A frozen gait recognition method and system based on acceleration motion data

Abstract

The application discloses a frozen gait recognition method and system based on acceleration motion data and relates to the frozen gait recognition field. The application solves the problems that the existing frozen gait data set is unbalanced in the number of normal gait and frozen gait samples, the model is inclined to normal gait characteristics and ignores frozen gait characteristics during learning, and the recognition effect is affected. The method comprises the following steps: obtaining frozen gait acceleration data of Parkinson's patients, and performing pretreatment to obtain a data sample set containing normal gait and frozen gait; a deep learning frozen gait recognition deep feature extraction model based on a residual network is constructed; a multi-scale convolution module and an attention module are introduced to construct a complete frozen gait recognition model; and the key feature set obtained after processing is subjected to frozen gait recognition on the complete frozen gait recognition model. The application is also applicable to the application fields of the deep learning frozen gait recognition method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of frozen gait recognition, and particularly relates to a frozen gait recognition method and system based on acceleration motion data. BACKGROUND

[0002] Parkinson's disease is a typical neurodegenerative disease, and patients often face many motor function disorders, of which frozen gait is the most common. The occurrence of frozen gait is random and unpredictable, and this symptom can seriously damage the patient's autonomous action ability and thus reduce the patient's quality of life. Frozen gait is a complex motor symptom, and its onset is easily affected by environmental factors, cognitive input state and drug use. Therefore, it is difficult to quantitatively evaluate it, and there is no uniform detection and evaluation standard. In the traditional recognition method, the patient's report questionnaire and evaluation scale are mainly relied on, but these means are mainly based on the subjective description of the patient and the clinical judgment of the doctor, so the evaluation result of the frozen gait lacks a certain accuracy, and the results between different evaluators also have differences. The traditional recognition method cannot meet the actual needs, so it is of great significance to objectively and accurately recognize the frozen gait and intervene and treat it.

[0003] The existing frozen gait data set has a serious imbalance between the number of normal gait samples and frozen gait samples, which will cause the model to be more inclined to normal gait features during learning, thereby ignoring frozen gait features, and thus affecting the final recognition effect of the model.

[0004] In addition, the current deep learning model for frozen gait recognition is mostly based on time recurrent network and simple convolutional neural network. These models often ignore the "spatial" distribution of limb motion information in the time dimension, i.e. local features, trend features, periodic features, etc. in the time dimension. These spatial features in the time dimension also have an important role in frozen gait recognition, for example, frozen gait is often accompanied by changes in specific action patterns, which show certain rules and characteristics in the space of the time dimension. SUMMARY

[0005] The present application is to solve the problems of the imbalance between the number of normal gait samples and frozen gait samples in the existing frozen gait data set, the model learning to be inclined to normal gait features, the frozen gait features being ignored, and the insufficient extraction of time-space features in the existing deep learning method in frozen gait recognition. To solve the above technical problems, the present application is realized by the following technical scheme:

[0006] Scheme one, the present application provides a frozen gait recognition method based on acceleration motion data, the method comprising the following steps:

[0007] Step 1, obtaining the freezing gait acceleration data of Parkinson's patients, and preprocessing the freezing gait acceleration data of the Parkinson's patients to obtain a data sample set containing normal gait and freezing gait;

[0008] Step 2, based on the acceleration data sample set containing normal gait and freezing gait obtained in step 1, input into a multi-scale convolutional network, extract features in different time scales in parallel, and fuse the features in different time scales to obtain preliminary fusion features;

[0009] Step 3, based on the preliminary fusion features extracted in step 2, input into a freezing gait recognition deep feature extraction model based on a residual network for deep feature extraction;

[0010] Step 4, based on the deep features output in step 3, sequentially input into a channel attention module and a spatial attention module, respectively, to focus on key features and suppress redundant information in two dimensions through adaptive weighting;

[0011] Step 5, the key feature set obtained in step 4 is passed through dimension reduction and a classifier to finally output the result;

[0012] Step 6, based on steps 1 to 5, the freezing gait acceleration data collected from three different collection positions is used to train, verify and test the freezing gait recognition deep feature extraction model to realize freezing gait recognition of acceleration data in different positions.

[0013] Further, a preferred embodiment is provided, in which Python software is used in step 1 to preprocess the freezing gait acceleration data, and the preprocessing includes filtering and denoising, window segmentation, window label addition, and freezing gait data sample balancing.

[0014] Further, a preferred embodiment is provided, in which the multi-scale convolutional network in step 2 is implemented by building a one-dimensional multi-scale feature extraction and fusion module;

[0015] Branch one, using one-dimensional convolution with a convolution kernel size of 1x1 for feature extraction;

[0016] Branch two, using one-dimensional convolution with a convolution kernel size of 1x1 and one-dimensional convolution with a convolution kernel size of 1x3 in series for feature extraction;

[0017] Branch three, using one-dimensional convolution with a convolution kernel size of 1x1 and two one-dimensional convolution with a convolution kernel size of 1x3 in series for feature extraction;

[0018] Branch four, using one-dimensional max pooling layer with a convolution kernel size of 1x3 and one-dimensional convolution with a convolution kernel size of 1x1 in series for feature extraction;

[0019] The outputs of the four branches are spliced in the channel dimension to build a comprehensive feature containing multi-scale information, and the output is represented as: Filter Concatenation = torch.cat([Branch_1, Branch_2, Branch_3, Branch_4], dim = 1).

[0020] Further, a preferred embodiment is provided, and the step 3 of constructing the frozen gait recognition deep feature extraction model based on the residual network is to construct a deep learning model based on a one-dimensional residual block, use one-dimensional convolution to process the time series gait signal, and form a feature extraction architecture from shallow to deep by continuously stacking multiple residual blocks.

[0021] Further, a preferred embodiment is provided, and the step 4 of the channel attention module and the spatial attention module sequentially realizes the weighting optimization method of the channel dimension feature and the spatial dimension feature in the channel and spatial dimensions.

[0022] The channel attention module realizes the optimization method of the channel dimension feature, which is:

[0023] The channel descriptor is generated by global average pooling and maximum pooling, the channel weight is generated by a shared multi-layer perception, and the key feature channel is weighted and optimized;

[0024] The weighting optimization method of the spatial dimension feature is:

[0025] The spatial information is fused by channel dimension double pooling and feature splicing, the spatial weight is generated by a convolution layer, and the feature focusing of the important region is realized.

[0026] Further, a preferred embodiment is provided, and the step 5 of constructing the dimension reduction and classifier module is:

[0027] Step 5.1, constructing a feature dimension reduction module, introducing a one-dimensional adaptive average pooling layer at the end of the attention module constructed in step 5, for reducing the dimension of the feature;

[0028] Step 5.2, constructing a classifier module, the classifier module realizes classification through a first full connection layer and a second full connection layer, that is, the feature vector after dimension reduction in the feature dimension reduction module in step 5.1 is input into the first full connection layer, and the feature is mapped to a low-dimensional space, and the feature is mapped to a category space through the second full connection layer.

[0029] Further, a preferred embodiment is provided, and the step 6 further includes the step of performing a frozen gait recognition experiment on the accelerometer motion data collected at three different positions using the complete frozen gait recognition model.

[0030] Scheme two, a frozen gait recognition system based on acceleration motion data, the system comprises:

[0031] A data acquisition module is configured to acquire acceleration data of frozen gait of Parkinson's patients, and to preprocess the acceleration data of frozen gait of Parkinson's patients to obtain a data sample set containing normal gait and frozen gait.

[0032] A multi-scale feature extraction and fusion module is configured to extract features of data at different scales based on the acceleration data sample set containing normal gait and frozen gait obtained by the data acquisition module, and to fuse the features by splicing to form a preliminary feature representation.

[0033] A deep feature extraction module is configured to extract deep features based on the fused feature map output by the multi-scale feature extraction module, and to extract more deep and discriminative high-level feature representations through multi-layer nonlinear transformation based on a residual network.

[0034] A feature optimization and selection module is configured to receive and process the high-level feature representation output by the deep feature extraction module, and to focus on key feature data sets and suppress redundant information through adaptive weighting, so as to optimize the feature representation.

[0035] A frozen gait classification module is configured to perform pooling and dimension reduction processing on the key feature set processed by the feature optimization and selection module, and to output a recognition result through a fully connected layer.

[0036] Scheme three, a computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is executed by the processor to realize the steps of the method of scheme one.

[0037] Scheme four, a computer device, comprising a memory and a processor, the memory stores a computer program, when the processor runs the computer program stored in the memory, the processor executes the method of scheme one.

[0038] The present application has the advantages of:

[0039] The application aims to provide a frozen gait recognition method and system based on acceleration motion data, taking gait data of Parkinson's patients collected by an accelerometer as the research object, and taking improving the recognition performance of frozen gait as the target.

[0040] The application uses a targeted data balancing method, such as downsampling for normal gait samples of the majority sample class and data enhancement for frozen gait samples of the minority sample class, so that the number of processed normal gait samples and frozen gait samples reaches a quantitative balance, and the influence of the unbalanced distribution of the number of sample classes on the classification performance of the model and the detection ability of frozen gait is reduced.

[0041] The application builds a deep learning model based on a residual network to recognize frozen gait, stacks residual blocks to build a shallow-to-deep feature extraction architecture, enhances the feature extraction ability of the model for frozen gait time series data, and overcomes the problem of ignoring data spatial feature extraction in existing deep learning-based frozen gait recognition methods.

[0042] The application introduces a multi-scale feature extraction and fusion and feature optimization strategy based on the frozen gait recognition model based on a residual network. Multi-scale feature extraction and fusion extract features from data through parallel convolution kernels of different sizes, capture features in different receptive fields in data at the same time, and enhance the feature extraction ability of the model. The feature optimization strategy further realizes the perception ability of different key feature information in the frozen gait acceleration signal by optimizing the channel and spatial dimension features in sequence.

[0043] The application is also applicable to the field of deep learning frozen gait recognition method and the like. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The multi-scale feature extraction module structure diagram according to the first embodiment.

[0045] Figure 2 The deep feature extraction model based on a residual network according to the first embodiment.

[0046] Figure 3The convolutional block attention module structure diagram according to the first embodiment.

[0047] Figure 4 The complete frozen gait recognition model diagram according to the first embodiment.

[0048] Figure 5 The complete frozen gait recognition framework diagram according to the first embodiment. DETAILED DESCRIPTION

[0049] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application.

[0050] In the first embodiment, a frozen gait recognition method based on acceleration motion data is provided, which comprises the following steps:

[0051] Step 1: Obtain the frozen gait acceleration data of Parkinson's patients, and pre-process the frozen gait acceleration data of Parkinson's patients to obtain a data sample set containing normal gait and frozen gait;

[0052] Step 2: Based on the acceleration data sample set containing normal gait and frozen gait obtained in step 1, input into a multi-scale convolutional network, extract features at different time scales in parallel, and fuse the features at different time scales to obtain preliminary fusion features;

[0053] Step 3: Based on the preliminary fusion features extracted in step 2, input into a frozen gait recognition deep feature extraction model based on a residual network for deep feature extraction;

[0054] Step 4: Based on the deep features output in step 3, input into a channel attention module and a spatial attention module in turn, respectively focus on key features in two dimensions through adaptive weighting, and suppress redundant information;

[0055] Step 5: The key feature set obtained in step 4 is input into a dimension reduction and classifier to finally output the result;

[0056] Step 6: Based on steps 1 to 5, frozen gait recognition deep feature extraction model training, verification and testing are implemented to realize frozen gait recognition of acceleration data at different positions by using frozen gait acceleration data collected at three different collection positions respectively.

[0057] Embodiment two, the embodiment is further limited to the frozen gait recognition method based on acceleration motion data of embodiment one, the pre-processing of frozen gait acceleration data in step 1 uses Python software, and the pre-processing includes filtering denoising, window segmentation, window adding label, and frozen gait data sample balancing.

[0058] Embodiment three, the embodiment is further limited to the frozen gait recognition method based on acceleration motion data of embodiment one, the multi-scale convolution network in step 2 is realized by adopting a one-dimensional multi-scale feature extraction fusion module;

[0059] That is, branch one, using one-dimensional convolution with a convolution kernel of 1x1 size for feature extraction;

[0060] Branch two, using one-dimensional convolution with a convolution kernel of 1x1 size and one-dimensional convolution with a convolution kernel of 1x3 size in series for feature extraction;

[0061] Branch three, using one-dimensional convolution with a convolution kernel of 1x1 size and two one-dimensional convolution with a convolution kernel of 1x3 in series for feature extraction;

[0062] Branch four, using one-dimensional max pooling layer with a convolution kernel of 1x3 size and one-dimensional convolution with a convolution kernel of 1x1 in series for feature extraction;

[0063] The outputs of the four branches are spliced in the channel dimension to construct a comprehensive feature containing multi-scale information, and the output is represented as: Filter Concatenation=torch.cat([Branch_1,Branch_2,Branch_3,Branch_4],dim=1).

[0064] Embodiment four, the embodiment is further limited to the frozen gait recognition method based on acceleration motion data of embodiment three, the frozen gait recognition deep feature extraction model based on residual network in step 3 is constructed, which is a deep learning model based on one-dimensional residual block, using one-dimensional convolution to process time series gait signals, and by continuously stacking multiple residual blocks, a feature extraction architecture from shallow to deep is formed.

[0065] Embodiment five, the embodiment is further limited to the frozen gait recognition method based on acceleration motion data of embodiment one, the channel attention module and the spatial attention module in step 4 realize the weighted optimization method of channel dimension features and spatial dimension features in channel and space dimensions in turn:

[0066] The method for the channel attention module to optimize the channel dimension features is:

[0067] The channel descriptor is generated by global average pooling and maximum pooling, the channel weight is generated by a shared multi-layer perception, and the key feature channel is weighted and optimized;

[0068] The method for weighted optimization of the spatial dimension feature is:

[0069] The spatial information is fused by channel dimension double pooling and feature splicing, the spatial weight is generated by a convolution layer, and the feature of an important region is focused.

[0070] Embodiment six, this embodiment is a further limitation of the frozen gait recognition method based on acceleration motion data according to embodiment one, and the method for constructing the dimension reduction and classifier module in step 5 is:

[0071] Step 5.1, constructing a feature dimension reduction module, a one-dimensional adaptive average pooling layer is introduced at the end of the attention module constructed in step 5, which is used for dimension reduction of the feature;

[0072] Step 5.2, constructing a classifier module, the classifier module realizes classification through a first full connection layer and a second full connection layer, that is, the feature vector after dimension reduction in the feature dimension reduction module in step 5.1 is input into the first full connection layer, and the feature is mapped to a low-dimensional space, and the feature is mapped to a class space through the second full connection layer.

[0073] Embodiment seven, this embodiment is a further limitation of the frozen gait recognition method based on acceleration motion data according to embodiment one, and step 6 further includes the step of using the complete frozen gait recognition model to perform a frozen gait recognition experiment on the accelerometer motion data collected at three different positions.

[0074] Embodiment eight, this embodiment proposes a frozen gait recognition system based on acceleration motion data, the system comprises:

[0075] A data acquisition module is used to acquire frozen gait acceleration data of Parkinson's patients, and to preprocess the frozen gait acceleration data of Parkinson's patients to obtain a data sample set containing normal gait and frozen gait;

[0076] A multi-scale feature extraction and fusion module is used to extract features of data at different scales based on the acceleration data sample set containing normal gait and frozen gait obtained by the data acquisition module, and to fuse the features by splicing to form a preliminary feature representation;

[0077] A deep feature extraction module is used to extract deep features based on the fused feature map output by the multi-scale feature extraction module, and to extract more deep and discriminative high-level feature representations through multi-layer nonlinear transformation based on a residual network;

[0078] a feature optimization selection module, configured to receive and process the high-level feature representation output by the deep feature extraction module, and focus on a key feature dataset and suppress redundant information through adaptive weighting, so as to optimize the feature representation;

[0079] a frozen gait classification module, configured to perform pooling dimension reduction processing on the key feature set processed by the feature optimization selection module, and finally output a recognition result through a fully connected layer.

[0080] Embodiment Nine, the embodiment provides a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to realize steps of the method in any one of the embodiments One to seven.

[0081] Embodiment Ten, the embodiment provides a computer device, including a memory and a processor, the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the method in any one of the embodiments One to seven.

[0082] Embodiment Eleven, the embodiment is used to explain the above-mentioned embodiments One to ten, and the embodiment is specifically:

[0083] Reference Figures 1-5 The embodiment is explained with reference to Figure 1 It is shown that the application provides a frozen gait recognition method based on acceleration motion data, balances samples through downsampling and data enhancement, constructs a residual network with multi-scale feature extraction and attention mechanism fusion, improves the spatial feature extraction capability, and thus improves the recognition accuracy and robustness.

[0084] The application builds a deep feature extraction model to recognize frozen gait, and innovatively introduces multi-scale feature extraction and fusion and feature optimization strategy, further enhances the feature extraction capability of the model, and realizes the capture capability of complex patterns in the frozen gait acceleration signal.

[0085] The application uses acceleration data collected under experimental conditions by 10 idiopathic Parkinson's patients with a history of frozen gait, each patient wears three accelerometers for data collection, the sampling frequency is 64HZ, and the accelerometers are respectively arranged at different three positions on the patient's body: the first is fixed on the left lower leg of the patient, i.e. the part above the ankle; the second accelerometer is fixed on the left thigh, i.e. the part above the knee; and the third accelerometer is fixed on the back of the patient, i.e. the part above the hip joint. The application recognizes frozen gait from the data collected by the three accelerometers, specifically including the following steps:

[0086] Step 1, obtain the acceleration data of Parkinson's patients with freezing gait, and preprocess the data to reduce noise and data sample imbalance.

[0087] Step 2, in order to comprehensively capture local key features and global trend features in the data and establish a correlation, a multi-scale feature extraction and fusion module is introduced, which realizes multi-scale feature extraction and fusion through different sizes of parallel convolution, and improves the model's understanding ability of complex patterns and structures in input data.

[0088] Step 3, in order to enhance the spatial feature extraction ability of freezing gait acceleration data, we construct a freezing gait recognition deep learning model based on one-dimensional residual block. The model uses one-dimensional convolution to process time series gait signals, and through continuous stacking of multiple residual blocks, it forms a feature extraction architecture from shallow to deep, improving the model's deep feature extraction ability.

[0089] Step 4, in order to enhance the feature optimization ability, the attention mechanism is introduced, which realizes the weighted optimization of features in channel and space dimensions through channel attention and spatial attention, and enhances the model's attention to key features.

[0090] Step 5, in order to realize the reduction of feature vector dimension in the input classifier and the final classification decision, a dimension reduction classification module is constructed.

[0091] Step 6, construct the complete freezing gait recognition framework; use the complete freezing gait recognition framework to conduct freezing gait recognition experiments on the accelerometer motion data collected at three different positions, and conduct comparative experiments with other methods.

[0092] Further, step 1 includes the following steps:

[0093] The freezing gait acceleration data is preprocessed using Python software, including:

[0094] Step 1.1, data extraction: the data collected by the three accelerometers in the data set are extracted respectively, which are used for subsequent freezing gait recognition of single acceleration sensor at different collection positions.

[0095] Step 1.2, filter denoising: in view of the high-frequency noise in the data, a second-order Butterworth low-pass filter with a cutoff frequency of 30Hz is used for denoising.

[0096] Step 1.3, window segmentation: the data is divided into windows with a time length of 2 seconds and an overlap rate of 50%, that is, a sliding window with a size of 128 and a step of 64 is used for data segment, obtaining three groups of data samples with a channel number of 3 and a length of 128.

[0097] Step 1.4, adding labels: the label with the highest frequency in the window is selected as the data segment label. This method is based on the statistical characteristics of the sample labels within the window, and can effectively reflect and summarize the main features or trends of the data segment. By following the "mode principle", the accuracy and representativeness of the label are enhanced.

[0098] Step 1.5, data balancing: due to the imbalance between the actual normal samples and the frozen gait samples in the data set, the number of two types of samples in the data needs to be balanced to ensure that the model learns the features of both types of samples equally during training. The specific settings are as follows:

[0099] Downsampling: for existing normal gait samples, use a combination of random downsampling and Tomek Links downsampling to reduce the number of normal gait samples.

[0100] Data augmentation: for existing frozen gait data samples, use arbitrary rotation to generate frozen gait samples. Since the gait signal collected by the acceleration sensor during actual data collection may be affected by factors such as sensor position and angle, causing a shift, appropriate transformation and rotation of the signal will not change the class label of the gait, and will expand the number of frozen gait samples. Data is the initial sample data, axis is the randomly generated rotation axis, angle is the randomly generated rotation angle, and rotation_matrix is the rotation matrix generated by axis and angle. By rotated_matrix = np.matmul(Data, rotation_matrix), the data points are multiplied by the rotation matrix to obtain new samples after 3D rotation of the frozen gait samples, achieving sample expansion. 3

[0101] Further, step 2 includes the following:

[0102] Build a multi-scale feature extraction module and create a Class Inception(nn.Module) class. Figure 1 For multi-scale feature extraction module. The specific settings are as follows:

[0103] Parallel multi-scale processing:

[0104] Branch one (self.Branch_1): one-dimensional convolution with a convolution kernel size of 1x1 is used for feature extraction.

[0105] Branch two (self.Branch_2): one-dimensional convolution with a convolution kernel size of 1x1 and one-dimensional convolution with a convolution kernel size of 1x3 are used in series for feature extraction.

[0106] Branch 3 (self.Branch_3): One-dimensional convolution with a convolution kernel of 1x1 size and two one-dimensional convolutions with a convolution kernel of 1x3 are connected in series for feature extraction.

[0107] Branch 4 (self.Branch_4): One-dimensional max-pooling layer with a convolution kernel of 1x3 size and one one-dimensional convolution with a convolution kernel of 1x1 are connected in series for feature extraction.

[0108] Feature concatenation: The outputs of the four branches are concatenated in the channel dimension to construct a comprehensive feature containing multi-scale information. The output can be represented as:

[0109] Filter Concatenation=torch.cat([Branch_1,Branch_2, Branch_3,Branch_4],dim=1)

[0110] Further, the deep feature extraction model based on residual network built in step 3 is Figure 2 , creating a ClassResNet(nn.Module) class, with specific settings:

[0111] Residual block (self.Basic Block): Five residual blocks (self.Block_1, self.Block_2, self.Block_3, self.Block_4, self.Block_5) are stacked in turn, each containing two one-dimensional convolutions with a convolution kernel size of 1x3. The residual block realizes feature reuse through jump connection and introduces batch normalization (nn.Batch Norm1d) and ReLU activation function (nn.ReLU).

[0112] Further, step 4 creates a Class CBMA(nn.Module) class, and the data passes through the channel attention module and the spatial attention module in turn. Figure 3 is a convolution block attention module. The specific settings are:

[0113] Step 4.1: Channel attention module realizes the optimization of channel dimension features:

[0114] Global pooling: The input feature map is aggregated through average pooling and max pooling operations to generate two different intermediate feature maps and .

[0115] Shared multi-layer perceptron: The intermediate feature maps and Meanwhile, a shared network composed of multiple layers of perceptron is fed into, and a channel attention vector is obtained after the shared network is applied to each feature map.

[0116] Activation function: a sigmoid activation function is applied to the channel attention vector to generate attention weights in the channel dimension.

[0117] Channel feature weighting: the channel weights are multiplied with the input feature map to obtain a weighted feature map. .

[0118] Step 4.2: The spatial attention module is implemented to optimize the spatial dimension features:

[0119] Channel dimension pooling: the feature map is averaged and max-pooled along the channel axis. The two generated feature maps are concatenated along the channel axis.

[0120] Convolution operation: the concatenated feature map is passed through a convolution layer with a 1x7 kernel to generate a spatial attention vector.

[0121] Activation function: a sigmoid activation function is applied to the spatial attention vector to generate attention weights in the spatial dimension.

[0122] Spatial feature weighting: the spatial weights are multiplied with the input feature map to obtain a weighted feature map. .

[0123] Further, step 5 constructs a dimension reduction classifier module. The specific settings are as follows:

[0124] Step 5.1, construct a feature dimension reduction module. The specific settings are as follows:

[0125] A one-dimensional adaptive average pooling layer (self.Global_Avg_Pool) is introduced at the end of the feature extraction module to reduce the dimension of the features. This reduces the dimension of the feature vector in the input classifier while preserving important feature information, thereby reducing the computational load.

[0126] Step 5.2, classifier module. In the classifier module, classification decisions are made through two fully connected layers. The specific settings are as follows:

[0127] self.fc1: the pooled feature vector is input into the fully connected layer self.fc1, which maps the features to a low-dimensional space.

[0128] self.fc2: finally, the features are mapped to the class space through the fully connected layer self.fc2, obtaining the frozen gait results of the model.

[0129] In summary, steps 1 to 5 create the Class ResGait(nn.Module) class, which builds the complete frozen gait recognition model. Figure 4 is the complete frozen gait recognition model. This section specifically includes:

[0130] Multi-scale feature extraction and fusion (self.inception): Different sizes of convolution kernels are used to extract multi-scale spatial features, which enhances the model's ability to express different gait features in time series without significantly increasing computational complexity.

[0131] Deep feature extraction (self.resnet): Deep feature extraction is performed through five residual blocks in succession, each composed of two convolution layers with a kernel size of 1x3, and batch normalization and ReLu activation functions are introduced.

[0132] Feature optimization (self.cbam): Channel attention modules are used to selectively emphasize important channels and reduce redundant information, and spatial attention modules are combined to focus on key position information in the input feature vector.

[0133] Adaptive average pooling (self.global_avg_pool): Dimensionality reduction is performed by the adaptive average pooling layer, which down-samples the feature map to a specified size while preserving important global feature information.

[0134] self.Dropout1(p=0.5): Dropout regularization layer is added with a dropout rate of 0.5, which suppresses overfitting by randomly deactivating neurons and improves the model's generalization ability on small gait datasets.

[0135] self.fc1: The fully connected layer self.fc1 is inputted with the pooled feature vector, which maps the features to a low-dimensional space.

[0136] self.Dropout2(p=0.5): Dropout regularization layer is added with a dropout rate of 0.5, which suppresses overfitting by randomly deactivating neurons and improves the model's generalization ability on small gait datasets.

[0137] self.fc2: Finally, the fully connected layer self.fc2 maps the features to the class space, resulting in the model's frozen gait results.

[0138] Finally, the freeze gait recognition experiments are performed on the accelerometer motion data collected from three different locations using the complete freeze gait recognition framework, and compared with other methods. Tables 1, 2 and 3 respectively show the freeze gait recognition results of the motion data collected from the same location using different methods, in which the freeze gait recognition method proposed in the present application has the best recognition performance.

[0139] Table 1

[0140]

[0141] Table 2

[0142]

[0143] Table 3

[0144]

[0145] Those skilled in the art can understand that the above description is merely preferred embodiments of the present application, and the features described in each embodiment and / or technical solution of the present disclosure can be combined or combined, even if such combination or combination is not explicitly described in the present disclosure. It is not intended to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, and those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacements for part of the technical features, and any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0146] Although the preferred embodiments of the present application have been described, those skilled in the art can make further changes and modifications to these embodiments once they know the basic creative concept. Therefore, the appended technical solutions are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and changes of the present application fall within the scope of the present application and its equivalent technology, the present application also intends to include these modifications and changes.

Claims

1. A method for recognizing frozen gait based on acceleration motion data, characterized by, The method comprises the following steps: Step 1, obtaining the freezing gait acceleration data of Parkinson's patients, and preprocessing the freezing gait acceleration data of the Parkinson's patients to obtain a data sample set containing normal gait and freezing gait; Step 2, based on the acceleration data sample set containing normal gait and freezing gait obtained in step 1, input into a multi-scale convolutional network, extract features at different time scales in parallel, and fuse the features at different time scales to obtain preliminary fusion features; Step 3, based on the preliminary fusion features extracted in step 2, input into a freezing gait recognition deep feature extraction model based on a residual network for deep feature extraction; Step 4, based on the deep features output in step 3, input into a channel attention module and a spatial attention module in turn, respectively, in two dimensions through adaptive weighting to focus on key features and suppress redundant information; Step 5, the key feature set obtained in step 4 is input into a dimension reduction and classifier to finally output the result; Step 6, based on steps 1 to 5, freezing gait recognition model training, verification and testing are implemented on freezing gait recognition of acceleration data at different positions by collecting freezing gait acceleration data at three different collection positions; The multi-scale convolutional network in step 2 is implemented by building a one-dimensional multi-scale feature extraction and fusion module; That is, branch one, using one-dimensional convolution with a convolution kernel size of 1x1 for feature extraction; Branch two, using one-dimensional convolution with a convolution kernel size of 1x1 and one-dimensional convolution with a convolution kernel size of 1x3 in series for feature extraction; Branch three, using one-dimensional convolution with a convolution kernel size of 1x1 and two one-dimensional convolution with a convolution kernel size of 1x3 in series for feature extraction; Branch four, using one-dimensional maximum pooling layer with a convolution kernel size of 1x3 and one one-dimensional convolution with a convolution kernel size of 1x1 in series for feature extraction; The outputs of the four branches are spliced in the channel dimension to construct a comprehensive feature containing multi-scale information, and the output is represented as: Filter Concatenation = torch.cat([Branch_1, Branch_2, Branch_3, Branch_4], dim=1); In step 3, the freezing gait recognition deep feature extraction model based on the residual network is constructed based on a one-dimensional residual block to build a deep learning model, and one-dimensional convolution is used to process the time sequence gait signal, and a plurality of residual blocks are stacked in series to form a feature extraction architecture from shallow to deep; In step 4, the channel attention module and the spatial attention module realize the weighted optimization method of the channel dimension features and the spatial dimension features in the channel and the space respectively: The method for optimizing the channel dimension features by the channel attention module is: Generate channel descriptors through global average pooling and maximum pooling, generate channel weights through shared multi-layer perceptron, and realize weighted optimization of key features channels; The method for weighted optimization of spatial dimension features is: Through channel dimension double pooling and feature splicing fusion of spatial information, generate spatial weights through convolution layer to focus on important region features.

2. The method of claim 1, wherein the method is based on acceleration motion data. In step 1, the frozen gait acceleration data is preprocessed using Python software, including filtering denoising, window segmentation, window labeling, and balancing of frozen gait data samples.

3. The method of claim 1, wherein the method is based on acceleration motion data. In step 5, the method for constructing the dimension reduction and classifier module is: In step 6.1, a feature dimension reduction module is constructed, and a one-dimensional adaptive average pooling layer is introduced at the end of the attention module constructed in step 5, which is used for dimension reduction of features. In step 6.2, a classifier module is constructed, which realizes classification through a first fully connected layer and a second fully connected layer, that is, the feature vector after dimension reduction in the feature dimension reduction module in step 6.1 is input into the first fully connected layer, and the features are mapped to a low-dimensional space, and the features are mapped to a class space through the second fully connected layer.

4. The method of claim 1, wherein the method is based on acceleration motion data. In step 7, it also includes the step of using the complete frozen gait recognition model to perform frozen gait recognition experiments on the accelerometer motion data collected at three different positions.

5. A frozen gait recognition system based on acceleration motion data, characterized by, The system is realized based on the method of claim 1, and the system comprises: A data acquisition module is configured to acquire frozen gait acceleration data of Parkinson's patients, and to preprocess the frozen gait acceleration data of the Parkinson's patients to obtain a data sample set containing normal gait and frozen gait; A multi-scale feature extraction and fusion module is configured to use convolution kernels of different sizes to simultaneously extract features of data at different scales based on the acceleration data sample set containing normal gait and frozen gait obtained by the data acquisition module, and to fuse the features by splicing to form a preliminary feature representation; A deep feature extraction module is configured to extract deep features based on the fused feature map output by the multi-scale feature extraction module, and to extract more deep and discriminative high-level feature representations through multi-layer nonlinear transformation based on a residual network; A feature optimization selection module is configured to receive and process the high-level feature representations output by the deep feature extraction module, and to focus on key feature data sets and suppress redundant information through adaptive weighting, thereby optimizing the feature representations; A frozen gait classification module is configured to perform pooling and dimension reduction processing on the key feature set processed by the feature optimization selection module, and to output the recognition result finally through the integration of information by a fully connected layer.

6. A computer storage medium having stored thereon a computer program, characterized in that The computer program is executed by a processor to realize the method of any one of claims 1-4.

7. A computer device, comprising: Comprise: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to realize the method of any one of claims 1-4.

Images