A trajectory recognition method based on distributed acoustic sensing

Abstract

This invention relates to a trajectory recognition method based on distributed acoustic sensing, belonging to the field of vehicle trajectory recognition technology. The trajectory recognition method based on distributed acoustic sensing utilizes a distributed acoustic sensing system to achieve high-precision quantitative measurement of external mechanical disturbances by demodulating the phase change of Rayleigh backscattered light. Its unique feature lies in that the trajectory recognition method includes the following steps: Step 1, acquiring low SNR vibration signals collected by the DAS system within a unit acquisition time; Step 2, in VTR Net, employing the Global-Local Feature Collaborative Modeling Mechanism (GL-FCM) and the Multi-Resolution Attention Feature Map Recovery Architecture (AG-FMR); Step 3, using the CEH-FA model robustness enhancement method for vehicle trajectory recognition to identify and finally output the vehicle trajectory.

Description

Technical Field

[0001] This invention relates to a trajectory recognition method based on distributed acoustic sensing, belonging to the field of vehicle trajectory recognition technology. Background Technology

[0002] With the rapid iteration of global transportation infrastructure and the continuous expansion of highway network coverage, more stringent requirements are being placed on the accuracy, real-time performance, and adaptability of long-distance traffic condition monitoring. Currently, the field of traffic monitoring has developed various technical solutions, covering single-modal sensing devices such as magnetic sensors, cameras, radar sensors, ultrasonic sensors, and vehicle-mounted GPS. However, these technical solutions have revealed significant limitations in practical applications: high deployment costs, difficulty in achieving both spatiotemporal resolution and environmental adaptability (such as performance degradation under severe weather or complex road conditions), making it difficult to meet the needs of large-scale long-distance traffic monitoring applications.

[0003] Against this backdrop, Distributed Acoustic Sensing (DAS) technology, with its core advantages such as low average cost, high spatiotemporal resolution, and strong resistance to electromagnetic interference, has been widely applied in monitoring scenarios for various modes of transportation, including trains, automobiles, ships, and pedestrians. Meanwhile, research on using machine learning (ML) and artificial intelligence (AI) algorithms to extract actionable information from fiber optic sensor data and support real-time autonomous decision-making is gaining momentum, providing new directions for the in-depth application of DAS technology in the field of traffic monitoring.

[0004] Current DAS-based vehicle trajectory recognition technology faces three common challenges: First, existing models lack sufficient global trajectory sensing capabilities, resulting in defects in the completeness of trajectory extraction; second, real DAS data possesses inherent noise characteristics such as signal-to-noise ratio fluctuations and channel failures, interfering with feature extraction and trajectory reconstruction; and third, deep learning models exhibit unstable robustness, making it difficult to maintain consistent detection performance across cross-regional and multi-scenario data. These technical bottlenecks directly restrict the large-scale application of DAS technology in long-distance traffic monitoring, becoming the core driving force behind the design of this invention and the improvement of existing technologies. Summary of the Invention

[0005] To address the problems existing in the prior art, a trajectory recognition method based on distributed acoustic sensing is provided.

[0006] The present invention solves the above-mentioned technical problems through the following technical solution: A trajectory recognition method based on distributed acoustic sensing utilizes a distributed acoustic sensing system to achieve high-precision quantitative measurement of external mechanical disturbances by demodulating the phase change of Rayleigh backscattered light. The method is characterized by the following steps: Step 1: Acquire the low SNR vibration signal collected by the DAS system within a unit acquisition time. First, use time-space double Gaussian smoothing and upper limit truncation to suppress random noise and abnormal high-value points, so that the vehicle trajectory appears as a continuous and smooth energy band in the spatiotemporal domain. Then, eliminate the influence of overall energy fluctuations by using line-direction L2 normalization, and use Hilbert envelope to amplify the amplitude changes caused by vehicle vibration, so that the preprocessed trajectory image has a higher signal-to-noise ratio and contrast, thus providing high-quality input for the subsequent VTR Net. Step 2: In VTR Net, the Global-Local Feature Collaborative Modeling Mechanism (GL-FCM) and the Multi-Resolution Attention Feature Map Recovery Architecture (AG-FMR) are adopted: GL-FCM combines the Mamba Like module with CNN to enhance the model's global attention to the trajectory; AG-FMR completes multi-scale and multi-channel feature fusion through the multi-resolution attention feature recovery architecture to eliminate the influence of noise on trajectory reconstruction. Step 3: The CEH-FA model robustness enhancement method for vehicle trajectory recognition is used. The CEH-FA is based on Hough transform and improves trajectory recognition accuracy through feature space mapping and energy focusing strategies, which alleviates the problem of weak model generalization ability caused by insufficient data. Finally, the full amount of vehicle trajectory data recognized within the unit acquisition time is output, and the unit acquisition time is 10s.

[0007] Based on the above technical solution, this application makes the following improvements and enhancements to the above technical solution: Furthermore, the local feature extraction module extracts information in the following manner: In GL-FCM, local feature extraction is performed by the Encoder module; let the input feature map of the Encoder be... Where C is the number of channels, H is the feature map height, and W is the feature map width. The core calculation process is as follows: , in, This is the local feature map output to Fusion or Decoder. Channel transformation convolution is used to map the input feature map to 2*C. This is a feature extraction convolution used for deep feature extraction in a 2*C channel dimension, and MaxPool is a downsampling pooling operation. This is the feature map that is input to the next level Encoder.

[0008] Furthermore, the global feature extraction module extracts information in the following manner: In GL-FCM, global feature modeling is implemented through ML Block, introducing the Mamba-Like LinearAttention (MLLA) module (built based on a state-space model, requiring the two-dimensional features to be converted into sequence input); VTR Net uses the inverse operations of Flatten and Fold to complete feature serialization and reconstruction. The Flatten calculation formula is as follows: , B represents the batch size. L The length of the feature sequence is denoted as ; after processing by the MLLA module, the spatial structure is restored by the inverse Fold transform to obtain the global feature map.

[0009] Furthermore, the global-local feature fusion module extracts information in the following manner: To fuse the local information extracted by the Encoder with the global information extracted by the ML Block, VTR Net uses element-wise addition in the Fusion module to achieve feature fusion, and combines convolution operations to adjust the channels, finally outputting to the Decoder. The calculation formula is as follows: , in Represents the output of the Fusion module. This represents the input feature map from the Encoder module. This represents the input feature map from the MLLA module. This represents convolution calculation.

[0010] Furthermore, the multi-resolution attention feature map recovery architecture includes the following modules: a multi-resolution feature map recovery module and a multi-resolution attention module.

[0011] Furthermore, the multi-resolution feature map recovery module includes the following modules: The core calculations of the Decoder module are as follows: , in, The feature map input via residual connections is used for the global-local feature collaborative modeling mechanism. This is the output feature map of the previous layer. The feature map is recovered after decoding. For feature map upsampling operation, This is a feature splicing operation along the channel dimension.

[0012] Furthermore, the multi-resolution attention module AG-FMR superimposes a lightweight ECA module on the output of each stage, calculated as follows: , Where X is the Decoder output feature map, For the Sigmoid function, This module performs channel-weighted operations and global average pooling operations. It effectively enhances cross-channel feature interaction, suppresses noise interference in trajectory recognition, and combines a multi-resolution attention pyramid architecture to extract semantic information at different levels, thus achieving high-precision vehicle trajectory recognition.

[0013] Furthermore, the trajectory detection using the Hough transform is as follows: (1) Straight lines in the spatiotemporal image are described in polar coordinates: in It is the perpendicular distance from the origin of the spacetime image coordinate system to the straight line. x is the angle between the horizontal axis and the perpendicular line; x and y represent the coordinates of each edge pixel in the image; the Hough line detection method calculates the coordinates of each edge pixel in the image according to the formula. and , construct array The voting is accumulated, and the peak vote corresponds to the pixels in the binary image most likely to be connected into a straight line; finally, the parameters are... Convert to Cartesian coordinates to reconstruct the detected straight line from the original image; the straight line vector can be formalized as: , representing the starting and ending coordinates of the line; (2) Vector space renormalization: Given the high sensitivity of cosine similarity calculation to vector direction, this application extends the feature dimension of straight line vectors: expanding the four-dimensional feature space of straight line vectors to a seven-dimensional vector space V, and extending the straight line vector from... Expand to (Added line length l (pixel value), line slope k (pixel value), and real space velocity v (km / h)); Using velocity as the evaluation criterion, line vectors with abnormal velocity are deleted, where the formula for calculating v is as follows: The spatial displacement is calculated for the spatiotemporal mapping of the vehicle trajectory map and the DAS signal. For time intervals; Calculate the degree of difference between any two vectors in vector space V using four parameters: Euclidean distance, slope, velocity, and length. Select the vector with the smallest average degree of difference as the normalized vector. Then calculate the intersection of any vector in V with... The similarity and difference are calculated as follows: in, Calculation of representative differences Represents the L2 norm. Represents a normalized vector; N is the total number of straight vectors contained in the vector space V; (3) Trajectory selection mechanism based on cosine similarity calculation: When using only the Hough transform to detect lines, a large number of redundant and falsely detected lines are prone to occur. To eliminate such interference, cosine similarity is introduced for trajectory line filtering: given two attributes... and cosine similarity The calculation formula is as follows: They are respectively and The modulus is calculated; the cosine similarity of the remaining vectors is calculated based on the normalized vector, and a threshold of 0.008 is set. Lines with similarity less than the threshold are retained, and lines with similarity greater than the threshold are deleted, which effectively improves the accuracy of trajectory recognition.

[0014] The main innovative aspects of this application are as follows: 1) Design the VTR Net deep neural network architecture, realize multi-scale spatiotemporal feature collaborative modeling through GL-FCM, and introduce AG-FMR to significantly improve trajectory recognition accuracy; 2) The CEH-FA algorithm is proposed to effectively enhance the stability and robustness of vehicle trajectory detection in complex environments; 3) Construct a signal preprocessing workflow for DAS that can stably output high-quality vehicle trajectory images. Attached Figure Description

[0015] Figure 1 Here is a flowchart of the vehicle trajectory detection algorithm; Figure 2 This is a structural diagram of a distributed acoustic sensing system. Figure 3 VTR Net architecture diagram; Figure 4 Here is the flowchart of the CEH-FA algorithm; Figure 5 This is the pseudocode for a trajectory selection mechanism based on cosine similarity calculation. Detailed Implementation

[0016] The following embodiments, in conjunction with the accompanying drawings, are merely for illustrating the technical solutions described in the claims and are not intended to limit the scope of protection of the claims.

[0017] Combination Figure 1 To extract complete and clear vehicle trajectories, this application proposes a trajectory recognition method based on distributed acoustic sensing, namely a joint optimization framework for DAS traffic monitoring, as follows: Step 1: Acquire low SNR vibration signals collected by the DAS system within a unit acquisition time. First, time-space double Gaussian smoothing and upper limit truncation are used to suppress random noise and abnormally high points, making the vehicle trajectory appear as a continuous and smooth energy band in the spatiotemporal domain. Subsequently, the influence of overall energy fluctuations is eliminated by directional L2 normalization, and the amplitude changes caused by vehicle vibration are amplified using Hilbert envelope, resulting in a preprocessed trajectory image with higher signal-to-noise ratio and contrast, thus providing high-quality input for subsequent VTR Net processing.

[0018] Step 2: In VTR Net, to address the lack of global trajectory awareness in traditional methods, this application innovatively designs a Global-Local Feature Collaborative Modeling Mechanism (GL-FCM) and a Multi-Resolution Attention-Guided Feature Map Restoration Architecture (AG-FMR). The former combines the Mamba-like module with CNN, effectively enhancing the model's global attention to the trajectory by leveraging its global modeling advantages while retaining the linear complexity (O(N)) of the Mamba structure. The latter completes multi-scale, multi-channel feature fusion through the multi-resolution attention feature restoration architecture, eliminating the impact of noise on trajectory reconstruction.

[0019] Step 3: Improving the model's generalization ability highly depends on the support of large-scale datasets. However, due to the scarcity of publicly available datasets in this field, training high-performance models using traditional data-driven paradigms faces significant challenges. This application proposes a robustness enhancement method for vehicle trajectory recognition—CEH-FA. This method is based on Hough transform and improves trajectory recognition accuracy through feature space mapping and energy focusing strategies, effectively alleviating the problem of weak model generalization ability caused by insufficient data. Finally, it outputs the full range of vehicle trajectory data identified within a unit acquisition time of 10 seconds.

[0020] 1. Distributed Acoustic Sensing (DAS) System: Combined with appendix Figure 2 The distributed acoustic sensing system structure of this application relies on the distributed acoustic sensing (DAS) technology to build an experimental device. By demodulating the phase change of Rayleigh backscattered light, it achieves high-precision quantitative measurement of external mechanical disturbances.

[0021] Signal acquisition utilizes a 1550nm distributed feedback (DFB) laser source, combined with chirped pulse modulation technology. After amplification by a two-stage erbium-doped fiber amplifier (EDFA), a 10dBm probe pulse is generated, enabling the system to achieve a vibration sensing performance of 5m spatial resolution and 7pε / √Hz noise floor. Based on the principle of coherent detection, a balanced detector (BPD) mixes the backscattered Rayleigh signal with the local oscillator light to generate an intermediate frequency electrical signal. After digital conversion at a 20kHz sampling frequency, the spatiotemporal distribution characteristics of strain disturbances along the fiber are reconstructed using a phase demodulation algorithm. The system has a dynamic range of up to 120dB and can simultaneously analyze vehicle vibration signal characteristics in the 20Hz-2.5kHz frequency band, providing high signal-to-noise ratio data support for machine learning-driven vehicle classification and trajectory reconstruction tasks.

[0022] 2. Data Preprocessing Methods Given the continuous characteristics of a vehicle in both time and space during its motion, a Gaussian window is introduced to smooth the signal, reducing the impact of random noise and enhancing its smoothness and continuity. The specific calculation formula is as follows: Where i (i=0,1,…,T-1) represents the time point of the signal in the time domain, and j (j=0,1,…,S-1) represents the spatial point of the signal in the spatial domain. A Gaussian kernel of size [2001, 2001] is used. Represents a Gaussian kernel of size [3,3]. Represents the L2 norm. This represents the convolution operation. Gaussian smoothing utilizes the continuity of vehicle motion in time and space to diffuse energy along the trajectory direction, making the energy band corresponding to the vehicle trajectory smoother and more continuous. Random noise is averaged and significantly attenuated after convolution. Row normalization further eliminates global amplitude variations, retaining only the relative strength relationships, which enhances trajectory contrast and provides stable input for subsequent envelope extraction and trajectory imaging.

[0023] The smoothed signal is truncated at its upper limit, and its envelope is calculated to extract feature information and enhance signal discernibility. The specific calculation formula is as follows: in Represents the cutoff value. represent The Hilbert transform is used to extract the vehicle's trajectory. DAS signals often contain maxima caused by mechanical vibrations or occasional instrument interference. These maxima can form "spots" during subsequent image processing, obscuring the true vehicle trajectory. Upper bound truncation limits these maxima to a reasonable range, preventing interference with subsequent analysis while preserving the core structure of the vehicle trajectory. The Hilbert transform converts the real signal into a complex analytic signal, whose magnitude can characterize the instantaneous energy envelope. When a vehicle passes, it generates a significant energy boost at the corresponding spatiotemporal location. Envelope operations can suppress phase oscillations while highlighting this energy change. The envelope[i,j] after envelope extraction is a non-negative real-valued matrix, which can be directly viewed as a grayscale image. Its high-value regions highly correspond to the vehicle's trajectory, allowing for visualization to obtain a clear and complete image of the vehicle trajectory, providing high-quality input for subsequent trajectory recognition algorithms.

[0024] 3. VTR-Net Architecture Design: Combination Figure 3 U-Net has become a benchmark model in the field of vehicle segmentation due to its feature reuse and pixel-level segmentation accuracy. However, its limited local receptive field makes it difficult to capture the non-stationary and discontinuous characteristics of vehicle motion in distributed fiber acoustic sensing (DAS) signals, which can easily lead to trajectory recognition breakpoints. To address this, this application proposes a novel vehicle trajectory recognition network, VTR Net (Vehicle Trajectory Recognition Network), with the specific structure as follows: Figure 2 As shown, the core innovative architecture includes a global-local feature collaborative modeling mechanism (GL-FCM) and a multi-resolution attention feature map recovery architecture (AG-FMR).

[0025] 1) Global-Local Feature Collaborative Modeling Mechanism (GL-FCM) The vehicle trajectory map contains both global and local information. To efficiently utilize both types of information, this application designs a hierarchical feature fusion architecture: the shallow network achieves global-local feature collaborative modeling through the Encoder-ML Block-Fusion module chain. The global features extracted by the ML Block and the local features generated by the Encoder are fused across modalities in the Fusion module and then output to the right-hand Decoder to participate in feature reconstruction. The fourth and fifth layers of the network are simplified into a single Encoder structure, focusing on extracting deep semantic information to reduce training complexity and accelerate inference.

[0026] (1) Local feature extraction module In GL-FCM, local feature extraction is performed by the Encoder module. Let the input feature map of the Encoder be... (C represents the number of channels, and H and W represent the height and width of the feature map, respectively). The core calculation process is as follows: in, This is the local feature map output to Fusion or Decoder. Channel transformation convolution is used to map the input feature map to 2*C. For feature extraction convolution, MaxPool is the downsampling pooling operation. This is the feature map that is input to the next level Encoder.

[0027] (2) Global Feature Extraction Module In GL-FCM, global feature modeling is implemented through the ML Block, which innovatively introduces the Mamba-LikeLinear Attention (MLLA) module (built based on a state-space model, requiring the two-dimensional features to be converted into sequence input). VTR Net uses the inverse operations of Flatten and Fold to complete feature serialization and reconstruction. The Flatten calculation formula is as follows: (B is the batch size, L=H×W is the sequence length). After processing by the MLLA module, the spatial structure is restored by the inverse Fold transform to obtain the global feature map.

[0028] MLLA is a linear attention module based on the Mamba architecture. Through parallelization and positional encoding enhancement, it retains the core advantages of Mamba's linear complexity and global modeling while solving its recursive computation bottleneck. To address the potential parameter count and latency issues of MLLA in deep networks, a layered strategy is adopted: the shallow layer utilizes MLLA's global perception advantage to model cross-regional trajectory associations of high-resolution features. After the feature maps are downsampled to a low-dimensional space, they are converted into convolutions to extract deep semantic information, achieving a dynamic balance between computational efficiency and representational capability.

[0029] (3) Global-Local Feature Fusion Module To fuse the local information extracted by the Encoder with the global information extracted by the ML Block, VTR Net uses element-wise addition in the Fusion module to achieve feature fusion, and combines convolution operations to adjust the channels, finally outputting to the Decoder. The calculation formula is as follows: in Represents the output of the Fusion module. This represents the input feature map from the Encoder module. This represents the input feature map from the MLLA module.

[0030] 2) Multi-resolution attention feature map recovery architecture (AG-FMR) The quality of feature map recovery directly affects the accuracy of vehicle trajectory recognition. To improve the model's recognition performance, VTR Net uses the Decoder and Efficient Channel Attention (ECA) modules to construct a multi-resolution attention pyramid architecture, as shown in the AG-FMR branch in Figure 1.

[0031] (1) Multi-resolution feature map recovery module The core calculations of the Decoder module are as follows: in, The feature map input via residual connections is used for the global-local feature collaborative modeling mechanism. This is the output feature map of the previous layer. For feature map upsampling operation, This is a feature splicing operation along the channel dimension.

[0032] (2) Multi-resolution attention module To enhance feature map recovery, AG-FMR superimposes a lightweight ECA module at each output stage. The calculation formula is as follows: Where X is the Decoder output feature map, For the Sigmoid function, This module performs channel-weighted operations. It effectively enhances cross-channel feature interaction, suppresses noise interference in trajectory recognition, and, combined with a multi-resolution attention pyramid architecture, achieves semantic information extraction at different levels, thus enabling high-precision vehicle trajectory recognition.

[0033] 4) CEH-FA Algorithm Implementation Combination Figure 3 Improving the generalization ability of a model highly depends on large-scale datasets, but currently there is a lack of publicly available datasets in this field, posing a significant challenge to training high-performance models using traditional data-driven methods. To address this, this application proposes a model robustness enhancement method suitable for vehicle trajectory recognition, and the algorithm flow is shown in Figure 3.

[0034] (1) Trajectory detection based on Hough transform Straight lines in a spacetime graph are described using polar coordinates: in It is the perpendicular distance from the origin of the spacetime image coordinate system to the straight line. It is the angle between the horizontal axis and the perpendicular line. The Hough line detection method calculates this angle for each edge pixel of the image according to a formula. and , construct array The voting process is performed, and the peak vote corresponds to the pixels in the binary image most likely to be connected into a straight line. Finally, the parameters are... Convert to Cartesian coordinates to reconstruct the detected lines in the original image. The line vector can be formalized as: , representing the starting and ending coordinates of the line.

[0035] (2) Vector space renormalization Given the high sensitivity of cosine similarity calculation to vector direction, this application extends the feature dimension of straight line vectors: expanding the four-dimensional feature space of straight line vectors to a seven-dimensional vector space V, and extending the straight line vector from... Expand to (Added line length l (pixel value), line slope k (pixel value), and real space velocity v (km / h)); Using velocity as the evaluation criterion, line vectors with abnormal velocity are deleted, where the formula for calculating v is as follows: , , , The spatial displacement is calculated for the spatiotemporal mapping of the vehicle trajectory map and the DAS signal. For time intervals.

[0036] Calculate the degree of difference between any two vectors in vector space V using four parameters: Euclidean distance, slope, velocity, and length. Select the vector with the smallest average degree of difference as the normalized vector. Then calculate the intersection of any vector in V with... The similarity and difference are calculated as follows: in, Calculation of representative differences Represents the L2 norm. Represents a canonical vector.

[0037] (3) Trajectory selection mechanism based on cosine similarity calculation Using only the Hough transform to detect lines easily leads to a large number of redundant and falsely detected lines. To eliminate this interference, cosine similarity is introduced for trajectory line filtering: given two attributes... and cosine similarity The calculation formula is as follows: They are respectively and The modulus length is determined. Cosine similarity is calculated for the remaining vectors based on the normalized vector, with a threshold of 0.008. Lines with similarity less than the threshold are retained, while lines with similarity greater than the threshold are deleted, effectively improving trajectory recognition accuracy. The specific processing procedure is as follows: Figure 5 As shown, the algorithm first sorts and augments the straight line data obtained from the Hough transform, and selects the optimal straight line as the reference vector. Then, it enters a filtering loop, calculating the cosine distance between each candidate vector and the reference vector. A distance filtering threshold of 0.008 is set, and a deduplication strategy is implemented: when the calculated cosine distance is greater than this threshold, it indicates that the current vector is effectively distinguishable from the reference vector, and it is retained in the output list; conversely, if the distance is less than or equal to the threshold, it is considered a redundant trajectory and is discarded.

[0038] Real-world experimental results show that, compared with traditional methods, this solution achieves significant improvements in trajectory recognition accuracy (IoU 56.13%) and structural integrity (SSIM 97.65%), providing an economical, efficient, and easy-to-deploy solution for intelligent traffic monitoring.

[0039] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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.

Claims

1. A trajectory recognition method based on distributed acoustic sensing, which utilizes a distributed acoustic sensing system to achieve high-precision quantitative measurement of external mechanical disturbances by demodulating the phase change of Rayleigh backscattered light, characterized in that... The trajectory recognition method includes the following steps: Step 1: Acquire the low SNR vibration signal collected by the DAS system within a unit acquisition time. First, use time-space double Gaussian smoothing and upper limit truncation to suppress random noise and abnormal high-value points, so that the vehicle trajectory appears as a continuous and smooth energy band in the spatiotemporal domain. Then, eliminate the influence of overall energy fluctuations by using line-direction L2 normalization, and use Hilbert envelope to amplify the amplitude changes caused by vehicle vibration, so that the preprocessed trajectory image has a higher signal-to-noise ratio and contrast, thus providing high-quality input for the subsequent VTR Net. Step 2: In VTR Net, the Global-Local Feature Collaborative Modeling Mechanism (GL-FCM) and the Multi-Resolution Attention Feature Map Recovery Architecture (AG-FMR) are adopted: GL-FCM combines the Mamba Like module with CNN to enhance the model's global attention to the trajectory; AG-FMR completes multi-scale and multi-channel feature fusion through the multi-resolution attention feature recovery architecture to eliminate the influence of noise on trajectory reconstruction. Step 3: The CEH-FA model robustness enhancement method for vehicle trajectory recognition is used. The CEH-FA is based on Hough transform and improves trajectory recognition accuracy through feature space mapping and energy focusing strategies, which alleviates the problem of weak model generalization ability caused by insufficient data. Finally, the full amount of vehicle trajectory data recognized within the unit collection time is output.

2. The trajectory recognition method based on distributed acoustic sensing according to claim 1, characterized in that, In the GL-FCM, the local feature extraction module extracts information in the following way: Local feature extraction is performed by the Encoder module; let the input feature map of the Encoder be... Where C is the number of channels, H is the feature map height, and W is the feature map width. The core calculation process is as follows: , in, This is the local feature map output to Fusion or Decoder. Channel transformation convolution is used to map the number of channels in the input feature map to 2*C. This is a feature extraction convolution used for deep feature extraction in a 2*C channel dimension, and MaxPool is a downsampling pooling operation. This is the feature map that is input to the next level Encoder.

3. The trajectory recognition method based on distributed acoustic sensing according to claim 1, characterized in that, In the full GL-FCM, the local feature extraction module extracts information in the following way: Global feature modeling is implemented through ML Block, introducing the MLLA module; VTR Net uses the inverse operation of Flatten and Fold to complete feature serialization and reconstruction. The Flatten calculation formula is as follows: ， B represents the batch size. L The length of the feature sequence; After processing by the MLLA module, the spatial structure is restored by the inverse Fold transform to obtain the global feature map.

4. The trajectory recognition method based on distributed acoustic sensing according to claim 1, characterized in that, In the GL-FCM, the global-local feature fusion module extracts information in the following way: To fuse the local information extracted by the Encoder with the global information extracted by the ML Block, VTR Net uses element-wise addition in the Fusion module to achieve feature fusion, and combines convolution operations to adjust the channels, finally outputting to the Decoder. The calculation formula is as follows: ， in Represents the output of the Fusion module. This represents the input feature map from the Encoder module. This represents the input feature map from the MLLA module. This represents convolution calculation.

5. The trajectory recognition method based on distributed acoustic sensing according to claim 1, characterized in that, The multi-resolution attention feature map recovery architecture includes the following modules: multi-resolution feature map recovery module and multi-resolution attention module.

6. The trajectory recognition method based on distributed acoustic sensing according to claim 5, characterized in that, The multi-resolution feature map recovery module includes the following modules: The core calculations of the Decoder module are as follows: ， in, The feature map input via residual connections is used for the global-local feature collaborative modeling mechanism. This is the output feature map of the previous layer. The feature map is recovered after decoding. For feature map upsampling operation, This is a feature splicing operation along the channel dimension.

7. The trajectory recognition method based on distributed acoustic sensing according to claim 5, characterized in that, The multi-resolution attention module AG-FMR superimposes a lightweight ECA module on the output of each stage, and the calculation formula is as follows: ， Where X is the Decoder output feature map, For the Sigmoid function, The module performs channel-weighted operations and global average pooling operations. This module can effectively enhance cross-channel feature interaction, suppress noise interference in trajectory recognition, and combine multi-resolution attention pyramid architecture to extract semantic information at different levels, thus achieving high-precision vehicle trajectory recognition.

8. The trajectory recognition method based on distributed acoustic sensing according to claim 1, characterized in that, The trajectory detection using the Hough transform is as follows: (1) Straight lines in the spatiotemporal image are described in polar coordinates: ， in It is the perpendicular distance from the origin of the spacetime image coordinate system to the straight line. It is the angle between the horizontal axis and the perpendicular line, where x and y represent the coordinates of each edge pixel in the image; the Hough line detection method calculates the coordinates of each edge pixel in the image according to the formula. and , construct array The voting is accumulated, and the peak vote corresponds to the pixels in the binary image most likely to be connected into a straight line; finally, the parameters are... Convert to Cartesian coordinates to reconstruct the detected straight line from the original image; the straight line vector can be formalized as: x 1 y 1 x 2 y 2 , representing the starting and ending coordinates of the line; (2) Vector space renormalization: Given the high sensitivity of cosine similarity calculation to vector direction, this application extends the feature dimension of straight line vectors: expanding the four-dimensional feature space of straight line vectors to a seven-dimensional vector space V, and extending the straight line vector from... x 1 y 1 x 2 y 2 Expand to x 1 y 1 x 2 y 2 l k v Add line length l, line slope k, and real space velocity v; using velocity as the evaluation criterion, delete line vectors with abnormal velocity, where the formula for calculating v is as follows: ， ， ， The spatial displacement is calculated for the spatiotemporal mapping of the vehicle trajectory map and the DAS signal. For time intervals; Calculate the degree of difference between any two vectors in vector space V using four parameters: Euclidean distance, slope, velocity, and length. Select the vector with the smallest average degree of difference as the normalized vector. Then calculate the intersection of any vector in V with... The similarity and difference are calculated as follows: ， , in, Calculation of representative differences Represents the L2 norm. Represents a canonical vector, where N is the total number of straight vectors contained in the vector space V; (3) Trajectory selection mechanism based on cosine similarity calculation: When using only the Hough transform to detect lines, a large number of redundant and falsely detected lines are prone to occur. To eliminate such interference, cosine similarity is introduced for trajectory line filtering: given two attributes... and cosine similarity The calculation formula is as follows: ， They are respectively and The modulus of the vector is determined; the cosine similarity of the remaining vectors is calculated based on the normalized vector, and a threshold of 0.008 is set. Lines with similarity less than the threshold are retained, and lines with similarity greater than the threshold are deleted, which effectively improves the accuracy of trajectory recognition.

Images