Freezing gait recognition method and system based on acceleration motion data

Abstract

The invention discloses a frozen gait recognition method and system based on acceleration motion data, and relates to the field of frozen gait recognition. The problems that the number of normal gaits and frozen gait samples in an existing frozen gait data set is unbalanced, so that the model is deviated to normal gait features during learning, frozen gait features are ignored, and the recognition effect is affected are solved. The method comprises the following steps: acquiring frozen gait acceleration data of a Parkinson's patient, and preprocessing the frozen gait acceleration data to obtain a data sample set containing normal gaits and frozen gaits; constructing a deep learning frozen gait recognition deep feature extraction model based on a residual network; and introducing a multi-scale convolution module and an attention module to construct a complete frozen gait recognition model, and performing frozen gait recognition on the key feature set obtained after processing on the complete frozen gait recognition model. The method is also suitable for the application fields of deep learning frozen gait recognition methods and the like.

Description

Technical Field

[0001] This invention relates to the field of frozen gait recognition technology, and more specifically to a frozen gait recognition method and system based on acceleration motion data. Background Technology

[0002] Parkinson's disease, a typical neurodegenerative disease, often presents patients with numerous motor dysfunctions, among which frozen gait is the most common. The occurrence of frozen gait is random and unpredictable, and this symptom severely impairs patients' ability to move independently, thus reducing their quality of life. Frozen gait is a complex motor symptom, and its onset is easily influenced by environmental factors, cognitive input, and medication use, making its quantitative assessment quite difficult. Currently, there is no unified detection and assessment standard. Traditional identification methods mainly rely on self-report questionnaires and assessment scales, but these methods are primarily based on the patient's subjective description and the physician's clinical judgment. Therefore, the evaluation results of frozen gait lack certain accuracy, and there are also differences in results among different evaluators. Traditional identification methods cannot meet practical needs; therefore, objective and accurate identification of frozen gait is of great significance for its intervention and treatment.

[0003] There is a serious imbalance between the number of normal gait samples and frozen gait samples in the existing frozen gait dataset. This causes the model to favor normal gait features during learning, thus ignoring frozen gait features and affecting the model's final recognition performance.

[0004] Furthermore, most current deep learning models used for frozen gait recognition are based on recurrent temporal networks and simple convolutional neural networks. These models often neglect the "spatial" distribution of limb motion information in the temporal dimension, i.e., local features, trend features, periodic features, etc., in the temporal dimension. These spatial features in the temporal dimension also play an indispensable role in frozen gait recognition. For example, frozen gait is often accompanied by changes in specific movement patterns, and these changes exhibit certain regularities and characteristics in the spatial dimension of time. Summary of the Invention

[0005] This invention addresses the problems of imbalanced sample sizes between normal and frozen gait in existing frozen gait datasets, leading to model learning bias towards normal gait features and neglect of frozen gait features, thus affecting recognition performance, and the inadequacy of existing deep learning methods in extracting temporal and spatial features in frozen gait recognition. To solve these technical problems, this invention employs the following technical solution: Option 1: This invention proposes a frozen gait recognition method based on acceleration motion data, the method comprising the following steps: Step 1: Obtain frozen gait acceleration data of Parkinson's patients, and preprocess the frozen gait acceleration data of Parkinson's patients to obtain a data sample set containing normal gait and frozen gait; Step 2: Based on the acceleration data sample set including normal gait and frozen gait obtained in Step 1, input it 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 fused features; Step 3: Based on the preliminary fusion features extracted in Step 2, input them into the frozen gait recognition deep feature extraction model based on residual network for deep feature extraction; Step 4: Based on the deep features output in Step 3, input them sequentially into the channel attention module and the spatial attention module, respectively, and use adaptive weighting to focus on key features in two dimensions and suppress redundant information. Step 5: Pass the key feature set obtained in Step 4 through a dimensionality reduction and classifier to finally output the result; Step 6: Based on steps 1 to 5, use frozen gait acceleration data collected from three different acquisition locations to train, validate, and test the frozen gait recognition deep feature extraction model to achieve frozen gait recognition for acceleration data at different locations.

[0006] Furthermore, a preferred embodiment is provided in which Python software is used to preprocess the frozen gait acceleration data in step 1. The preprocessing includes filtering and denoising, window segmentation, adding labels to windows, and balancing frozen gait data samples. Furthermore, a preferred embodiment is provided, wherein the multi-scale convolutional network described in step 2 is implemented by constructing a one-dimensional multi-scale feature extraction and fusion module; Branch 1 uses a one-dimensional convolution with a kernel size of 1×1 for feature extraction. Branch 2 uses concatenated one-dimensional convolutions with kernel size 1×1 and one-dimensional convolutions with kernel size 1×3 for feature extraction; Branch 3 uses a one-dimensional convolution with a kernel size of 1×1 and two concatenated one-dimensional convolutions with kernel size of 1×3 for feature extraction; Branch 4 uses a one-dimensional max pooling layer with a 1×3 kernel and a one-dimensional convolutional layer with a 1×1 kernel in series for feature extraction; The outputs of the four branches are concatenated along the channel dimension to construct a comprehensive feature containing multi-scale information. The output is represented as: Filter Concatenation=torch.cat([Branch_1,Branch_2,Branch_3,Branch_4],dim=1).

[0007] Furthermore, a preferred implementation is provided, in which the frozen gait recognition deep feature extraction model based on residual network is constructed in step 3. This is based on building a deep learning model with one-dimensional residual blocks, using one-dimensional convolution to process temporal gait signals, and forming a feature extraction architecture from shallow to deep by continuously stacking multiple residual blocks.

[0008] Furthermore, a preferred embodiment is provided, wherein the method for the channel attention module and the spatial attention module to perform weighted optimization of channel dimension features and spatial dimension features in step 4 is as follows: The channel attention module optimizes channel-dimensional features using the following method: Channel descriptors are generated by global average pooling and max pooling, and channel weights are generated by a shared multilayer perceptron to achieve weighted optimization of key feature channels; The weighted optimization method for the spatial dimension features is as follows: By fusing spatial information through channel-dimensional dual pooling and feature concatenation, spatial weights are generated via convolutional layers to achieve feature focusing on important regions.

[0009] Furthermore, a preferred embodiment is provided, wherein the method for constructing the dimensionality reduction and classifier module in step 5 is as follows: Step 5.1: Construct a feature dimensionality reduction module. At the end of the attention module constructed in step 5, a one-dimensional adaptive average pooling layer is introduced to reduce the dimensionality of the features. Step 5.2: Construct a classifier module. The classifier module achieves classification through a first fully connected layer and a second fully connected layer. Specifically, the feature vectors after dimensionality reduction in the feature dimensionality reduction module in Step 5.1 are input into the first fully connected layer, and the features are mapped to a low-dimensional space. Then, the features are mapped to the category space through the second fully connected layer.

[0010] Furthermore, in a preferred embodiment, step 6 further includes a step of conducting a frozen gait recognition experiment on accelerometer motion data collected at three different locations using a complete frozen gait recognition model.

[0011] Option 2: A frozen gait recognition system based on acceleration motion data, the system comprising: The data acquisition module is used to acquire frozen gait acceleration data of Parkinson's patients and preprocess the frozen gait acceleration data of Parkinson's patients to obtain a data sample set containing normal gait and frozen gait. The multi-scale feature extraction and fusion module is used to extract features from the acceleration data sample set containing normal gait and frozen gait obtained by the data acquisition module at different scales simultaneously using convolutional kernels of different sizes, and then fuse them by splicing to form a preliminary feature representation. The deep feature extraction module is used to extract deep features based on the fused feature map output by the multi-scale feature extraction module. Based on the residual network, it extracts deeper and more discriminative high-level feature representations through multi-layer nonlinear transformation. The feature optimization and selection module receives and processes the high-level feature representation output from the deep feature extraction module, and optimizes the feature representation by focusing on key feature datasets and suppressing redundant information through adaptive weighting. The frozen gait classification module is used to perform pooling and dimensionality reduction on the key feature set processed by the feature optimization and selection module, and then integrate the information through a fully connected layer to finally output the recognition result.

[0012] Option 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in Option 1.

[0013] Option 4: A computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the method described in Option 1.

[0014] The advantages of this invention are: The purpose of this invention is to propose a frozen gait recognition method and system based on acceleration motion data. Taking gait data of Parkinson's patients collected by accelerometers as the research object, and aiming to improve the recognition performance of frozen gait, the invention employs downsampling and data augmentation techniques to balance the data samples during the data preprocessing stage. In the model part, a deep learning frozen gait recognition model based on residual networks is built. By stacking residual blocks, a feature extraction architecture from shallow to deep is constructed to enhance the model's feature extraction capability for frozen gait temporal data. Furthermore, based on this, an improved scheme combining multi-scale feature extraction and an attention mechanism is proposed to further enhance the model's ability to extract and optimize spatial features from frozen gait temporal data, thereby improving the accuracy and robustness of frozen gait recognition.

[0015] This invention addresses the problem of imbalanced data sample classes in existing datasets by employing targeted data balancing methods. For example, it downsamples the majority of normal gait samples and augments the minority of frozen gait samples. After processing, the number of normal gait samples and frozen gait samples are balanced, reducing the impact of the imbalanced distribution of sample classes on the model's classification performance and its ability to detect frozen gait.

[0016] This invention constructs a deep learning model based on residual networks to identify frozen gait. By stacking residual blocks, a feature extraction architecture from shallow to deep is built, which enhances the model's feature extraction capability for frozen gait time series data and overcomes the problem of neglecting data space feature extraction in existing deep learning-based frozen gait recognition methods.

[0017] This invention introduces multi-scale feature extraction and fusion, as well as feature optimization strategies, into a frozen gait recognition model based on residual networks. Multi-scale feature extraction and fusion extracts features from the data by using convolutional kernels of different sizes in parallel, simultaneously capturing features from different receptive fields and enhancing the model's feature extraction capabilities. The feature optimization strategy further enhances the perception of different key feature information in the frozen gait acceleration signal by sequentially optimizing channel and spatial dimension features.

[0018] This invention is also applicable to fields such as deep learning frozen gait recognition methods. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the multi-scale feature extraction module structure described in Implementation Method 1.

[0020] Figure 2 This is a schematic diagram of the deep feature extraction model based on residual networks described in Implementation Method 1.

[0021] Figure 3 This is a schematic diagram of the convolutional block attention module structure described in Implementation Method 1.

[0022] Figure 4 This is a schematic diagram of the complete frozen gait recognition model described in Implementation Method 1.

[0023] Figure 5 This is a schematic diagram of the complete frozen gait recognition framework described in Implementation Method 1. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0025] Implementation Method 1: This implementation method proposes a frozen gait recognition method based on acceleration motion data. The method includes the following steps: Step 1: Obtain frozen gait acceleration data of Parkinson's patients, and preprocess the frozen gait acceleration data of Parkinson's patients to obtain a data sample set containing normal gait and frozen gait; Step 2: Based on the acceleration data sample set including normal gait and frozen gait obtained in Step 1, input it 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 fused features; Step 3: Based on the preliminary fusion features extracted in Step 2, input them into the frozen gait recognition deep feature extraction model based on residual network for deep feature extraction; Step 4: Based on the deep features output in Step 3, input them sequentially into the channel attention module and the spatial attention module, respectively, and use adaptive weighting to focus on key features in two dimensions and suppress redundant information. Step 5: Pass the key feature set obtained in Step 4 through a dimensionality reduction and classifier to finally output the result; Step 6: Based on steps 1 to 5, use frozen gait acceleration data collected from three different acquisition locations to train, validate, and test the frozen gait recognition deep feature extraction model to achieve frozen gait recognition for acceleration data at different locations.

[0026] Implementation Method 2: This implementation method further defines the frozen gait recognition method based on acceleration motion data described in Implementation Method 1. In step 1, Python software is used to preprocess the frozen gait acceleration data. The preprocessing includes filtering and denoising, window segmentation, adding labels to windows, and balancing frozen gait data samples.

[0027] Implementation Method 3: This implementation method further defines the frozen gait recognition method based on acceleration motion data described in Implementation Method 1. The multi-scale convolutional network in step 2 is implemented by constructing a one-dimensional multi-scale feature extraction and fusion module. Branch 1 uses a one-dimensional convolution with a kernel size of 1×1 for feature extraction. Branch 2 uses concatenated one-dimensional convolutions with kernel size 1×1 and one-dimensional convolutions with kernel size 1×3 for feature extraction; Branch 3 uses a one-dimensional convolution with a kernel size of 1×1 and two concatenated one-dimensional convolutions with kernel size of 1×3 for feature extraction; Branch 4 uses a one-dimensional max pooling layer with a 1×3 kernel and a one-dimensional convolutional layer with a 1×1 kernel in series for feature extraction; The outputs of the four branches are concatenated along the channel dimension to construct a comprehensive feature containing multi-scale information. The output is represented as: Filter Concatenation=torch.cat([Branch_1,Branch_2,Branch_3,Branch_4],dim=1).

[0028] Implementation Method 4: This implementation method further defines the frozen gait recognition method based on acceleration motion data described in Implementation Method 3. In step 3, the frozen gait recognition deep feature extraction model based on residual network is constructed. This model is based on one-dimensional residual blocks to build a deep learning model. One-dimensional convolution is used to process the temporal gait signal. By continuously stacking multiple residual blocks, a feature extraction architecture from shallow to deep is formed.

[0029] Implementation Method 5: This implementation method further defines the frozen gait recognition method based on acceleration motion data described in Implementation Method 1. The method for weighted optimization of channel dimension features and spatial dimension features in step 4, implemented sequentially by the channel attention module and the spatial attention module, is as follows: The channel attention module optimizes channel-dimensional features using the following method: Channel descriptors are generated by global average pooling and max pooling, and channel weights are generated by a shared multilayer perceptron to achieve weighted optimization of key feature channels; The weighted optimization method for the spatial dimension features is as follows: By fusing spatial information through channel-dimensional dual pooling and feature concatenation, spatial weights are generated via convolutional layers to achieve feature focusing on important regions.

[0030] Implementation Method Six: This implementation method further defines the frozen gait recognition method based on acceleration motion data described in Implementation Method One. The method for constructing the dimensionality reduction and classifier module in step 5 is as follows: Step 5.1: Construct a feature dimensionality reduction module. At the end of the attention module constructed in step 5, a one-dimensional adaptive average pooling layer is introduced to reduce the dimensionality of the features. Step 5.2: Construct a classifier module. The classifier module achieves classification through a first fully connected layer and a second fully connected layer. Specifically, the feature vectors after dimensionality reduction in the feature dimensionality reduction module in Step 5.1 are input into the first fully connected layer, and the features are mapped to a low-dimensional space. Then, the features are mapped to the category space through the second fully connected layer.

[0031] Implementation Method Seven: This implementation method further defines the frozen gait recognition method based on acceleration motion data described in Implementation Method One. Step 6 also includes a step of conducting a frozen gait recognition experiment on accelerometer motion data collected at three different locations using a complete frozen gait recognition model.

[0032] Implementation Method 8: This implementation method proposes a frozen gait recognition system based on acceleration motion data. The system includes: The data acquisition module is used to acquire frozen gait acceleration data of Parkinson's patients and preprocess the frozen gait acceleration data of Parkinson's patients to obtain a data sample set containing normal gait and frozen gait. The multi-scale feature extraction and fusion module is used to extract features from the acceleration data sample set containing normal gait and frozen gait obtained by the data acquisition module at different scales simultaneously using convolutional kernels of different sizes, and then fuse them by splicing to form a preliminary feature representation. The deep feature extraction module is used to extract deep features based on the fused feature map output by the multi-scale feature extraction module. Based on the residual network, it extracts deeper and more discriminative high-level feature representations through multi-layer nonlinear transformation. The feature optimization and selection module receives and processes the high-level feature representation output from the deep feature extraction module, and optimizes the feature representation by focusing on key feature datasets and suppressing redundant information through adaptive weighting. The frozen gait classification module is used to perform pooling and dimensionality reduction on the key feature set processed by the feature optimization and selection module, and then integrate the information through a fully connected layer to finally output the recognition result.

[0033] Implementation Method Nine: This implementation method provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in any one of Implementation Methods One to Seven.

[0034] Implementation Method 10: This implementation method provides a computer device, including 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 described in any one of Implementation Methods 1 to 7.

[0035] Implementation Method Eleven: This implementation method provides an example, which is used to explain the above-described implementation methods one through ten. The specific example is as follows: See Figures 1-5 This implementation method is described with reference to Figure 1 As shown, this invention proposes a frozen gait recognition method based on acceleration motion data. By downsampling and data augmentation to balance samples, and constructing a residual network with multi-scale feature extraction and attention mechanism fusion, the spatial feature extraction capability is improved, thereby enhancing recognition accuracy and robustness.

[0036] This invention establishes a deep feature extraction model for identifying frozen gait, and innovatively introduces multi-scale feature extraction, fusion, and feature optimization strategies to further enhance the model's feature extraction capabilities and enable it to capture complex patterns in frozen gait acceleration signals.

[0037] This invention uses acceleration data collected under experimental conditions from 10 idiopathic Parkinson's disease patients, all with a history of frozen gait. Each patient wore three accelerometers for data collection, with a sampling frequency of 64 Hz. The accelerometers were placed in three different locations on the patient's body: the first was fixed on the left calf, above the ankle; the second on the left thigh, above the knee; and the third on the back, above the hip joint. This invention performs frozen gait identification on the data collected from the three accelerometers, specifically including the following steps: Step 1: Obtain frozen gait acceleration data from Parkinson's patients, and preprocess the data to reduce the impact of noise and data sample imbalance on the experiment.

[0038] Step 2: In order to fully capture local key features and global trend features in the data and establish their correlation, a multi-scale feature extraction and fusion module is introduced. Multi-scale feature extraction and fusion are achieved through parallel convolutions of different sizes, which improves the model's ability to understand complex patterns and structures in the input data.

[0039] Step 3: To enhance the spatial feature extraction capability of frozen gait acceleration data, we construct a frozen gait recognition deep learning model based on one-dimensional residual blocks. This model uses one-dimensional convolution to process temporal gait signals, and by continuously stacking multiple residual blocks, a feature extraction architecture from shallow to deep is formed, improving the model's deep feature extraction capability.

[0040] Step 4: To enhance feature optimization capabilities, an attention mechanism is introduced. Through channel attention and spatial attention, features are weighted and optimized in both channel and spatial dimensions, thereby enhancing the model's focus on key features.

[0041] Step 5: In order to reduce the dimension of the feature vectors in the input classifier and the final classification decision, a dimensionality reduction classification module is constructed.

[0042] Step 6: Construct a complete frozen gait recognition framework; use the complete frozen gait recognition framework to conduct frozen gait recognition experiments on accelerometer motion data collected at three different locations, and conduct comparative experiments with other methods.

[0043] Furthermore, step 1 includes the following steps: Preprocessing of frozen gait acceleration data using Python software included: Step 1.1, Data Extraction: Extract the data collected by the three accelerometers in the dataset separately, which will be used for subsequent frozen gait recognition of individual accelerometers at different acquisition locations.

[0044] Step 1.2, Filtering and Denoising: To address the high-frequency noise in the data, a second-order Butterworth low-pass filter with a cutoff frequency of 30Hz was used for noise reduction.

[0045] Step 1.3, Window Segmentation: The data is divided using a window with a duration of 2 seconds and an overlap rate of 50%. That is, a sliding window with a size of 128 and a step size of 64 is used to segment the data, resulting in three sets of data samples with 3 channels and a length of 128.

[0046] Step 1.4: Add labels: Select the label that appears most frequently in the window as the data segment label. This method is based on the statistical characteristics of the sample labels in the window, which can effectively reflect and summarize the main characteristics or trends of the data segment. By following the "mode principle", the accuracy and representativeness of the labels are enhanced.

[0047] Step 1.5, Data Balancing: Due to the imbalance between the actual normal samples and frozen gait samples in the dataset, it is necessary to balance the number of samples of the two classes to ensure that the model learns the features of both classes equally during training. Specifically, the settings are as follows: Downsampling: For existing normal gait samples, a combination of random downsampling and Tomek Links downsampling is used to reduce the number of normal gait samples.

[0048] Data augmentation: Existing frozen gait data samples are used to generate frozen gait samples through arbitrary rotation. Since the gait signals acquired by the accelerometer are affected by factors such as sensor position and angle during actual data acquisition, appropriate transformation and rotation of the signal will not change the gait category label, thus expanding the number of frozen gait samples. `Data` represents the initial sample data, `axis` is a randomly generated rotation axis, and `angle` is a randomly generated rotation angle. 3D models are generated using `axis` and `angle`. The rotation matrix rotation_matrix is ​​used to multiply the data points by the rotation matrix through rotated_matrix=np.matmul(Data,rotation_matrix), resulting in a new 3D rotated sample of the frozen gait sample, thus expanding the sample.

[0049] Furthermore, step 2 includes the following: Build a multi-scale feature extraction module and create the Class Inception(nn.Module) class. Figure 1 This is a multi-scale feature extraction module. Specific settings: Parallel multi-scale processing: Branch 1 (self.Branch_1): Uses a one-dimensional convolution with a kernel size of 1×1 for feature extraction.

[0050] Branch 2 (self.Branch_2): Feature extraction is performed by concatenating a one-dimensional convolution with a kernel size of 1×1 and a one-dimensional convolution with a kernel size of 1×3.

[0051] Branch 3 (self.Branch_3): Feature extraction is performed by concatenating a one-dimensional convolution with a kernel size of 1×1 and two one-dimensional convolutions with kernel size of 1×3.

[0052] Branch 4 (self.Branch_4): Feature extraction is performed by concatenating a one-dimensional max pooling layer with a 1×3 kernel and a one-dimensional convolution with a 1×1 kernel.

[0053] Feature concatenation: The outputs of the four branches are concatenated along the channel dimension to construct a comprehensive feature containing multi-scale information. The output can be represented as: Filter Concatenation=torch.cat([Branch_1,Branch_2, Branch_3,Branch_4],dim=1) Furthermore, the deep feature extraction model based on residual networks built in step 3 is... Figure 2 Create a ClassResNet(nn.Module) class, with the following settings: 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 sequentially. Each residual block contains two 1×3 one-dimensional convolutions. Feature reuse is achieved through skip connections within the residual block, and batch normalization (nn.Batch Norm1d) and ReLU activation function (nn.ReLU) are introduced.

[0054] Further, in step 4, the Class CBMA(nn.Module) is created, which contains the channel attention module and the spatial attention module in sequence. Figure 3 It is a convolutional block attention module. Specifically, it is configured as follows: Step 4.1: The channel attention module optimizes channel-dimensional features: Global pooling: First, the input feature map Aggregation is performed using average pooling and max pooling operations to generate two distinct intermediate feature maps. and .

[0055] Shared multilayer perceptron: using intermediate feature maps and Simultaneously, a shared network consisting of multilayer perceptrons is fed in, and the channel attention vector is obtained after applying the shared network to each feature map.

[0056] Activation function: Apply the Sigmoid activation function to the channel attention vector to generate channel-dimensional attention weights.

[0057] Channel feature weighting: Channel weights and input feature map The weighted feature map is obtained by multiplying each channel. .

[0058] Step 4.2: The spatial attention module optimizes spatial dimension features. Channel-dimensional pooling: pooling the feature map along the channel axis. Perform average pooling and max pooling operations, and then stitch the two generated feature maps together along the channel axis.

[0059] Convolution operation: The concatenated feature map is passed through a convolutional layer with a kernel of 1×7 to generate a spatial attention vector.

[0060] Activation function: Apply the Sigmoid activation function to the spatial attention vector to generate spatial attention weights.

[0061] Spatial Feature Weighting: Spatial Weights and Input Feature Map The weighted feature map is obtained by multiplying each position. .

[0062] Further, step 5 constructs the dimensionality reduction classifier module. Specifically, it is set as follows: Step 5.1: Construct the feature dimensionality reduction module. Specific settings: A one-dimensional adaptive average pooling layer (self.Global_Avg_Pool) is introduced at the end of the feature extraction module to reduce the dimensionality of the features. This reduces the dimension of the feature vectors in the input classifier while retaining important feature information, thereby reducing the computational cost.

[0063] Step 5.2, Classifier Module. The classifier module uses two fully connected layers to make classification decisions. Specific settings: self.fc1: Input the pooled feature vector into the fully connected layer self.fc1 to map the features to a low-dimensional space.

[0064] self.fc2: Finally, through the fully connected layer self.fc2, the features are mapped to the class space to obtain the frozen gait result of the model.

[0065] Combining steps 1 to 5, the Class ResGait(nn.Module) class was created, constructing a complete frozen gait recognition model. Figure 4 This is the complete frozen gait recognition model provided by Ning. This part specifically includes: Multi-scale feature extraction and fusion (self.inception): Using convolutional kernels of different sizes to extract multi-scale spatial features, enhancing the model's ability to express different time-series features without significantly increasing computational cost.

[0066] Deep feature extraction (self.resnet): Deep feature extraction is performed through five consecutive residual blocks. Each residual block consists of two convolutional layers with a kernel size of 1×3. Batch normalization and ReLU activation function are introduced.

[0067] Feature optimization (self.cbam): Selectively emphasizes important channels through the channel attention module, reduces redundant information, and combines it with the spatial attention module to focus on key positional information in the input feature vector.

[0068] Adaptive average pooling (self.global_avg_pool): The adaptive average pooling layer performs dimensionality reduction, downsampling the feature map to a specified size while retaining important global feature information.

[0069] self.Dropout1(p=0.5): Adds a Dropout regularization layer with a dropout rate of 0.5. This suppresses overfitting by randomly deactivating neurons, thus improving the model's generalization ability on small gait datasets.

[0070] self.fc1: Input the pooled feature vector into the fully connected layer self.fc1 to map the features to a low-dimensional space.

[0071] self.Dropout2(p=0.5): Adds a Dropout regularization layer with a dropout rate of 0.5. This suppresses overfitting by randomly deactivating neurons, thus improving the model's generalization ability on small gait datasets.

[0072] self.fc2: Finally, through the fully connected layer self.fc2, the features are mapped to the class space to obtain the frozen gait result of the model.

[0073] Finally, a frozen gait recognition experiment was conducted using the complete frozen gait recognition framework on accelerometer motion data collected at three different locations, and the results were compared with other methods. Tables 1, 2, and 3 show the experimental results of frozen gait recognition using different methods on motion data from the same acquisition location, among which the frozen gait recognition method proposed in this invention has the best recognition performance.

[0074] Table 1

[0075] Table 2

[0076] Table 3

[0077] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or technical solutions of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. This is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0078] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended technical solutions are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the present invention. Clearly, those skilled in the art can make various modifications and variations to the present invention without departing from its spirit and scope. Thus, if these modifications and variations of the present invention fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A frozen gait recognition method based on acceleration motion data, characterized in that, The method includes the following steps: Step 1: Obtain frozen gait acceleration data of Parkinson's patients, and preprocess the frozen gait acceleration data of Parkinson's patients to obtain a data sample set containing normal gait and frozen gait; Step 2: Based on the acceleration data sample set including normal gait and frozen gait obtained in Step 1, input it 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 fused features; Step 3: Based on the preliminary fusion features extracted in Step 2, input them into the frozen gait recognition deep feature extraction model based on residual network for deep feature extraction; Step 4: Based on the deep features output in Step 3, input them sequentially into the channel attention module and the spatial attention module, respectively, and use adaptive weighting to focus on key features in two dimensions and suppress redundant information. Step 5: Pass the key feature set obtained in Step 4 through a dimensionality reduction and classifier to finally output the result; Step 6: Based on steps 1 to 5, use frozen gait acceleration data collected from three different acquisition locations to train, validate, and test the frozen gait recognition model to achieve frozen gait recognition for acceleration data at different locations.

2. The frozen gait recognition method based on acceleration motion data according to claim 1, characterized in that, In step 1, Python software is used to preprocess the frozen gait acceleration data. The preprocessing includes filtering and denoising, window segmentation, adding labels to windows, and balancing frozen gait data samples.

3. The frozen gait recognition method based on acceleration motion data according to claim 1, characterized in that, The multi-scale convolutional network described in step 2 is implemented by constructing a one-dimensional multi-scale feature extraction and fusion module; Branch 1 uses a one-dimensional convolution with a kernel size of 1×1 for feature extraction. Branch 2 uses concatenated one-dimensional convolutions with kernels of size 1×1 and 1×3 to extract features; Branch 3 uses a one-dimensional convolution with a kernel size of 1×1 and two concatenated one-dimensional convolutions with kernel size of 1×3 for feature extraction; Branch 4 uses a one-dimensional max pooling layer with a 1×3 kernel and a one-dimensional convolutional layer with a 1×1 kernel in series for feature extraction; The outputs of the four branches are concatenated along the channel dimension to construct a comprehensive feature containing multi-scale information. The output is represented as: Filter Concatenation = torch.cat([Branch_1, Branch_2, Branch_3, Branch_4], dim=1).

4. The frozen gait recognition method based on acceleration motion data according to claim 3, characterized in that, In step 3, a deep feature extraction model for frozen gait recognition based on residual networks is constructed. This model is built on one-dimensional residual blocks, and one-dimensional convolution is used to process temporal gait signals. By continuously stacking multiple residual blocks, a feature extraction architecture from shallow to deep is formed.

5. The frozen gait recognition method based on acceleration motion data according to claim 1, characterized in that, In step 4, the channel attention module and the spatial attention module sequentially implement weighted optimization of channel-dimensional features and spatial-dimensional features in both the channel and spatial dimensions as follows: The channel attention module optimizes channel-dimensional features using the following method: Channel descriptors are generated by global average pooling and max pooling, and channel weights are generated by a shared multilayer perceptron to achieve weighted optimization of key feature channels; The weighted optimization method for the spatial dimension features is as follows: By fusing spatial information through channel-dimensional dual pooling and feature concatenation, spatial weights are generated via convolutional layers to achieve feature focusing on important regions.

6. The frozen gait recognition method based on acceleration motion data according to claim 1, characterized in that, The method for constructing the dimensionality reduction and classifier modules in step 5 is as follows: Step 6.1: Construct a feature dimensionality reduction module. At the end of the attention module constructed in step 5, a one-dimensional adaptive average pooling layer is introduced to reduce the dimensionality of the features. Step 6.2: Construct a classifier module. The classifier module achieves classification through a first fully connected layer and a second fully connected layer. Specifically, the feature vectors after dimensionality reduction in the feature dimensionality reduction module in Step 6.1 are input into the first fully connected layer, and the features are mapped to a low-dimensional space. The features are then mapped to the category space through the second fully connected layer.

7. The frozen gait recognition method based on acceleration motion data according to claim 1, characterized in that, Step 7 also includes a step of conducting a frozen gait recognition experiment on accelerometer motion data collected at three different locations using a complete frozen gait recognition model.

8. A frozen gait recognition system based on acceleration motion data, characterized in that, The system includes: The data acquisition module is used to acquire frozen gait acceleration data of Parkinson's patients and preprocess the frozen gait acceleration data of Parkinson's patients to obtain a data sample set containing normal gait and frozen gait. The multi-scale feature extraction and fusion module is used to extract features from the acceleration data sample set containing normal gait and frozen gait obtained by the data acquisition module at different scales simultaneously using convolutional kernels of different sizes, and then fuse them by splicing to form a preliminary feature representation. The deep feature extraction module is used to extract deep features based on the fused feature map output by the multi-scale feature extraction module. Based on the residual network, it extracts deeper and more discriminative high-level feature representations through multi-layer nonlinear transformation. The feature optimization and selection module receives and processes the high-level feature representation output from the deep feature extraction module, and optimizes the feature representation by focusing on key feature datasets and suppressing redundant information through adaptive weighting. The frozen gait classification module is used to perform pooling and dimensionality reduction on the key feature set processed by the feature optimization and selection module, and then integrate the information through a fully connected layer to finally output the recognition result.

9. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-7.

10. A computer device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method of any one of claims 1-7.

Images