, and avoids subsequent calculation redundancy. The global average pooling layer calculates the average value of the convolutional feature map channel by channel, aggregates the spatially distributed feature information (such as energy changes at different time points) into a feature vector, and the feature vector reflects the overall trend characteristics (such as the energy attenuation rate of the whole period). The fully connected layer maps the aggregated feature vector to a class score through weight connection, and each output node corresponds to a gait class. Since it is a binary classification task of "aging gait" and "healthy gait", the output dimension of the fully connected layer is set to 2, corresponding to the two classes of "aging gait" and "healthy gait". The fully connected layer first outputs "original scores (logits)", and then converts them into probability values through the Softmax activation function, and the sum of all class probabilities is 1.

[0029] In one example of the present application, in the step S30, a YOLOv11 model is used for classification training, specifically including the following steps: S31: dividing the data set into a training set and a validation set according to a proportion; S32: inputting the training set into the YOLOv11 model for model training, and after the model training is completed, inputting the validation set into the trained YOLOv11 model to evaluate the model performance on the validation set; S33: using the trained YOLOv11 model to perform batch classification prediction operation on new image data, and outputting the accuracy rate of batch prediction.

[0030] In the training stage, the difference between the "predicted probability" and the "true class label" is calculated through the cross-entropy loss function, and then the loss is back-propagated to update all parameters of the backbone network, Neck and classification head, so that the model gradually learns the mapping relationship of "energy distribution of time-Doppler map → gait class".

[0031] Finally, in the inference stage, the model will choose the class with the highest probability value as the final prediction result. For example, when the input is a time-Doppler plot of a gait of the elderly, the model will output "gait of the elderly"; when the input is a time-Doppler plot of a healthy gait, the model will output "healthy gait".

[0032] The cross-entropy loss function formula is: In the formula, y is the true label, which is 0 or 1, 0 represents a healthy gait, and 1 represents a gait of the elderly; and is the class probability predicted by the model, which is the value processed by the Softmax activation function, and the range is between 0 and 1, representing the probability that the sample belongs to "gait of the elderly"; log represents the natural logarithm.

[0033] The essence of this formula is to measure the deviation between the model's predicted value and the true label. Specifically, if the true label y is 1 (i.e., a gait of the elderly), and the model's predicted is very small (such as 0.1), then log( ) will be a very large negative number, which will increase the loss L, thus punishing the model more strongly. Conversely, if the true label y is 0 (i.e., a healthy gait), but the model's predicted is very large (such as 0.9), then log(1− ) will be a very large negative number, which will also cause the loss L to increase. Only when the model's predicted probability is close enough to the true label y, such as when y = 1 and = 0.9, will the loss be very small, and the model will consider this a relatively "satisfactory" prediction result.

[0034] Backpropagation uses the gradient descent algorithm to update network parameters. For each parameter in the model (such as the weights of the convolutional layer, the weights and biases of the fully connected layer, etc.), the update formula is: In the formula, is the current parameter (such as the weight of a certain convolution kernel, the weight of a fully connected layer); is the updated parameter; is the learning rate; is the partial derivative of the loss function L with respect to the parameter : In the formula, y is the true label, is the class probability predicted by the model.

[0035] In one example of the present application, the evaluation of the model performance on the validation set includes: a training loss indicator, a validation loss indicator, and an accuracy rate.

[0036] According to the second aspect of the present application, a millimeter wave radar-based aging gait detection system comprises: A data acquisition module is configured to use a millimeter wave radar to collect human stair climbing and descending activity data; for example, an IWR1843 millimeter wave radar is selected, radar parameters are configured, and a DCA1000EVM is matched; A data processing module is configured to perform Fourier transform (1D-FFT) on the sampled activity data to obtain signal spectrum and target distance information; wherein continuous Chirp signals with a fixed interval are stored as a row vector, and each row is subjected to 1D-FFT to obtain the receiving frequency response of different distance targets; a plurality of Chirp signals in the same distance unit are subjected to fast Fourier transform (2D-FFT) to obtain target speed information, i.e., a time-Doppler map; A data training module is configured to build a YOLOv11 model, construct a data set from the time-Doppler map, use the YOLOv11 model for classification training, and perform batch classification prediction on new time-Doppler maps, and finally output the prediction accuracy.

[0037] The detection system cooperates with the IWR1843 millimeter wave radar and the DCA1000EVM to collect stair climbing and descending gait data, analyzes distance and speed information through 1D-FFT and 2D-FFT, filters out static interference and enhances effective gait signals using MTI technology, constructs a data set based on the time-Doppler map, and classifies and trains using YOLOv11, effectively solving the problems of traditional gait detection, such as interference of contact-type devices with natural gait, vulnerability of visual devices to light and shielding, and low classification accuracy due to noise remaining after signal processing. This method has high environmental adaptability and stronger robustness to light conditions and shielding, can maintain high detection accuracy in complex environments, and reduces false positives. At the same time, using a non-contact detection method avoids invasion of user privacy, reduces the psychological burden of users, and improves user acceptance. Moreover, YOLOv11 has fast processing speed and can efficiently process gait data and quickly feedback detection results, laying a solid foundation for subsequent dynamic monitoring of gait changes in the aging process.

[0038] In one example of the present application, it also includes a noise suppression module configured to use the MTI method to process target speed information to suppress environmental static interference and improve signal-to-noise ratio, distinguish between gait key actions and non-target movements, and enhance low-speed gait signals to avoid feature loss.

[0039] In one example of the present application, the YOLOv11 model comprises: The backbone network comprises a C2f module, an SPPF module and a spatial attention module, wherein the C2f module is configured to efficiently extract multi-scale features in the input image through residual connection and multi-branch design; the SPPF module is configured to enhance the receptive field, and the spatial attention module is configured to further improve the spatial context expression capability of the feature map. The neck network comprises a multi-scale feature fusion structure PAFPN (Path Aggregation Feature Pyramid Network), which is configured to realize the transmission of high-level semantics to low levels through a top-down feature pyramid path, and fuse the details from the bottom up, so that the model has both detail perception and semantic understanding capabilities. The detection head comprises a convolution layer, a global average pooling layer and a fully connected layer, wherein the convolution layer is configured to perform weighted calculation on the feature map through a sliding convolution kernel, compress the high-dimensional feature map to a low-dimensional feature map while retaining the strong energy patches of the aging gait and the stable energy band of the healthy gait, and avoid subsequent calculation redundancy; the global average pooling layer is configured to calculate the average value of the convolution feature map channel by channel, aggregate the spatially distributed feature information (such as energy changes at different time points) into a feature vector, and the feature vector reflects the overall trend feature (such as the energy attenuation rate of the whole period); and the fully connected layer is configured to map the aggregated feature vector to a class score through weight connection, and each output node corresponds to a gait class.

[0040] In an example of the present application, the data training module comprises the following steps: A data division unit configured to divide the data set into a training set and a validation set according to a proportion; A data training unit configured to input the training set into the YOLOv11 model for model training, and after the model training is completed, input the validation set into the trained YOLOv11 model, and evaluate the model performance on the validation set; A data prediction unit configured to use the trained YOLOv11 model to perform batch classification prediction operation on new image data, and output the accuracy rate of batch prediction.

[0041] It should be noted that the millimeter wave radar-based aging gait detection system of the present application can also perform any processing in the millimeter wave radar-based aging gait detection method as previously described, and the specific details are not repeated here.

[0042] Specific case: The method selects IWR1843 millimeter wave radar as the core perception device. The radar has 3 transmission channels and 4 receiving channels, can accurately capture the micro-motion characteristics of human stair descent gait, and is matched with DCA1000EVM data acquisition card. The radar parameters are configured through the mmWaveStudio platform. The specific parameters are as follows: the number of sampling points is 128, the bandwidth is 3997.56 MHz, the number of Chirps in the frame is 128, the frame period is 40 ms, the sampling frequency is 4000 ksps, and the stair descent gait data of 5 seconds of volunteers of different ages (21 people in the aging group, 50-76 years old; 5 people in the healthy group, 20-25 years old) is collected. Each person collects more than 20 groups, and a total of 579 groups of effective data are obtained.

[0043] After sampling the original echo signal collected by the radar, 1D-FFT is performed to convert the time domain signal to the frequency domain signal to obtain the signal spectrum, and the distance information of the target and the radar is extracted therefrom. The continuously transmitted and fixedly spaced Chirp signals are stored as a row vector, 1D-FFT preprocessing is performed on each row of data to obtain the receiving frequency response of different distance targets, and the distance unit related to human gait is preliminarily screened out. For the same distance unit after 1D-FFT screening, 2D-FFT is performed on multiple Chirp signals in the unit to extract the speed information of human gait by analyzing the Doppler shift characteristics of the signals, and a distance-speed two-dimensional spectrum is generated. However, there are still ground, wall and other static clutter interferences in the spectrum at this time, which need to be further filtered. The results are as shown in Figure 3 (a).

[0044] An MTI filter is constructed by using a double-pulse canceller, and the signals processed by 2D-FFT are filtered based on the difference in Doppler characteristics between moving targets and static backgrounds. By setting a high-attenuation stop band for zero-frequency components and integer multiple frequency points of the pulse repetition period, static clutter is efficiently suppressed, while maintaining low suppression intensity at other frequency points to retain effective motion signals of human gait (such as foot contact and lower limb swing), and to enhance low-speed gait signals, so that gait echo signals can be accurately obtained, and the signal distortion problem in complex environments in traditional methods can be effectively avoided. The effect is as shown in Figure 3 (b).

[0045] Figure 3 In (a), the color gradually changes from red (high energy) to blue (low energy), and there are obvious red and bright yellow patches in the marked areas such as "alpha1=0.95" and "alpha2=0.8". These high-energy areas correspond to specific frequency-wave number energy peaks formed in the frequency domain by effective motion signals such as limb swing and trunk displacement during human stair descent, and are surrounded by relatively weak yellow and blue environmental clutter energy such as stair structure and equipment noise, which reflects the energy mixing state of effective signals and interference in the original signal. Figure 3(b) The color bar value range is significantly expanded from the 10-16 interval of the left figure to the -20-60 interval, and the overall energy distribution is mainly yellow and red, with only a small amount of strong red area on the left, which reflects that the filtering operation suppresses the low-energy components of the environmental clutter (the original blue and dark yellow areas are removed), highlights the energy distribution of the effective motion signal of the human body going down the stairs, and because of the interference being filtered, the effective signal energy is in a more uniform high-energy state (colors mainly yellow and red) in the frequency domain, facilitating the subsequent precise extraction of key parameters such as frequency and wave number of human motion.

[0046] Based on the gait time-Doppler graph as the basic data, the "aging gait" and "healthy gait" two types of labels are labeled to construct the aging gait detection data set, and the data set is divided into training set for model parameter learning and test set for model performance verification according to the ratio of 8:2, ensuring balanced data distribution; under the Windows 10 operating system, Python development tool, PyCharm platform and AMD Ryzen7 5800U CPU, 32GB memory hardware configuration, YOLOv11 deep learning model is used for classification training of data set, YOLOv11 continues and innovates the three-section architecture of Backbone-Neck-Head. Backbone (main network) is based on C2f module, through residual connection and multi-branch design, it efficiently extracts multi-scale features in input images, covering low-order details (such as edges, local textures) and high-order semantics (such as overall patterns, energy trends). In addition, the model retains the SPPF module to enhance the receptive field, and introduces the spatial attention module (such as C2PSA), further improving the spatial context expression ability of the feature map. Neck (neck network) uses PAFPN (Path Aggregation Feature Pyramid Network). That is, the multi-scale feature fusion structure widely used in the YOLO series, through the top-down feature pyramid path to realize the transmission of high-level semantics to low-level, and fuse the details from bottom to top, so that the model has both detail perception and semantic understanding ability. The detection head (Head) uses a deep separable convolution structure, where the classification branch uses DWConv (deep separable convolution) to reduce computational complexity, the regression branch uses standard convolution and deformable convolution to improve target positioning accuracy, and includes a center-ness branch to suppress low-quality prediction boxes, achieving a good balance between model accuracy, inference speed and computational efficiency. Specifically, it can be divided into the following five steps: 1、Build the basic environment First, create a virtual environment with a specified Python version using Anaconda. Then, download the official source code for YOLOv11. Finally, in the created virtual environment, use the pip command to install all the necessary dependencies for YOLOv11 training, thus completing the setup of the basic environment.

[0047] 2. Prepare the dataset and configuration files First, split the dataset into a training set (train) and a validation set (val) in an 8:2 ratio. Then, create a new folder named "data" in the YOLOv11 source code directory. This folder contains two subfolders named "train" and "val". Each of the "train" and "val" folders should also have two subdirectories to store the "aging gait" and "healthy gait" data, respectively. Finally, create a file named "data.yaml" in the YOLOv11 source code directory. This file should explicitly specify the paths to the training and validation sets, as well as the category information for "aging gait" and "healthy gait".

[0048] 3. Configure the model and training parameters First, obtain the pre-trained YOLOv11 classification model yolo11n-cls.pt. Then, write a training script. The script should include the core content of importing YOLO model-related libraries, initializing and loading the yolo11n-cls.pt pre-trained model, configuring the path of the data.yaml dataset file, and setting training parameters such as the number of training iterations, input image size, and the number of images processed per batch.

[0049] 4. Model Training and Performance Evaluation First, run the prepared training script to start model training based on the configured dataset and parameters. After training is complete, evaluate the model performance on the validation set, focusing on three key metrics during the evaluation: training loss (train_loss), validation loss (val_loss), and Top-1 accuracy (accuracy_top1).

[0050] 5. Batch classification prediction Using the trained model, perform batch classification prediction on new image data, and finally output the accuracy of this batch prediction.

[0051] Training results as follows Figure 5As shown, the figure shows the training and validation indicators of YOLOv11 in the "aging gait and healthy gait" binary classification task: the train / loss curve on the top left continues to decline to near 0, indicating that the model fits the training data and the training converges; the val / loss curve on the top right overall declines and stabilizes at a low level, indicating that the model has good generalization ability and no obvious overfitting; the metrics / accuracy_top1 curve on the bottom left rises rapidly and can reach 0.97, reflecting the high classification accuracy of the model; the metrics / accuracy_top5 on the bottom right is always stable at 1.0 due to the characteristics of the binary classification task, only used to assist in verifying stability. In summary, the model has good training convergence, strong generalization ability, and high classification accuracy in this classification task, and can effectively complete the gait classification task.

[0052] A batch of new data is used to perform batch classification prediction operation using the trained model, and the new batch of data is the gait data of ten new volunteers, with 20-25 gait data graphs per person. The results are as shown in Figure 6 As shown, the accuracy rate means the ratio of the model's prediction of the tester's gait condition consistent with the actual gait, proving the reliability of the method for detecting aging gait.

[0053] Therefore, the method has the following advantages: high signal processing efficiency, clear and easy-to-understand technical logic, and high gait detection accuracy. Through the complete technical process of "1D-FFT+2D-FFT+MTI filtering+YOLOv11 classification" designed to detect human aging and healthy gait signals, without increasing the hardware running load, compared with the traditional gait detection method, this method effectively avoids the problems of contact interference with natural gait, light blocking affecting signal quality, and fuzzy classification boundary.

[0054] The above describes the exemplary embodiments of the aging gait detection method and system based on millimeter wave radar proposed by the present application in detail with reference to the preferred embodiments, however, those skilled in the art can understand that various modifications and improvements can be made to the above specific embodiments without departing from the concept of the present application, and various technical features and structures proposed by the present application can be combined without exceeding the protection scope of the present application, the protection scope of the present application is determined by the appended claims.

Claims

1. A method for detecting aging gait based on millimeter-wave radar, characterized in that, Includes the following steps: S10: Employs millimeter-wave radar to collect data on human activities when going up and down stairs; S20: After sampling the activity data, perform Fourier transform to obtain the signal spectrum and target distance information; among them, the continuous Chirp signals with fixed intervals are stored as row vectors, and the receiving frequency response of targets at different distances can be obtained by 1D-FFT for each row; perform Fast Fourier Transform on multiple Chirp signals of the same distance unit to obtain target velocity information, i.e., time-Doppler plot; S30: Build a YOLOv11 model, construct a dataset from time-Doppler graphs, use the YOLOv11 model for classification training, perform batch classification prediction on the new time-Doppler graphs, and finally output the prediction accuracy.

2. The aging gait detection method based on millimeter-wave radar according to claim 1, characterized in that, The process after step S20 further includes: using the MTI method to process target velocity information to suppress static environmental interference in order to improve the signal-to-noise ratio, distinguishing between key gait movements and non-target movements, and enhancing low-speed gait signals to avoid feature loss.

3. The aging gait detection method based on millimeter-wave radar according to claim 1, characterized in that, The MTI method uses a two-pulse canceller for MTI filtering, and its expression is: In the formula, This is the output signal of the dual-pulse canceller. For input signals; This is the delay time.

4. The aging gait detection method based on millimeter-wave radar according to claim 1, characterized in that, In step S30, building the YOLOv11 model specifically includes: The backbone network is based on the C2f module. Through residual connections and multi-branch design, it efficiently extracts multi-scale features from the input image, covering low-level details and high-level semantics. The network enhances the receptive field through the SPPF module and introduces a spatial attention module to further improve the spatial contextual representation ability of the feature map. The neck network adopts the multi-scale feature fusion structure PAFPN, which realizes the transmission of high-level semantics to low-level through the top-down feature pyramid path, and integrates bottom-up detailed information, so that the model has both detail perception and semantic understanding capabilities. The detection head consists of 1-2 convolutional layers, a global average pooling layer, and a fully connected layer. The convolutional layers use sliding convolutional kernels to perform weighted calculations on the feature maps, compressing high-dimensional feature maps to low dimensions while preserving strong energy patches in aging gait and stable energy bands in healthy gait. The global average pooling layer calculates the average value of each channel of the convolutional feature map, aggregating the spatially distributed feature information into a feature vector. The fully connected layer maps the aggregated feature vector into class scores through weighted connections, with each output node corresponding to a gait class.

5. The aging gait detection method based on millimeter-wave radar according to claim 1, characterized in that, In step S30, the YOLOv11 model is used for classification training, which specifically includes the following steps: S31: Divide the dataset into training and validation sets according to the proportions; S32: Input the training set into the YOLOv11 model for model training. After the model training is completed, input the validation set into the trained YOLOv11 model and evaluate the model performance on the validation set. S33: Using the trained YOLOv11 model, perform batch classification prediction on new image data and output the batch prediction accuracy.

6. The aging gait detection method based on millimeter-wave radar according to claim 5, characterized in that, Evaluating model performance on the validation set includes: training loss metric, validation loss metric, and accuracy.

7. An aging gait detection system based on millimeter-wave radar, characterized in that, include: The data acquisition module is configured to use millimeter-wave radar to collect data on human activities when going up and down stairs; The data processing module is configured to perform Fourier transform on sampled activity data to obtain signal spectrum and target distance information. Specifically, continuous Chirp signals with fixed intervals are stored as row vectors, and each row can obtain the received frequency response of targets at different distances by performing 1D-FFT. Fast Fourier transform is performed on multiple Chirp signals of the same distance unit to obtain target velocity information, i.e., time-Doppler plot. The data training module is configured to build a YOLOv11 model. It constructs a dataset using time-Doppler graphs, performs classification training using the YOLOv11 model, performs batch classification predictions on the new time-Doppler graphs, and finally outputs the prediction accuracy.

8. The aging gait detection system based on millimeter-wave radar according to claim 7, characterized in that, It also includes a noise suppression module, configured to use the MTI method to process target velocity information, suppress static environmental interference to improve the signal-to-noise ratio, distinguish key gait movements from non-target movements, and enhance low-speed gait signals to avoid feature loss.

9. The aging gait detection system based on millimeter-wave radar according to claim 7, characterized in that, The YOLOv11 model includes: The backbone network includes a C2f module, an SPPF module, and a spatial attention module. The C2f module is configured to efficiently extract multi-scale features from the input image through residual connections and multi-branch design; the SPPF module is configured to enhance the receptive field; and the spatial attention module is configured to further improve the spatial contextual representation ability of the feature map. The neck network includes a multi-scale feature fusion structure PAFPN, which is configured to transmit high-level semantics to low-level structures through a top-down feature pyramid path and fuse bottom-up detail information, enabling the model to have both detail perception and semantic understanding capabilities. The detection head includes convolutional layers, global average pooling layers, and fully connected layers. The convolutional layers are configured to perform weighted calculations on the feature maps using sliding convolutional kernels, compressing high-dimensional feature maps to low dimensions while preserving strong energy patches in aging gait and stable energy bands in healthy gait. The global average pooling layers are configured to calculate the average value of each channel of the convolutional feature maps, aggregating spatially distributed feature information into a feature vector. The fully connected layers are configured to map the aggregated feature vector into class scores through weighted connections, with each output node corresponding to a gait class.

10. The aging gait detection system based on millimeter-wave radar according to claim 7, characterized in that, The data training module includes the following steps: The data partitioning unit is configured to divide the dataset into training and validation sets proportionally. The data training unit is configured to input the training set into the YOLOv11 model for model training. After the model training is completed, the validation set is input into the trained YOLOv11 model to evaluate the model performance on the validation set. The data prediction unit is configured to use a trained YOLOv11 model to perform batch classification prediction on new image data and output the accuracy of the batch prediction.

Images