Fiber-optic smart carpet gait recognition system based on strain-profile dual-modal network

Abstract

The application discloses a fiber intelligent carpet gait recognition system based on strain-contour bimodal network and belongs to the technical field of intelligent sensors, which comprises an OFDR sensing system and can realize continuous high-sensitivity strain collection on a long-distance optical fiber; the optical fiber sensing carpet arranges sensing optical fibers on the carpet to form a dense distributed strain sensing array and collect two-dimensional strain distribution of a foot bottom in real time; a local strain-global contour bimodal attention network, a local strain graph attention module for modeling dynamic coupling relationship between optical fiber sections and extracting key stress features, a gradient-guided global contour non-local attention module for obtaining a foot bottom structure contour and overall stress form, and finally integrating the two kinds of information to recognize a gait category; the fiber intelligent carpet gait recognition system based on the strain-contour bimodal network solves the problems of insufficient spatial resolution, low recognition accuracy and limitation by environmental conditions of the existing gait recognition method.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensor technology, and in particular to an optical fiber intelligent carpet gait recognition system based on a strain-profile dual-modal network. Background Technology

[0002] Gait recognition, a technology that analyzes human walking or posture changes to achieve identity verification, behavior monitoring, and health assessment, is widely used in medical rehabilitation, security monitoring, and other fields. Compared with other biometric identification methods, it has significant advantages such as being non-contact, having strong continuity, and being easy to collect data.

[0003] Existing gait recognition technologies are mainly divided into three categories: vision-based methods rely on cameras to obtain human information, but are easily affected by lighting, occlusion, shooting angle and privacy restrictions, and lack robustness; wearable device-based methods can provide high-precision motion data, but require users to wear additional hardware, affecting comfort and ease of use; among ground pressure or strain sensing methods, traditional ground sensors (such as piezoresistive arrays and low-density fiber arrays) have problems such as low spatial resolution, insufficient signal-to-noise ratio and inability to continuously sample, making it difficult to capture fine-grained changes in foot strain, thus limiting the accuracy and generalization ability of gait recognition.

[0004] With the development of distributed fiber optic sensing (DFOS) technology, optical frequency domain reflectance (OFDR) technology, with its millimeter-level spatial resolution and micro-strain-level sensitivity, provides a new direction for building high-precision smart carpets. This technology can achieve continuous, high-density, and high-sensitivity strain acquisition over long-distance optical fibers, and has advantages such as resistance to electromagnetic interference and flexible installation. It can construct a high-resolution two-dimensional strain sensing network to accurately reflect the foot contact area, pressure distribution, and gait dynamics.

[0005] However, relying solely on local strain signals from optical fibers has limitations: firstly, complex mechanical coupling relationships exist between different fiber segments, which traditional algorithms struggle to model, requiring the integration of relevant information for accurate identification; secondly, local strain signals alone are insufficient for extracting global features such as plantar pressure profiles and overall stress structures, and lack the ability to identify subtle differences in similar gaits. Therefore, a dual-modal gait recognition technology that integrates local strain information from optical fibers with global plantar contour features is urgently needed to fully leverage the sensing advantages of OFDR smart carpets and improve the accuracy and robustness of gait recognition. Summary of the Invention

[0006] The purpose of this invention is to provide a fiber optic intelligent carpet gait recognition system based on strain-profile dual-modal networks, which solves the problems of insufficient robustness, poor comfort and convenience, and limited accuracy and generalization ability of existing gait recognition technologies based on vision, wearable devices and traditional ground sensors.

[0007] To achieve the above objectives, the present invention provides an optical fiber intelligent carpet gait recognition system based on a strain-contour dual-modal network, including an OFDR sensing system, an optical fiber sensing carpet, and a local strain-global contour dual-modal attention network SCDA-Net;

[0008] The OFDR sensing system is composed of a main interferometer and an auxiliary interferometer. The two interferometers share the same tuned laser source and achieve optical path distribution, reflection and interference signal output through fiber couplers and circulators with different splitting ratios. The main interferometer is used to inject swept light into the fiber under test and receive backscattered Rayleigh signals. The auxiliary interferometer is used to generate K-clock interference signals to compensate for the linearity error of the swept laser.

[0009] The fiber optic sensing carpet uses silicone as a flexible base. The entire OFDR strain sensing fiber is laid parallel on the flexible base at a fixed interval and fixed with silicone glue. Another layer of silicone pad is covered on top and pressed together with the silicone glue and the flexible base. The OFDR strain sensing fiber forms a dense distributed strain sensing point array on the entire fiber optic sensing carpet.

[0010] The Local Strain-Global Contour Dual-Modal Attention Network (SCDA-Net) comprises a data preparation section, a strain map attention module (SGAM), a pressure profile-guided nonlocal attention module (PGGNM), and a final prediction module. The data preparation section preprocesses the raw strain signal, inputting the original signal in two-dimensional matrix form into SGAM and the high-resolution image data after two-dimensional matrix interpolation into PGGNM. SGAM extracts and processes features from the raw strain signal, capturing dynamic correlations across fiber segments and weighting key node features. PGGNM extracts and processes features from the image modal data, fusing gradient information from the image with the nonlocal attention mechanism. The final prediction module maps the combination of these two feature sets to the final gait recognition result.

[0011] Preferably, the main interferometer includes a tuned laser source, fiber coupler one, fiber coupler two, fiber coupler three, fiber circulator one, fiber circulator two, polarization controller, polarization beam splitter, photodetector one, and the fiber under test. The output end of the tuned laser source is connected to the input end of fiber coupler one. 5% of the light output from fiber coupler one is introduced into port one of fiber circulator one, and 95% of the light is output to fiber coupler two. Fiber coupler two inputs 10% of the light to the polarization controller and inputs 90% of the incident signal to port one of fiber circulator two. Port two of fiber circulator two is connected to the input end of the fiber under test. The Rayleigh scattering signal output from port three and the output signal from port two of the polarization controller enter fiber coupler three together and interfere. The two output ends of fiber coupler three are connected to the two ports of photodetector one through the polarization beam splitter. The electrical signal output from photodetector one enters the data acquisition system.

[0012] Preferably, the auxiliary interferometer includes a fiber optic circulator, a fiber optic coupler, a delay fiber, a Faraday rotator mirror, a Faraday rotator mirror, a photodetector, and a photodetector. Port 1 of the fiber optic circulator receives reference light from the 5% output end of the fiber optic coupler, and port 2 is connected to port 1 of the fiber optic coupler. Port 3 of the fiber optic coupler is connected to the delay fiber, and the end of the delay fiber is connected to the Faraday rotator mirror, forming the long arm of the auxiliary interferometer. Port 4 of the fiber optic coupler is connected to the Faraday rotator mirror, forming the short arm of the auxiliary interferometer. The interference beat frequency signal generated by the mixing and interference of the short arm light and the long arm light in the fiber optic coupler is output from port 2 of the fiber optic coupler and received by the photodetector. The output electrical signal of the photodetector enters the data acquisition system as the K-clock time base correction signal of the system.

[0013] Preferably, in the data preparation section, the raw strain signal collected by the fiber optic sensing carpet is processed to remove outliers, and the strain data of the fiber segment is stitched into a two-dimensional matrix and input into SGAM; after interpolation, the two-dimensional matrix is ​​used to construct a high-resolution image data, which is then normalized and input into PGGNM.

[0014] Preferably, SGAM constructs the raw strain data into graphical data, treating each fiber segment as a node in the graph, with each node characterized by the strain sequence along its length. All nodes are The graph data is aggregated through a graph attention layer after the K-nearest neighbor layer to build connections between nodes. Each of its neighboring nodes The attention coefficient is:

[0015]

[0016] in, For nodes For neighboring nodes The original attention coefficient; It is a learnable attention parameter. This represents vector concatenation. For nodes Node characteristics;

[0017] For neighboring nodes The attention coefficients are normalized using the Softmax function:

[0018]

[0019] in, This represents the normalized attention coefficient; Represents a node The set of neighbors;

[0020] Aggregate and update node features:

[0021]

[0022] in, Represents the node after aggregation New features;

[0023] Finally, a Softmax layer is used to calculate the importance of all nodes and weight them to obtain the output features. :

[0024]

[0025] in, This represents the total number of nodes in the graph structure.

[0026] Preferably, PGGNM performs feature extraction and downsampling on the input image using a convolutional encoder:

[0027]

[0028] in, This represents the input image, obtained by interpolation from the original strain signal. express The data format is a three-dimensional array of real numbers. For the number of channels, For height, Width;

[0029] Calculate the transverse and longitudinal pressure edge gradients using the Sobel operator:

[0030]

[0031]

[0032] in, The horizontal gradient of the image. The vertical gradient of the image. For the horizontal gradient operator, For the vertical gradient operator, the Sobel kernel is defined as:

[0033] , ;

[0034] Learn the overall gradient factor:

[0035]

[0036] in, For the Sigmoid function, The kernel size is Convolution;

[0037] After obtaining the gradient factor, it is injected into the nonlocal attention mechanism to process image features. The mapping process yields the Query vector, Key vector, and Value vector:

[0038]

[0039]

[0040]

[0041] in, For the query vector, For the key vector, For the Value vector;

[0042] Add the gradient factor to the Key vector:

[0043]

[0044] in, This represents the enhanced key vector;

[0045] Flatten the three feature maps into vectors with dimensions of . ;

[0046] Calculate global attention for the image:

[0047]

[0048] Among them, the attention weight matrix ;

[0049] Weighting of image features:

[0050] ;

[0051] Finally, the gradient-enhanced features are subjected to dimensionality transformation and residual connection to obtain the output features. .

[0052] Preferably, the final prediction module outputs gait recognition results; after SGAM and PGGNM, two types of output features are obtained. and The two vectors are transformed into one-dimensional vectors through a dimension transformation operation; after being processed by a linear layer, they are added together and then mapped to the gait recognition result through a fully connected layer.

[0053] Therefore, the fiber optic intelligent carpet gait recognition system based on strain-profile dual-mode network described above has the following beneficial effects:

[0054] (1) This invention uses OFDR technology to construct a high-density fiber optic smart carpet, with the fiber embedded in a flexible substrate to form a large-area two-dimensional distributed strain sensing base. Compared with traditional pressure arrays or discrete sensing structures, OFDR fiber optic carpets can achieve continuous sampling, strong anti-interference ability, and high stability, providing a high-quality raw data foundation for gait recognition;

[0055] (2) This invention proposes a local strain-global contour dual-modal attention network. The local strain map attention module extracts the dynamic coupling relationship between fiber segments, enhancing fine-grained features related to key gait regions. Simultaneously, a pressure gradient-guided nonlocal module captures the overall shape and contour structure of the foot, achieving effective encoding of global information. This significantly improves the accuracy of gait recognition.

[0056] (3) Experiments were conducted on 16 types of gait. The present invention achieved a classification accuracy of 97.55% and an AUC of 0.9994, which is significantly better than the traditional method. At the same time, the model has only 0.156M parameters and has real-time recognition capability.

[0057] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0058] Figure 1 This is an overall system framework diagram of an embodiment of the present invention;

[0059] Figure 2 This is a schematic diagram of an OFDR system according to an embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram of a fiber optic sensing carpet according to an embodiment of the present invention;

[0061] Figure 4 This is a wiring diagram according to an embodiment of the present invention;

[0062] Figure 5 This is a diagram of the local strain-global contour bimodal attention network structure according to an embodiment of the present invention.

[0063] Figure 6 This is a diagram of 16 types of gait data signals collected by the OFDR fiber optic carpet according to an embodiment of the present invention;

[0064] Figure 7 This is an image showing the extraction of key regions from two-dimensional strain data and image data using a local strain-global contour dual-modal attention network according to an embodiment of the present invention.

[0065] Figure Labels

[0066] 1. Tuned laser source; 2. Fiber optic coupler one; 3. Fiber optic coupler two; 4. Fiber optic coupler three; 5. Fiber optic coupler four; 6. Fiber optic circulator one; 7. Fiber optic circulator two; 8. Polarization controller; 9. Polarization beam splitter; 10. Delay fiber; 11. Faraday rotating mirror one; 12. Faraday rotating mirror two; 13. Photodetector one; 14. Photodetector two; 15. And the fiber under test. Detailed Implementation

[0067] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0068] The fiber optic intelligent carpet gait recognition system based on strain-profile dual-mode network of the present invention, such as... Figure 1 As shown, the process is implemented step by step through the following steps, with the structure, data, and network functions involved in each step all combined with the appendix. Figure 1-7 Detailed explanation:

[0069] I. Constructing an OFDR Sensing System

[0070] like Figure 2As shown, the OFDR sensing system of the present invention consists of core components such as a tuned laser source 1, fiber coupler 1 (5:95), fiber coupler 2 (10:90), fiber coupler 3 (50:50), fiber coupler 4 (50:50), fiber coupler 4 (5), fiber circulator 1 (6), fiber circulator 2 (7), polarization controller 8, polarization beam splitter 9, delay fiber 10, Faraday rotating mirror 11, Faraday rotating mirror 2 (12), photodetector 1 (13), photodetector 2 (14), and fiber under test 15. High-precision strain signal acquisition is achieved through the coordinated operation of the main interferometer and the auxiliary interferometer.

[0071] (I) Setup of the main interferometer

[0072] The output of tuned laser source 1 is connected to the input of fiber coupler 2, providing a swept-frequency laser source for the entire system. Fiber coupler 2 distributes the optical signal at a split ratio of 5:95, with 5% of the light introduced into port 1 of fiber circulator 6 to supply reference light for the interferometer; 95% of the light is transmitted to fiber coupler 3, which directs most of the optical power into the main path of the main interferometer. Fiber coupler 3 further distributes the optical signal at a split ratio of 10:90, with 10% of the light input to polarization controller 8 to adjust the polarization state of the reference arm light, ensuring the stability of the interference signal; 90% of the incident signal is input to port 1 of fiber circulator 7. Port 2 of fiber circulator 7 is connected to the input of fiber 15 under test. The Rayleigh scattering signal reflected back from fiber 15 is output through port 3 of fiber circulator 7 and enters the input of fiber coupler 4 together with the output signal from port 2 of polarization controller 8, interfering with the Rayleigh scattered light phase information and converting it into a measurable interference light intensity signal. The two outputs of fiber coupler 34 are connected to the two ports of photodetector 13 via polarization beam splitter 9 to suppress common-mode noise in a balanced detection mode. The electrical signal output by photodetector 13 is connected to the data acquisition system to acquire the Rayleigh scattering interference signal formed by the main interferometer.

[0073] (II) Construction of the auxiliary interferometer

[0074] Port 1 of fiber optic circulator 6 receives reference light from the 5% output of fiber optic coupler 2, providing a light source for the auxiliary interferometer. Port 2 of fiber optic circulator 6 is connected to port 1 of fiber optic coupler 5. Port 3 of fiber optic coupler 5 is connected to delay fiber 10, and the end of delay fiber 10 is connected to Faraday rotating mirror 11, together forming the long arm of the auxiliary interferometer, introducing a fixed optical path difference. The reflected light returns to port 3 of fiber optic coupler 5. Port 4 of fiber optic coupler 5 is connected to Faraday rotating mirror 12, forming the short arm of the auxiliary interferometer. The reflected light returns to port 4 of fiber optic coupler 5. The short-arm and long-arm beams mix and interfere in fiber coupler 4.5, generating an interference beat frequency signal whose frequency is proportional to the laser sweep speed. This signal is output from port 2 of fiber coupler 4.5 and is received by photodetector 2.14 as an auxiliary interference signal. The electrical signal output from photodetector 2.14 is connected to the data acquisition system as the system's K-clock time base correction signal, providing an external trigger clock for the data acquisition system. This ensures that the main interferometer signal is sampled at equal intervals in the wavenumber domain, compensating for the effects of laser sweep nonlinearity.

[0075] II. Constructing a fiber optic sensing carpet

[0076] like Figure 3 , Figure 4 As shown, the fiber optic sensing carpet of the present invention uses flexible silicone as a substrate and adopts a dense fiber optic cabling method to achieve high-resolution strain sensing.

[0077] A 1mm thick silicone substrate is used as a flexible base. The entire OFDR strain sensing fiber is laid parallel to the substrate at a fixed spacing, with a longitudinal spacing of 1cm between fibers. Silicone adhesive is used to fix the fibers to the substrate, and a 0.5mm thick silicone pad is placed on top, then pressed together with the silicone adhesive and the lower silicone pad to form an integrated structure. The sensing fibers form a dense, distributed strain sensing point array on the carpet. The OFDR host has a strain sampling resolution of 2.6mm along the fiber direction, and the carpet width is 50cm, ensuring comprehensive coverage and fine-grained strain acquisition of the foot contact area. Furthermore, this invention can also employ other dense wiring methods, adjusting the fiber distribution according to the actual application scenario to ensure the uniformity and integrity of strain sensing.

[0078] III. Constructing a local strain-global contour bimodal attention network

[0079] like Figure 5 As shown, a local strain-global contour bimodal attention network SCDA-Net is constructed. This network includes a data preparation part, a strain map attention module SGAM, a pressure contour-guided nonlocal module PGGNM, and a final prediction module. High-precision gait recognition is achieved through bimodal feature fusion.

[0080] (a) Data preparation

[0081] The raw strain signal acquired by the strain-sensing fiber optic carpet is a two-dimensional matrix. First, the raw strain signal is preprocessed to remove outliers. The strain data of the fiber segment is then stitched into a 38×146 two-dimensional matrix and directly input into SGAM. The two-dimensional matrix is ​​then interpolated to construct a 256×256 high-resolution image data, which is then normalized and input into PGGNM to provide adaptive data for dual-modal feature extraction.

[0082] (II) Feature Extraction of SGAM Module

[0083] SGAM constructs two-dimensional strain data into a graph structure. Since the dimension of the two-dimensional strain data is 38×146, the feature length of each node is 146, and the number of nodes is 38. Each fiber segment is regarded as a node in the graph, and the node feature is the strain sequence of the fiber along its length. All nodes are The connection relationship between nodes is constructed through K-nearest neighbors, set to 5. The graph data undergoes node aggregation through the graph attention layer. Each of its neighboring nodes The attention coefficient is:

[0084]

[0085] in, For nodes For neighboring nodes The original attention coefficient; It is a learnable attention parameter. This represents vector concatenation. For nodes Node characteristics, Since the linear layer has a dimension of 256, the dimension of the node features is transformed into 256.

[0086] For neighboring nodes The attention coefficients are normalized using the Softmax function:

[0087]

[0088] in, This represents the normalized attention coefficient; Represents a node The set of neighbors;

[0089] Aggregate and update node features:

[0090]

[0091] in, Represents the node after aggregation New features;

[0092] Finally, a Softmax layer is used to calculate the importance of all nodes and weight them to obtain the output features. :

[0093]

[0094] in, This represents the total number of nodes in the graph structure. The dimensions are 38×256.

[0095] (III) Feature Extraction of PGGNM Module

[0096] PGGNM extracts and downsamples features from the input image using a convolutional encoder:

[0097]

[0098] in, This represents the input image, obtained by interpolation from the original strain signal, with dimensions of 256×256. express The data format is a three-dimensional array of real numbers. For the number of channels, For height, This refers to the width, specifically 64 for all values.

[0099] Calculate the transverse and longitudinal pressure edge gradients using the Sobel operator:

[0100]

[0101]

[0102] in, The horizontal gradient of the image. The vertical gradient of the image. For the horizontal gradient operator, For the vertical gradient operator, the Sobel kernel is defined as:

[0103] , ;

[0104] Learn the overall gradient factor:

[0105]

[0106] in, For the Sigmoid function, The kernel size is Convolution;

[0107] After obtaining the gradient factor, it is injected into the nonlocal attention mechanism to process image features. The mapping process yields the Query vector, Key vector, and Value vector:

[0108]

[0109]

[0110]

[0111] in, For the query vector, For the key vector, For the Value vector;

[0112] Add the gradient factor to the Key vector:

[0113]

[0114] in, This represents the enhanced key vector;

[0115] Flatten the three feature maps into vectors with dimensions of . ;

[0116] Calculate global attention for the image:

[0117]

[0118] Among them, the attention weight matrix ,here It is a 4096×4096 square matrix, representing spatial relationships;

[0119] Weighting of image features:

[0120] ;

[0121] Finally, the gradient-enhanced features are subjected to dimensionality transformation and residual connection to obtain the output features. Its dimensions are 64×64×64.

[0122] (iv) Prediction module output

[0123] For two output features and The vector is transformed into a one-dimensional vector. After processing by a linear layer, the vectors are summed, and then mapped to the gait recognition result through a fully connected layer.

[0124] IV. Collect real gait data using the OFDR sensing system and fiber optic sensing carpet, set the hyperparameters of the training network, and begin training.

[0125] Step 1: Connect the constructed OFDR sensing system and the constructed fiber optic sensing carpet, and simultaneously connect a computer for data collection. The collected data is divided into training data and test data according to a certain ratio. Subjects are instructed to perform various specific gait movements on the fiber optic carpet, and the OFDR system is used to collect their contact and force changes on the carpet surface, thereby obtaining a high-density two-dimensional strain distribution sequence. The actual collected dataset contains 2350 gait samples, covering 16 different gait categories, including 8 standing gaits and 8 sitting gaits. Gait types include feet together, feet apart, heel raised, forefoot raised, inward-pointing, outward-pointing, and single-leg. Figure 6 The dataset showcases 16 representative gait samples collected. It is divided into a 60% training set, 15% validation set, and 25% test set, and can be used for network training, tuning, and performance evaluation.

[0126] Step 2: Initialize network parameters. Set the hyperparameters for the training process. Training uses a batch iteration method with a batch size of 32 and a total of 15 iterations. Set the learning rate to 0.0005 and the optimizer to the Adam optimizer. Start training.

[0127] 5. Perform the same preprocessing procedure on the test set data as on the training set, and input it into the trained network for testing. This process does not update the network parameters. Calculate the accuracy, specificity, AUC, and other performance metrics of the test results. Tables 1 and 2 show a comparison between the proposed method and existing methods, demonstrating that the proposed method exhibits superior performance. Figure 7 The proposed method is demonstrated to extract key regions from two modalities of data, showing that the proposed method can effectively focus on key regions.

[0128] Table 1. Comparison of results between the present invention and traditional machine learning methods on 16 gait signals.

[0129]

[0130] Table 2 Comparison of results between the present invention and deep learning methods on 16 gait signals.

[0131]

[0132] Therefore, this invention employs the aforementioned fiber optic intelligent carpet gait recognition system based on a strain-contour dual-modal network. It constructs a high-density, high-sensitivity fiber optic sensing carpet using OFDR sensing technology, and combines a dual-modal network composed of a local strain map attention module and a pressure gradient-guided non-local module. This effectively integrates the dynamic coupling relationship between fiber segments and the global contour features of the foot, solving problems such as insufficient robustness, poor comfort, limited accuracy, and inadequate global feature extraction in existing gait recognition technologies. Ultimately, it achieves a high classification accuracy of 97.55% and an AUC of 0.9994 for 16 gait types, with a model of only 0.156M parameters. It combines high accuracy, high robustness, and real-time recognition capabilities, providing a reliable gait recognition solution for fields such as medical rehabilitation and security monitoring.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A fiber optic intelligent carpet gait recognition system based on strain-profile dual-mode network, characterized in that, This includes OFDR sensing systems, fiber optic sensing carpets, and the local strain-global profile dual-modal attention network SCDA-Net; The OFDR sensing system is composed of a main interferometer and an auxiliary interferometer. The two interferometers share the same tuned laser source and achieve optical path distribution, reflection and interference signal output through fiber couplers and circulators with different splitting ratios. The main interferometer is used to inject swept light into the fiber under test and receive backscattered Rayleigh signals. The auxiliary interferometer is used to generate K-clock interference signals to compensate for the linearity error of the swept laser. The fiber optic sensing carpet uses silicone as a flexible base. The entire OFDR strain sensing fiber is laid parallel on the flexible base at a fixed interval and fixed with silicone glue. Another layer of silicone pad is covered on top and pressed together with the silicone glue and the flexible base. The OFDR strain sensing fiber forms a dense distributed strain sensing point array on the entire fiber optic sensing carpet. The Local Strain-Global Contour Dual-Modal Attention Network (SCDA-Net) comprises a data preparation section, a strain map attention module (SGAM), a pressure profile-guided nonlocal attention module (PGGNM), and a final prediction module. The data preparation section preprocesses the raw strain signal, inputting the original signal in two-dimensional matrix form into SGAM and the high-resolution image data after two-dimensional matrix interpolation into PGGNM. SGAM extracts and processes features from the raw strain signal, capturing dynamic correlations across fiber segments and weighting key node features. PGGNM extracts and processes features from the image modal data, fusing gradient information from the image with the nonlocal attention mechanism. The final prediction module maps the combination of these two feature sets to the final gait recognition result.

2. The fiber optic intelligent carpet gait recognition system based on strain-profile dual-mode network according to claim 1, characterized in that, The main interferometer includes a tuned laser source, fiber coupler 1, fiber coupler 2, fiber coupler 3, fiber circulator 1, fiber circulator 2, polarization controller, polarization beam splitter, photodetector 1, and the fiber under test. The output of the tuned laser source is connected to the input of fiber coupler 1. 5% of the light output from fiber coupler 1 is introduced into port 1 of fiber circulator 1, and 95% of the light is output to fiber coupler 2. Fiber coupler 2 inputs 10% of the light to the polarization controller and 90% of the incident signal to port 1 of fiber circulator 2. Port 2 of fiber circulator 2 is connected to the input of the fiber under test. The Rayleigh scattering signal output from port 3 and the output signal from port 2 of the polarization controller enter fiber coupler 3 together and interfere. The two outputs of fiber coupler 3 are connected to the two ports of photodetector 1 through the polarization beam splitter. The electrical signal output from photodetector 1 enters the data acquisition system.

3. The fiber optic intelligent carpet gait recognition system based on strain-profile dual-mode network according to claim 2, characterized in that, The auxiliary interferometer includes fiber optic circulator I, fiber optic coupler IV, delay fiber, Faraday rotator mirror I, Faraday rotator mirror II, photodetector II, and photodetector II. Port 1 of fiber optic circulator I receives reference light from the 5% output end of fiber optic coupler I, and port 2 is connected to port 1 of fiber optic coupler IV. Port 3 of fiber optic coupler IV is connected to the delay fiber, and the end of the delay fiber is connected to Faraday rotator mirror I, forming the long arm of the auxiliary interferometer. Port 4 of fiber optic coupler IV is connected to Faraday rotator mirror II, forming the short arm of the auxiliary interferometer. The interference beat frequency signal generated by the mixing and interference of the short arm light and the long arm light in fiber optic coupler IV is output from port 2 of fiber optic coupler IV and received by photodetector II. The output electrical signal of photodetector II enters the data acquisition system as the K-clock time base correction signal of the system.

4. The fiber optic intelligent carpet gait recognition system based on strain-profile dual-mode network according to claim 1, characterized in that, In the data preparation section, the raw strain signals collected by the fiber optic sensing carpet are processed to remove outliers. The strain data of the fiber segments are stitched into a two-dimensional matrix and input into SGAM. After interpolation, the two-dimensional matrix is ​​used to construct a high-resolution image data, which is then normalized and input into PGGNM.

5. The fiber optic intelligent carpet gait recognition system based on strain-profile dual-mode network according to claim 4, characterized in that, SGAM constructs graph-structured data from the raw strain data, treating each fiber segment as a node in the graph, with each node characterized by the strain sequence along its length. All nodes are The graph data is aggregated through a graph attention layer after the K-nearest neighbor layer to build connections between nodes. Each of its neighboring nodes The attention coefficient is: in, For nodes For neighboring nodes The original attention coefficient; It is a learnable attention parameter. This represents vector concatenation. For nodes Node characteristics; For neighboring nodes The attention coefficients are normalized using the Softmax function: in, This represents the normalized attention coefficient; Represents a node The set of neighbors; Aggregate and update node features: in, Represents the node after aggregation New features; Finally, a Softmax layer is used to calculate the importance of all nodes and weight them to obtain the output features. : in, This represents the total number of nodes in the graph structure.

6. The fiber optic intelligent carpet gait recognition system based on strain-profile dual-mode network according to claim 5, characterized in that, PGGNM extracts and downsamples features from the input image using a convolutional encoder: in, This indicates that the input image is obtained by interpolation from the original strain signal. express The data format is a three-dimensional array of real numbers. For the number of channels, For height, Width; Calculate the transverse and longitudinal pressure edge gradients using the Sobel operator: in, The horizontal gradient of the image. The vertical gradient of the image. For the horizontal gradient operator, For the vertical gradient operator, the Sobel kernel is defined as: ， ； Learn the overall gradient factor: in, For the Sigmoid function, The kernel size is Convolution; After obtaining the gradient factor, it is injected into the nonlocal attention mechanism to process image features. The mapping process yields the Query vector, Key vector, and Value vector: in, For the query vector, For the key vector, For the Value vector; Add the gradient factor to the Key vector: in, This represents the enhanced key vector; Flatten the three feature maps into vectors with dimensions of . ; Calculate global attention for the image: Among them, the attention weight matrix ; Weighting of image features: ； Finally, the gradient-enhanced features are subjected to dimensionality transformation and residual connection to obtain the output features. .

7. The fiber optic intelligent carpet gait recognition system based on strain-profile dual-mode network according to claim 6, characterized in that, The final prediction module outputs gait recognition results; after SGAM and PGGNM, two types of output features are obtained. and The two vectors are transformed into one-dimensional vectors through a dimension transformation operation; after being processed by a linear layer, they are added together and then mapped to the gait recognition result through a fully connected layer.

Images