A das vehicle weak trajectory extraction method based on space-time feature enhancement
By constructing a DAS vehicle weak trajectory extraction network based on spatiotemporal feature enhancement, and employing residual spatial attention denoising and a hybrid feature bottleneck layer, combined with a soft skeleton loss function, the problems of low signal-to-noise ratio and poor trajectory connectivity in the DAS system are solved, achieving efficient weak vehicle trajectory segmentation and continuity extraction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-19
- Publication Date
- 2026-07-21
AI Technical Summary
In complex traffic scenarios, vehicle trajectory segmentation in DAS systems faces problems such as extremely low signal-to-noise ratio, weak signals being easily submerged by noise, and poor trajectory connectivity. Existing methods are unable to effectively extract and maintain the integrity and continuity of weak vehicle trajectories.
A spatiotemporal feature-based enhancement method is adopted, which constructs an end-to-end trajectory extraction network by combining residual spatial attention denoising, hybrid feature bottleneck layer and multi-scale Hough fusion output module with soft skeleton loss function, so as to achieve directional enhancement and topological continuity constraint of weak signals.
It effectively suppresses background noise, maintains the integrity of weak signals, improves the topological connectivity and extraction accuracy of the trajectory, solves the problem of signal loss and breakage in DAS data, and achieves high-resolution vehicle trajectory segmentation.
Smart Images
Figure CN122435560A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to vehicle trajectory extraction and recognition technology, specifically to a DAS vehicle weak trajectory extraction method based on spatiotemporal feature enhancement, belonging to the interdisciplinary field of distributed fiber optic sensing data processing and computer vision. Background Technology
[0002] Distributed Acoustic Sensing (DAS) technology has shown great potential in highway traffic monitoring due to its advantages of long-distance and continuous monitoring. Data collected by DAS systems is typically converted into a vibration spatiotemporal image (waterfall plot), where vehicle trajectories exhibit specific diagonal line characteristics. However, vehicle trajectory segmentation faces several challenges in complex real-world traffic scenarios: First, the signal-to-noise ratio is extremely low. DAS data often contains strong background noise and horizontal interference stripes, and traditional hard denoising methods (such as thresholding and Frangi filtering) easily erase weak vehicle trajectory signals while suppressing noise. Second, the target scale is extremely small and features are singular. Vehicle trajectories typically occupy only a small number of pixels in the entire image and appear as thin, elongated lines; conventional convolutional neural networks easily lose these high-resolution features after multiple downsampling. Finally, trajectory connectivity is poor. Conventional pixel-based classification loss functions (such as IoU loss and cross-entropy loss) lack attention to topology, leading to frequent misclassifications in weak signal areas and resulting in discontinuous trajectories that severely impact subsequent vehicle speed estimation and traffic flow statistics. Therefore, there is an urgent need for a DAS vehicle trajectory segmentation method that can achieve end-to-end soft denoising, maintain high resolution of linear features, and constrain topological continuity. Summary of the Invention
[0003] To address the aforementioned shortcomings of existing technologies, the present invention aims to propose a method for extracting weak vehicle trajectories in DAS (Digital Automotive Image) based on spatiotemporal feature enhancement. This invention can solve the problems of weak signal loss and trajectory breakage under strong noise interference in DAS spatiotemporal images.
[0004] The technical solution of this invention is implemented as follows:
[0005] A method for extracting weak trajectories of DAS vehicles based on spatiotemporal feature enhancement includes the following steps:
[0006] 1) Acquire spatiotemporal images of highway vibrations from distributed fiber optic acoustic sensors with known vehicle trajectories, and preprocess them; divide the preprocessed spatiotemporal images of vibrations into training and testing sets;
[0007] 2) Input the spatiotemporal vibration images from the training set into the pre-constructed trajectory extraction network to train the trajectory extraction network. The trajectory extraction network identifies and outputs the vehicle trajectory corresponding to the spatiotemporal vibration images, thus obtaining the trained trajectory extraction network. The trajectory extraction network includes a denoising module, an encoder, a decoder, and a multi-scale Hough fusion output module connected in sequence. The denoising module uses a residual space attention mechanism to directionally enhance weak vehicle trajectory signals and suppress background speckle noise.
[0008] 3) After performing the same preprocessing as in step 1), the vibration spatiotemporal image acquired by the actual distributed fiber acoustic sensor to be tested is input into the trained trajectory extraction network. The trajectory extraction network then outputs a vehicle binary trajectory segmentation mask with topological continuity.
[0009] Furthermore, the specific implementation process of the preprocessing in step 1) is as follows:
[0010] Background subtraction: The vibration spatiotemporal image is converted into a floating-point two-dimensional matrix, and the median vector of each row of the two-dimensional matrix is calculated; the original two-dimensional matrix is subtracted from the median vector and multiplied by a preset coefficient to remove horizontal stripe interference caused by fiber Rayleigh scattering or laser frequency drift.
[0011] Intelligent black level clamping: Calculate the percentile of the overall brightness distribution of the image matrix after stripe removal; extract the pixel value with brightness at a certain percentile as the dynamic noise baseline; set the pixel values in the image matrix that are less than the baseline to 0 to remove dark background speckle noise;
[0012] Adaptive enhancement: The processed matrix is Min-Max normalized to the range of 0-255, and contrast-limited adaptive histogram equalization is applied to obtain the final preprocessed input image. .
[0013] Furthermore, considering the characteristics of strong background speckle noise and extremely weak linear signal accompanying DAS vibration spatiotemporal images, the denoising module includes:
[0014] Feature pooling unit: The input DAS vibration spatiotemporal image features are subjected to average pooling and max pooling along the channel dimension respectively; average pooling is used to obtain the background color trend information of a large area in the image, and max pooling is used to capture the significant texture features of weak vehicle trajectories that present a diagonal linear distribution.
[0015] Mask generation unit: After concatenating the acquired background color trend information and the faint vehicle trajectory significant texture features in the channel dimension, a spatial attention mask with a value range of 0 to 1 is generated through a convolutional layer and a Sigmoid activation function. This mask represents the spatial distribution probability of the vehicle trajectory in the spatiotemporal image.
[0016] The residual enhancement unit multiplies the original input DAS vibration spatiotemporal image features with the spatial attention mask element-wise to obtain a primary feature map that suppresses background noise. Then, the primary feature map is added to the original input DAS vibration spatiotemporal image features as a residual, thereby achieving directional numerical multiplication of the weak vehicle trajectory region with high confidence while preserving the original signal background, and outputting the denoised and enhanced feature map to the subsequent encoder.
[0017] Furthermore, considering the physical characteristics of DAS vehicle trajectories in spatiotemporal images, which have long spans, varying thicknesses, and different inclination angles, a hybrid feature bottleneck layer is set between the encoder and the decoder. The hybrid feature bottleneck layer includes a parallel CNN feature extraction branch and a Transformer self-attention branch, both of which are used to receive deep features output by the encoder.
[0018] CNN Feature Extraction Branch: Contains multiple parallel multi-scale dilated convolution branches, combined with square, horizontal and vertical receptive fields, to match and extract local vehicle trajectory texture and geometric orientation features with different tilt angles and widths; Transformer Self-Attention Branch: Preserves the spatial location prior of the spatiotemporal image through two-dimensional absolute position encoding, and uses a multi-head self-attention mechanism to capture long-distance, global trajectory coherence dependencies across the entire DAS vibration spatiotemporal image;
[0019] Finally, the outputs of the two branches are concatenated along the channel dimension, and local trajectory details and global coherence information are fused through a convolutional layer before being output to the decoder.
[0020] Furthermore, in response to the problem that the weak trajectory of DAS vehicles appears as a thin, elongated line structure occupying only a few pixels in the spatiotemporal image, and that multiple downsampling can easily lead to the loss of this high-resolution feature, the encoder adopts a feature extraction network based on dilated convolution to maintain feature resolution.
[0021] The encoder maintains regular convolution and downsampling in the shallow network to extract basic features. In the deep network, regular convolution is replaced with dilated convolution, and the downsampling stride is modified to 1. The feature map size reduction of the deep network is canceled.
[0022] Specifically, a dilation rate of 2 is set in the penultimate stage of the deep network and a dilation rate of 4 is set in the final stage. This expands the receptive field without increasing the number of additional parameters, ensuring that the resolution of the deep feature map is always maintained at the preset ratio. This guarantees that the extremely fine vehicle trajectory features are not over-compressed and disappear in the deep network.
[0023] Furthermore, the decoder employs a 4-stage cascaded striped convolutional decoder; each stage of the decoder receives the upsampled features from the previous stage and the skip connection features corresponding to the encoder; the fused features are fed into parallel horizontal and vertical branches:
[0024] Horizontal branches: containing kernel size of Asymmetric convolution;
[0025] Vertical branches: contain kernel size of Asymmetric convolution;
[0026] The outputs of the two branches are concatenated and batch normalized before entering the next level.
[0027] Furthermore, the multi-scale Hough fusion output module collects multi-scale features output from the 4-level decoder; each scale feature is equipped with a prior Hough block, and the prior Hough block is average pooled along the height and width dimensions to obtain a one-dimensional vector. Attention weights are calculated through one-dimensional convolution and broadcast and added. Finally, a weight map is formed by Sigmoid activation; all weighted features are bilinearly upsampled to the highest resolution size, stitched together, and then fused using a convolutional layer for final feature fusion. Finally, a classification logarithmic map is output through a 1×1 convolution.
[0028] Furthermore, during the training phase, a combined loss function including soft skeleton loss is used to optimize the parameters of the trajectory extraction network; the soft skeleton loss is used to constrain the network at the topological level, penalizing the fragmentation of vehicle trajectories; the calculation process of the soft skeleton loss includes:
[0029] Morphological erosion and dilation operations are approximated by differentiable minima pooling and maxma pooling operations, respectively.
[0030] By approximate erosion and dilation operations, the predicted soft skeleton and the real label skeleton are iteratively extracted from the trajectory probability map and the real label map predicted by the trajectory extraction network, respectively.
[0031] The degree of overlap between the predicted trajectory probability map and the real label skeleton is calculated as the topological accuracy, and the degree of overlap between the predicted soft skeleton and the real label map is calculated as the topological sensitivity.
[0032] The centerline intersection-union ratio is obtained by taking the harmonic mean of the topology accuracy and topology sensitivity. , This is known as soft skeleton loss.
[0033] Furthermore, during training, auxiliary convolutional heads are added to the feature outputs of the second and third levels of the decoder to generate auxiliary prediction maps. These auxiliary prediction maps are then scaled to the original image size, and auxiliary loss is calculated based on them as part of the total loss.
[0034] Furthermore, it also includes an auxiliary mask prediction module for training. The Layer 3 output of the encoder is connected to the auxiliary mask prediction module, which first generates a random binary mask with a preset coverage. ; Combine the Layer 3 output features of the encoder with After random occlusion is achieved through multiplication, the original features are reconstructed by the predictor. During training, the auxiliary mask reconstruction loss is calculated based on the reconstructed original features so that the encoder can learn the contextual geometric continuity of the vibration spatiotemporal image, thereby improving the self-repair capability of the fracture trajectory. The auxiliary mask reconstruction loss is part of the total loss.
[0035] Compared with the prior art, the present invention has the following beneficial effects:
[0036] 1. Fidelity-Preserving Noise Reduction and Signal Enhancement: This invention abandons traditional destructive preprocessing and innovatively designs a spatial attention-based denoising front-end at the network input end. Unlike traditional denoising, this invention multiplies the input image with a mask and then adds the residuals to the original image. This allows the network to adaptively suppress background noise while performing numerical "directional additive enhancement" on weak vehicle trajectory regions, ensuring the survival rate of target features under extremely low signal-to-noise ratios and achieving directional enhancement of weak signals.
[0037] 2. Significantly Improved Topological Connectivity: Addressing the weakness of conventional pixel-level loss functions in their insensitivity to "broken lines," this invention introduces a differentiable soft skeleton loss during the training phase, in addition to conventional IoU and Focal Loss. This loss utilizes differentiable Min-Pooling and Max-Pooling approximate morphological erosion and dilation to iteratively extract the skeleton of the predicted image and the true label, calculate the topological overlap, and shift the penalty focus from independent pixel errors to the topological structure errors of the trajectory. This forces the network to learn contextual continuity and effectively closes trajectory breakpoints.
[0038] 3. Perfect Integration of Global and Local Features: This invention constructs a hybrid bottleneck layer based on parallel CNN and Transformer between the encoder and decoder. The CNN branch uses multi-scale receptive fields to capture local textures such as trajectory thickness; the Transformer branch uses a self-attention mechanism to capture long-distance dependencies across the entire spatiotemporal image. The fusion of the two takes into account both local details and global continuity, further consolidating the completeness of ultra-long trajectory extraction. Attached Figure Description
[0039] Figure 1 - Overall flowchart of the DAS weak trajectory extraction method based on spatiotemporal feature enhancement of this invention.
[0040] Figure 2 -Schematic diagram of the trajectory extraction network structure of this invention.
[0041] Figure 3 - Visual effect comparison diagram of the test set in the embodiments of the present invention; wherein, (a) original input; (b) prediction result of the traditional method U-Net; (c) prediction result of the present invention. Detailed Implementation
[0042] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0043] See Figure 1 This invention proposes a method for extracting weak vehicle trajectories using DAS based on spatiotemporal feature enhancement. The complete execution flow of this method includes the following steps:
[0044] Step 1: Acquisition and Preprocessing of DAS Vibration Spatiotemporal Images. Spatiotemporal images of highway vibrations collected by distributed fiber optic acoustic sensors with known vehicle trajectories are acquired and preprocessed. The preprocessed vibration spatiotemporal images are divided into training and testing sets. In actual traffic monitoring, the raw data collected by distributed fiber optic acoustic sensing (DAS) systems typically contains a large amount of horizontal stripe noise. To remove strong noise while preserving weak vehicle trajectory signals, this invention employs a Smart Clamp preprocessing algorithm. The specific processing procedure is as follows:
[0045] Background subtraction: The original input vibration spatiotemporal image is converted into a floating-point two-dimensional matrix, and the median vector of each row of the image two-dimensional matrix is calculated; the original two-dimensional matrix is subtracted from the median vector and then multiplied by a preset coefficient (preferably 0.9), thereby effectively removing horizontal stripe interference caused by fiber Rayleigh scattering or laser frequency drift;
[0046] Intelligent black level clamping: Calculate the percentile of the overall brightness distribution of the image matrix after stripe removal. Given that background pixels occupy the vast majority of DAS traffic images, this embodiment extracts the pixel value with the brightness ranking of 20% as the dynamic noise baseline (Noise Floor); pixel values in the image matrix with a noise floor lower than this baseline are directly set to 0 (pure black). This operation can completely remove dark background speckle noise without damaging the high-brightness effective signal.
[0047] Adaptive enhancement: The processed matrix is Min-Max normalized to the range of 0~255, and contrast-limited adaptive histogram equalization (CLAHE, preferably with a cropping limit of 2.0 and a grid size of 8×8) is applied to obtain the final preprocessed input image. .
[0048] Step 2: Construct a trajectory extraction network based on spatiotemporal feature enhancement. This involves processing the pre-processed spatiotemporal vibration images from the training set. (in For batch size, and The spatial resolution (preferably 512×512) is input into a pre-built trajectory extraction network to train the trajectory extraction network. The trajectory extraction network identifies and outputs the vehicle trajectory corresponding to the vibration spatiotemporal image, thus obtaining the trained trajectory extraction network, which is then tested using a test set.
[0049] Step 3: After performing the same preprocessing as in Step 1, the vibration spatiotemporal image acquired by the actual distributed fiber acoustic sensor to be tested is input into the trained trajectory extraction network. The trajectory extraction network then outputs a vehicle binary trajectory segmentation mask with topological continuity.
[0050] The core of this invention lies in the construction of the trajectory extraction network and the setting of the loss function during training. The trajectory extraction network of this invention includes a learnable denoising front-end module, an encoder, a hybrid feature bottleneck layer, a decoder, and a multi-scale Hough fusion output module, connected sequentially. Specific descriptions of each module are as follows, and can be found at the same time. Figure 2 .
[0051] 1) Learnable denoising front-end module (Spatial Attention)
[0052] To achieve end-to-end soft denoising and overcome the trajectory fragmentation caused by traditional hard threshold filtering, this invention sets up a residual space attention mechanism at the front end of the network. This mechanism directionally enhances weak vehicle trajectory signals and suppresses background speckle noise, outputting an enhanced feature map. The learnable denoising front-end module includes...
[0053] Feature pooling unit: for input features Mean pooling and max pooling are performed on the channel dimension respectively to capture background color information and local salient texture features:
[0054]
[0055] Mask generation unit: After concatenating average pooling and max pooling along the channel dimension, the result is fed into the receptive field. In the convolutional layer, and activated by the Sigmoid function, a value range of [value range missing] is generated. Spatial mask matrix between .
[0056] Residual Enhancement Unit: Multiplies the input image with the spatial attention mask to obtain a denoised feature map, then adds the residual to the original input image to complete feature enhancement, and outputs it to the subsequent encoder. The residual connection formula is as follows:
[0057]
[0058] in This indicates element-wise multiplication. This design allows the network to perform targeted numerical multiplication on high-confidence trajectory regions while preserving the original signal.
[0059] 2) Dilated ResNet Encoder
[0060] The encoder is used to perform multi-scale feature extraction on the enhanced feature map to obtain semantic features at different levels. The encoder in this invention is a dilated convolutional encoder. The input is fed into a dilated encoder based on an improved ResNet-34 for multi-scale feature extraction. To avoid the loss of extremely fine vehicle trajectories (typically only 1-3 pixels wide) in deep feature maps due to continuous downsampling, this invention modifies the stride of Layer 3 in ResNet-34 to 1 and sets the dilation rate to 2; it also modifies the stride of Layer 4 to 1 and sets the dilation rate to 4. This improvement ensures that the resolution of the deep feature maps remains at 1 / 8 of the original image (i.e., 64×64), greatly expanding the receptive field without increasing the number of additional parameters and guaranteeing the integrity of the fine trajectory features. The encoder of this invention uses ResNet as its backbone, replacing the regular convolutions in its deep network with dilated convolutions to maintain the feature map at 1 / 8 of its original size, preventing the disappearance of fine trajectories in deep networks due to over-downsampling.
[0061] 3) Hybrid Bottleneck Layer
[0062] The hybrid feature bottleneck layer is positioned between the encoder and decoder and includes parallel CNN feature extraction branches and Transformer self-attention branches for receiving the output features of Layer 4.
[0063] The CNN branch (Multi-Scale Context Extractor (MSCE)) employs multi-scale dilated convolutions to extract local trajectory texture and geometric orientation features. Specifically, it utilizes features containing squares ( )vertical( ) and level ( Multi-scale dilated convolution of the receptive field, as well as horizontal and vertical pooling branches, comprehensively capture the local texture and directional features of vehicle trajectories;
[0064] Transformer branch: Used to serialize feature maps and capture long-range global trajectory dependency information; specifically, channel projection is performed through 1×1 convolution, and after adding two-dimensional absolute position encoding, the features are serialized and input into a 2-layer Transformer encoder containing multi-head self-attention, which is used to capture global trajectory dependencies of up to hundreds of pixels.
[0065] Finally, the outputs of the two branches are concatenated along the channel dimension and fused through a convolutional layer before being output to the decoder.
[0066] 4) Stripe Conv Decoder (SCD)
[0067] Because vehicle trajectories in DAS spatiotemporal images exhibit distinct diagonal lines, the feature matching efficiency of square convolution kernels is low. This invention employs a striped convolution structure in its decoder, specifically a striped convolution decoder, which includes horizontal... Asymmetric convolution kernel and vertical direction Asymmetric convolutional kernels capture horizontal and vertical features of the trajectory, restoring the spatial resolution of the image. This invention replaces square convolutional kernels with striped convolutional kernels that have extremely asymmetric receptive fields, allowing for a natural fit and extraction of vehicle trajectory features exhibiting a diagonally linear distribution in vibration spatiotemporal images. This invention actually employs a four-stage cascaded striped convolutional decoder. Each decoder stage receives the upsampled features from the previous stage and the skip connection features corresponding to the encoder. The fused features are fed into two parallel striped convolutional branches:
[0068] Horizontal branches: containing kernel size of Asymmetric convolution;
[0069] Vertical branches: contain kernel size of Asymmetric convolution.
[0070] The outputs of the two branches are concatenated and batch normalized before entering the next stage. Striped convolution acts like a "compass," accurately aligning and extracting linear features while filtering out irregular speckle noise.
[0071] 5) Multi-scale Hough Fusion Output Module (HHTFD)
[0072] The multi-scale Hough fusion output module collects multi-scale features from the output of the 4-level decoder. For each scale of features, a Prior Hough Block is provided: the Prior Hough Block is average pooled along the height and width dimensions to obtain a one-dimensional vector, attention weights are calculated through one-dimensional convolution and broadcast and added, and finally a weight map is formed by sigmoid activation. All weighted features are bilinearly upsampled to the highest resolution size, concatenated, and then fused using a 3×3 convolutional layer, followed by a 1×1 convolution to output a classification logits map.
[0073] See Figure 2 The invention also includes a training-aided mask prediction module (MPM). The MPM is connected to the Layer 3 output of the encoder, and generates a random binary mask with a coverage of 30%. The Layer 3 output features of the encoder are compared with... After random occlusion is achieved through multiplication, the original features are attempted to be reconstructed through a predictor (containing two convolutional layers). During training, an auxiliary mask reconstruction loss (mean squared error of the reconstructed region MSE Loss) is calculated based on the reconstructed original features, and the auxiliary mask reconstruction loss is included as part of the total loss. This self-supervised task forces the encoder to learn the contextual geometric continuity of the spatiotemporal image, thereby improving the self-repair capability of broken trajectories.
[0074] Loss function construction and model training
[0075] To address the issues of extremely low pixel ratios in DAS trajectory (significant imbalance between positive and negative samples) and susceptibility to fragmentation, this invention designs a multi-dimensional combined loss function for network training supervision. Let the network's output probability map be... The real label is The combined loss function of the main output is defined as:
[0076] ;
[0077] in, For soft skeleton loss, To compare the losses, For the focus of loss, For Dice's loss, The loss is the Hausdorff distance transformation loss; to These are the weighting coefficients corresponding to the loss of each component, and the weighting coefficients of the soft skeleton loss. The weighting coefficients are greater than those for other losses. The following is a specific example of the weighting coefficients for each loss.
[0078]
[0079] The calculation method for the loss of each component is as follows:
[0080] 1) Soft Skeleton Loss This loss is used to constrain the network at the topological level, penalizing the breakage of vehicle trajectories and forcing the model to repair broken trajectories. Differentiable min-pooling and max-pooling operations are used to approximate morphological erosion and dilation operations, respectively. Through approximate erosion and dilation operations, the predicted soft skeleton is iteratively extracted from the network's predicted trajectory probability map and the ground truth label map, respectively. and real label skeleton .
[0081] Soft erosion (soft_erode): utilizes a kernel size of and Max pooling is used to process negative images, approximating local minima.
[0082]
[0083] Soft dilate: utilizing Max pooling implementation: .
[0084] The opening operation is defined as a soft erosion followed by a soft expansion. This is achieved through iterative calculations. ,
[0085] Next, the degree of overlap between the predicted trajectory probability map and the ground truth label skeleton is calculated as the topological precision (T-prec), and the degree of overlap between the predicted soft skeleton and the ground truth label map is calculated as the topological sensitivity (T-sens); the formula for calculating topological precision is... Topology sensitivity calculation formula .
[0086] Finally, the harmonic mean of the topology accuracy and topology sensitivity is used to obtain the centerline intersection-union ratio (CUNR). ), and will The value of is included in the backpropagation as a soft skeleton loss term, i.e. .
[0087] 2) Local classification loss: where Focal Loss( To address the imbalance between easy and difficult samples, Dice Loss ( ) and Soft IoU Loss Constrain global area and shape intersection-union ratio consistency.
[0088] 3) Boundary distance loss (Hausdorff DT Loss) ): For real tags Perform a distance transformation to generate a distance field. Pixels that make incorrect predictions are penalized, and the penalty weight increases exponentially the further the pixel is from the true trajectory, thereby completely eliminating isolated false positive noise points that are far from the true trajectory.
[0089] In the training phase, in addition to calculating the combined loss on the main output, this invention also introduces a deep supervision mechanism. An auxiliary convolutional head is added to the feature outputs of the second (Dec2) and third (Dec3) stages of the striped convolutional decoder to generate an auxiliary prediction map, which is then scaled to the original image size. The deep supervision auxiliary loss is then calculated based on the auxiliary prediction map.
[0090] Therefore, the total loss of this invention is a weighted sum of the combined loss (soft skeleton loss + intersection-over-union loss + focal point loss + Dice loss + Hausdorff distance transform loss), deep supervision auxiliary loss, and auxiliary mask reconstruction loss (MPM Loss). This invention uses the Adam or AdamW optimizer for gradient descent and parameter updates.
[0091] In actual testing, the auxiliary mask prediction module (MPM module) and the branch containing the auxiliary convolution head are turned off. The preprocessed spatiotemporal image of the DAS vibration under test is propagated forward once through the trajectory extraction network. The Logits map output by the network is activated by the Sigmoid function and a confidence threshold (preferably 0.4 to 0.5) is set to obtain a high-fidelity, topologically connected binary vehicle trajectory mask map, thus completing the extraction of weak signals.
[0092] To verify the effectiveness of the proposed method, a real highway DAS traffic monitoring dataset was constructed for comparative experiments in this embodiment. This dataset was collected from a 10-kilometer fiber optic sensing section of a provincial highway, covering traffic scenarios under different time periods, traffic volumes, and weather conditions. The dataset contains 2590 labeled spatiotemporal vibration images, randomly divided into training and test sets at a 9:1 ratio, with 2331 images in the training set and 259 images in the test set. All images underwent a standardized preprocessing procedure, and trajectory annotation was performed by three traffic professionals. The annotation results were cross-validated to ensure accuracy.
[0093] This invention selects mainstream models in the current semantic segmentation field as comparison baselines, including the classic U-Net, DeepLabV3+ based on dilated convolution, and TransUNet combined with Transformer. Experiments use the following four evaluation metrics to comprehensively assess model performance:
[0094] 1. IoU (Intersection over Union): Measures the pixel-level overlap between the predicted trajectory and the true trajectory;
[0095] 2. Relaxed IoU: Allows the predicted trajectory to deviate from the true trajectory by up to 3 pixels, which better meets the practical application requirements of DAS trajectory extraction;
[0096] 3. Conn (connectivity index): Specifically measures the topological connectivity of a trajectory, calculating the proportion of complete and continuous trajectory segments to the total trajectory length;
[0097] 4. Qual (Quality Index): A comprehensive evaluation index that takes into account the integrity, accuracy, and noise level of the trajectory.
[0098] The quantitative performance comparison results of each model on the test set are shown in Table 1.
[0099] Table 1 Performance comparison between the present invention and existing mainstream models
[0100]
[0101] As shown in Table 1, this invention significantly outperforms other comparative models across all evaluation metrics. Particularly noteworthy are the Relaxed IoU and Conn metrics, crucial for DAS trajectory extraction, where this invention achieves 94.44% and 32.88% respectively, representing improvements of 6.44 and 7.28 percentage points compared to the second-best performing TransUNet. This demonstrates that the proposed combination of a learnable denoising front-end, a hybrid feature bottleneck layer, a striped convolutional decoder, and a soft skeleton loss function effectively addresses the issues of low signal-to-noise ratio and easily fragmented trajectories in DAS data.
[0102] To more intuitively demonstrate the advantages of the present invention, the embodiments further selected a typical strong noise and weak signal scenario from the test set for qualitative comparison, and the results are as follows. Figure 3 As shown.
[0103] Figure 3 These are visual effect comparison images of the test set in this embodiment of the invention; where (a) is the original input; (b) is the prediction result of the traditional U-Net method; and (c) is the prediction result of this invention. Figure 3It can be clearly seen that the original input image (a) contains a large amount of background speckle noise and horizontal interference stripes, the vehicle trajectory signal is very weak, and some areas are almost completely submerged by noise. Although the traditional method U-Net (b) can detect some trajectories, it suffers from severe discontinuity and falsely detects a large number of noise points. In contrast, the method of this invention (c) not only successfully suppresses background noise but also extracts the continuous vehicle trajectory completely. It maintains good topological connectivity even in areas with extremely weak signals, fully demonstrating the superior ability of this invention to extract weak vehicle trajectories under strong noise interference.
[0104] Finally, it should be noted that the above examples of the present invention are merely illustrative and not intended to limit the implementation of the invention. Although the applicant has described the present invention in detail with reference to preferred embodiments, those skilled in the art can make other variations and modifications based on the above description. It is impossible to exhaustively list all possible implementations here. All obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A method for extracting weak vehicle trajectories using DAS based on spatiotemporal feature enhancement, characterized in that: Includes the following steps: 1) Acquire spatiotemporal images of highway vibrations from distributed fiber optic acoustic sensors with known vehicle trajectories, and perform preprocessing. The preprocessed spatiotemporal images of vibration were divided into training and testing sets. 2) Input the spatiotemporal vibration images in the training set into the pre-constructed trajectory extraction network to train the trajectory extraction network. The trajectory extraction network identifies and outputs the vehicle trajectory corresponding to the spatiotemporal vibration image, thus obtaining the trained trajectory extraction network. The trajectory extraction network includes a denoising module, an encoder, a decoder, and a multi-scale Hough fusion output module connected in sequence. The denoising module uses a residual space attention mechanism to directionally enhance weak vehicle trajectory signals and suppress background speckle noise. 3) After performing the same preprocessing as in step 1), the vibration spatiotemporal image acquired by the actual distributed fiber acoustic sensor to be tested is input into the trained trajectory extraction network. The trajectory extraction network then outputs a vehicle binary trajectory segmentation mask with topological continuity.
2. The method for extracting weak vehicle trajectories based on spatiotemporal feature enhancement according to claim 1, characterized in that: The specific implementation process of preprocessing in step 1) is as follows: Background subtraction: The vibration spatiotemporal image is converted into a floating-point two-dimensional matrix, and the median vector of each row of the two-dimensional matrix is calculated; the original two-dimensional matrix is subtracted from the median vector and multiplied by a preset coefficient to remove horizontal stripe interference caused by fiber Rayleigh scattering or laser frequency drift. Intelligent black level clamping: Calculates the overall brightness distribution percentile of the image matrix after stripe removal; Extract the pixel value with brightness at a certain percentile as the dynamic noise baseline; Pixel values in the image matrix that are smaller than the noise level are set to 0 to remove dark background speckle noise. Adaptive enhancement: The processed matrix is Min-Max normalized to the range of 0-255, and contrast-limited adaptive histogram equalization is applied to obtain the final preprocessed input image. .
3. The method for extracting weak vehicle trajectories based on spatiotemporal feature enhancement according to claim 1, characterized in that: The noise reduction module includes: Feature pooling unit: The input DAS vibration spatiotemporal image features are subjected to average pooling and max pooling along the channel dimension respectively; average pooling is used to obtain the background color trend information of a large area in the image, and max pooling is used to capture the significant texture features of weak vehicle trajectories that present a diagonal linear distribution. Mask generation unit: After concatenating the acquired background color trend information and the faint vehicle trajectory significant texture features in the channel dimension, a spatial attention mask with a value range of 0 to 1 is generated through a convolutional layer and a Sigmoid activation function. This spatial attention mask represents the spatial distribution probability of the vehicle trajectory in the spatiotemporal image. The residual enhancement unit multiplies the original input DAS vibration spatiotemporal image features with the spatial attention mask element-wise to obtain a primary feature map that suppresses background noise. Then, the primary feature map is added to the original input DAS vibration spatiotemporal image features as a residual, thereby achieving directional numerical multiplication of the weak vehicle trajectory region with high confidence while preserving the original signal background, and outputting the denoised and enhanced feature map to the subsequent encoder.
4. The method for extracting weak trajectories of DAS vehicles based on spatiotemporal feature enhancement according to claim 1, characterized in that: A hybrid feature bottleneck layer is provided between the encoder and the decoder. The hybrid feature bottleneck layer includes a parallel CNN feature extraction branch and a Transformer self-attention branch. Both branches are used to receive deep features from the encoder's output. CNN Feature Extraction Branch: It contains multiple parallel multi-scale dilated convolution branches, which are combined with square, horizontal and vertical receptive fields to match and extract local vehicle trajectory texture and geometric orientation features with different tilt angles and widths. Transformer self-attention branch: preserves the spatial location prior of the spatiotemporal image through two-dimensional absolute position encoding, and uses a multi-head self-attention mechanism to capture long-distance, global trajectory coherence dependencies across the entire DAS vibration spatiotemporal image; Finally, the outputs of the two branches are concatenated along the channel dimension, and local trajectory details and global coherence information are fused through a convolutional layer before being output to the decoder.
5. The method for extracting weak vehicle trajectories based on spatiotemporal feature enhancement according to claim 1, characterized in that: The encoder employs a feature extraction network based on dilated convolution to maintain feature resolution. The encoder maintains regular convolution and downsampling in the shallow network to extract basic features. In the deep network, regular convolution is replaced with dilated convolution, and the downsampling stride is modified to 1. The feature map size reduction of the deep network is canceled. Specifically, a dilation rate of 2 is set in the penultimate stage of the deep network and a dilation rate of 4 is set in the final stage. This expands the receptive field without increasing the number of additional parameters, ensuring that the resolution of the deep feature map is always maintained at the preset ratio. This guarantees that the extremely fine vehicle trajectory features are not over-compressed and disappear in the deep network.
6. The method for extracting weak trajectories of DAS vehicles based on spatiotemporal feature enhancement according to claim 1, characterized in that: The decoder employs a 4-stage cascaded striped convolutional decoder; each stage of the decoder receives the upsampled features from the previous stage and the skip connection features corresponding to the encoder. The fused features are fed into parallel horizontal and vertical branches: Horizontal branches: containing kernel size of Asymmetric convolution; Vertical branches: contain kernel size of Asymmetric convolution; The outputs of the two branches are concatenated and batch normalized before entering the next level.
7. The method for extracting weak vehicle trajectories based on spatiotemporal feature enhancement according to claim 6, characterized in that: The multi-scale Hough fusion output module collects multi-scale features output from the 4-level decoder; each scale feature is equipped with a prior Hough block, and the prior Hough block is average pooled along the height and width dimensions to obtain a one-dimensional vector. Attention weights are calculated through one-dimensional convolution and broadcast and added. Finally, a weight map is formed by Sigmoid activation; all weighted features are bilinearly upsampled to the highest resolution size, stitched together, and finally fused using a convolutional layer. Then, a 1×1 convolution is used to output a classification logarithmic map.
8. The method for extracting weak trajectories of DAS vehicles based on spatiotemporal feature enhancement according to claim 1, characterized in that: During the training phase, a combined loss function including soft skeleton loss is used to optimize the parameters of the trajectory extraction network. The soft skeleton loss is used to constrain the network at the topology level, penalizing the breakage of vehicle trajectories. The calculation process for the soft skeleton loss includes: Morphological erosion and dilation operations are approximated by differentiable minima pooling and maxma pooling operations, respectively. By approximate erosion and dilation operations, the predicted soft skeleton and the real label skeleton are iteratively extracted from the trajectory probability map and the real label map predicted by the trajectory extraction network, respectively. The degree of overlap between the predicted trajectory probability map and the real label skeleton is calculated as the topological accuracy, and the degree of overlap between the predicted soft skeleton and the real label map is calculated as the topological sensitivity. The centerline intersection-union ratio is obtained by taking the harmonic mean of the topology accuracy and topology sensitivity. , This is known as soft skeleton loss.
9. The method for extracting weak trajectories of DAS vehicles based on spatiotemporal feature enhancement according to claim 6, characterized in that: During training, auxiliary convolution heads are added to the feature outputs of the second and third levels of the decoder to generate auxiliary prediction maps. The auxiliary prediction maps are then scaled to the original image size, and the auxiliary loss is calculated based on the auxiliary prediction maps as part of the total loss.
10. The method for extracting weak vehicle trajectories based on spatiotemporal feature enhancement according to claim 5, characterized in that: It also includes an auxiliary mask prediction module for training. The Layer 3 output of the encoder is connected to the auxiliary mask prediction module, which first generates a random binary mask with a preset coverage. ; Combine the Layer 3 output features of the encoder with After multiplication to achieve random occlusion, the original features are reconstructed through a predictor; During training, the auxiliary mask reconstruction loss is calculated based on the reconstructed original features to enable the encoder to learn the contextual geometric continuity of the vibration spatiotemporal image, thereby improving the self-repair capability of the fracture trajectory. The auxiliary mask reconstruction loss is part of the total loss.