Image Detection and Localization Method for Smoke Forefront in Highway Tunnel Fires

Abstract

This invention discloses a method for detecting and locating the smoke front in a highway tunnel fire, comprising: S1. performing data augmentation processing on the acquired fire smoke images to obtain an image dataset; S2. inputting the image dataset into a fire smoke detection network model for network model training to obtain a trained detection network model; S3. inputting the fire smoke image to be detected into the trained detection network model and outputting the fire smoke target area; S4. determining the position of the smoke front in the fire smoke target area; S5. determining the longitudinal distance between the smoke front position and the monitoring position. This invention can obtain richer subsequent disaster information and has high fire detection accuracy and timeliness.

Description

Technical Field

[0001] This invention relates to the field of highway tunnel fire detection, specifically to a method for detecting and locating smoke front images in highway tunnel fires. Background Technology

[0002] With the rapid increase in the number and mileage of operating tunnels in my country, the risk of fire in tunnels is becoming increasingly prominent. In the event of a fire, delayed traffic control instructions inside and outside the tunnel can cause traffic congestion, which can not only damage the tunnel structure and ancillary facilities, but also endanger the lives and property of drivers and passengers.

[0003] Existing highway tunnel fire detection systems or methods are severely limited in terms of interference resistance, accuracy, and real-time performance due to the influence of numerous factors such as airflow, wind speed, and cross-sectional height within the tunnel. They can only detect initial fire points and lack the ability to perceive and acquire information about the development of the fire within the tunnel and related information. Therefore, a new image detection and localization method for smoke fronts in highway tunnel fires is needed to solve the above problems. Summary of the Invention

[0004] In view of this, the purpose of this invention is to overcome the defects in the prior art and provide a method for detecting and locating smoke front images of highway tunnel fires, which can obtain richer information on subsequent disasters and has higher fire detection accuracy and timeliness.

[0005] The method for detecting and locating smoke front images in highway tunnel fires of the present invention includes the following steps:

[0006] S1. Perform data augmentation processing on the collected fire smoke images to obtain an image dataset;

[0007] S2. Input the image dataset into the fire smoke detection network model to train the network model and obtain the trained detection network model;

[0008] S3. Input the fire smoke image to be tested into the trained detection network model and output the fire smoke target area;

[0009] S4. Determine the location of the smoke front in the target area of ​​the fire smoke;

[0010] S5. Determine the longitudinal distance between the smoke front location and the monitoring location.

[0011] Furthermore, data augmentation processing is performed on the collected fire smoke images to obtain an image dataset, specifically including:

[0012] S11. Use rectangles to mark the locations of smoke areas in the fire smoke image to obtain the marked data image;

[0013] S12. Randomly select several labeled data images to form a subset of data images;

[0014] S13. Perform random cropping and scaling on the data images in the data image subset to obtain the processed data image subset;

[0015] S14. Randomly arrange the data images in the processed data image subset and stitch them together to form a data image;

[0016] S15. Repeat steps S12-S14 to obtain several data images.

[0017] Furthermore, YOLOv5s is used as the base network to construct a fire smoke detection network model. Specifically, a Transformer Encoder module is introduced into the backbone network; a lightweight convolutional method GSConv module is used in the neck network; a lightweight and efficient channel attention ECA module is added to the multi-scale feature branch; and the CIoU loss function and SiLU activation function are used.

[0018] Furthermore, a Transformer Encoder module is introduced into the backbone network, specifically including:

[0019] In the last layer of the backbone network, a Transformer Encoder is introduced and integrated with the C3 module. The input of the Transformer Encoder module comes from the feature map obtained by the previous Conv convolution operation, which is divided into Patches with positional information by the Embedded Patches layer, then linearly transformed and passed to the Multi-HeadAttention and MLP layers. Finally, the feature map with rich contextual information is passed to the next layer of the network.

[0020] Furthermore, the lightweight convolutional method GSConv module is used in the neck network, specifically including:

[0021] The lightweight convolutional method GSConv combines DSC with SC, using shuffle to permeate the information generated by SC into each part of the information generated by DSC.

[0022] Furthermore, a lightweight and efficient channel attention ECA module is added to the multi-scale feature branch, specifically including:

[0023] An efficient channel attention (ECA) module is added to the multi-scale feature branch of the Neck network. The ECA module first takes a feature map as input, with dimensions H1*W1*C1, where H1 is the feature map height, W1 is the feature map width, and C1 is the number of channels. Then, global average pooling is used to compress the spatial features of the feature map, resulting in a 1*1*C1 feature map. Next, a 1*1 convolution is used to learn the channel features of the compressed feature map, studying the importance of different channels. The output feature map has dimensions of 1*1*C1, thus incorporating the channel attention features. Figure 1 *1*C1, the original input feature map H1*W1*C1, is multiplied channel by channel to finally output a feature map with channel attention.

[0024] Furthermore, the CIoU loss function is determined according to the following formula.

[0025]

[0026] in, J represents the predicted bounding box area, K represents the ground truth bounding box area, ρ(·) represents the Euclidean distance function between the two center points, and b represents the center point of the predicted bounding box. gt denoted by , c represents the diagonal distance of the smallest enclosing region that can simultaneously contain two boxes, θ is a weighting coefficient, and ζ is a parameter that measures the consistency of aspect ratio.

[0027]

[0028] w gt Indicates the width of the actual bounding box, i gt represents the length of the ground truth bounding box, w represents the width of the predicted bounding box, and i represents the length of the predicted bounding box.

[0029] Furthermore, the SiLU activation function SiLU(x) is determined according to the following formula:

[0030] SiLU(x) = x·sigmoid(x);

[0031] Where x is the input value, and sigmoid(·) is a sigmoid function.

[0032] Furthermore, determining the smoke front position of the target area of ​​the fire smoke specifically includes: taking the point closest to the tunnel entrance in the target area as the target corner point, selecting a target point along the tunnel length direction starting from the target corner point, and taking the target point as the smoke front position; wherein, the distance between the target point and the target corner point is 1 / n of the width of the target area along the tunnel length direction.

[0033] Furthermore, the longitudinal distance d between the smoke front position and the monitoring position is determined according to the following formula:

[0034]

[0035] Among them, the location of the smoke front is taken as reference point B, and the location of the monitoring equipment is taken as the monitoring location;

[0036]

[0037] H is the height of the tunnel arch, h′ is the distance between the monitoring equipment and the ground, v is the longitudinal coordinate value of the reference point B, and v0 is the longitudinal coordinate value of the center point on the imaging plane of the monitoring equipment; f y The focal length in the vertical direction of the monitoring device is α, and the angle between the optical axis of the monitoring device and the horizontal plane is α.

[0038]

[0039] f is the focal length of the monitoring device, b is the pixel width of coordinate point F, h is the pixel height of coordinate point F, and coordinate point F is the coordinate point of reference point B on the imaging plane.

[0040] δ is the deflection angle, which is defined as arctan(l2 / l1), where l2 is the vertical distance from point D to the tunnel wall, l1 is the horizontal distance from point D to the bottom of the monitoring device, and point D is the intersection of the optical axis of the monitoring device and the tunnel surface.

[0041] The beneficial effects of this invention are as follows: The method for detecting and locating the smoke front in a highway tunnel fire disclosed in this invention acquires fire smoke images through traffic monitoring cameras inside the tunnel and identifies the fire smoke images. It can accurately and timely obtain the target area of ​​the fire smoke, and use computer vision positioning technology to dynamically perceive the spread of smoke across the entire tunnel section after the fire occurs. It also determines the longitudinal distance between the smoke front position and the monitoring position in the target area of ​​the fire smoke, enriching the disaster perception information and providing technical support for tunnel fire rescue and evacuation and fire emergency decision-making. Attached Figure Description

[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0043] Figure 1 This is a schematic diagram of the detection method of the present invention;

[0044] Figure 2 This is a schematic diagram of the YOLOv5 network structure of the present invention;

[0045] Figure 3 This is a structural diagram of the fire smoke detection network model based on YOLOv5s of the present invention;

[0046] Figure 4 This is an example of smoke image data from various types of tunnel fires according to the present invention;

[0047] Figure 5 This is a schematic diagram illustrating image annotation using LabelImg image annotation software according to the present invention;

[0048] Figure 6 The diagram shows the effect of the improved Mosaic data augmentation method of the present invention.

[0049] Figure 7 This is a structural diagram of the improved Transformer and C3 module of the present invention;

[0050] Figure 8 This is a structural diagram of the GSConv module introduced in this invention;

[0051] Figure 9 This is a structural diagram of the ECA module introduced in this invention;

[0052] Figure 10 This is a flowchart of the camera calibration process of the present invention;

[0053] Figure 11 This is an example diagram showing the position of the smoke front in this invention;

[0054] Figure 12 This is a schematic diagram illustrating the principle of smoke front position ranging based on monocular vision positioning technology of the present invention;

[0055] Figure 13 A diagram illustrating the fire smoke detection effect of this invention. Detailed Implementation

[0056] The present invention will be further described below with reference to the accompanying drawings, as shown in the figures:

[0057] The method for detecting and locating smoke front images in highway tunnel fires of the present invention includes the following steps:

[0058] S1. Perform data augmentation processing on the collected fire smoke images to obtain an image dataset;

[0059] S2. Input the image dataset into the fire smoke detection network model to train the network model and obtain the trained detection network model;

[0060] S3. Input the fire smoke image to be tested into the trained detection network model and output the fire smoke target area;

[0061] S4. Determine the location of the smoke front in the target area of ​​the fire smoke;

[0062] S5. Determine the longitudinal distance between the smoke front location and the monitoring location.

[0063] In this embodiment, in step S1, various types of fire smoke image data in a highway tunnel scenario are acquired, the smoke area locations in all data images are labeled, a labeled dataset is obtained, and the dataset is further expanded using data augmentation methods.

[0064] Data augmentation processing was performed on the collected fire smoke images to obtain an image dataset, which specifically includes:

[0065] S11. Use rectangles to mark the locations of smoke areas in the fire smoke image to obtain the marked data image;

[0066] S12. Randomly select several labeled data images to form a subset of data images;

[0067] S13. Perform random cropping and scaling on the data images in the data image subset to obtain the processed data image subset;

[0068] S14. Randomly arrange the data images in the processed data image subset and stitch them together to form a data image;

[0069] S15. Repeat steps S12-S14 to obtain several data images. Ordinary traffic monitoring cameras are used, with their video images as the target. Traffic monitoring cameras in tunnels are generally spaced 100-150m apart, achieving full coverage. Randomly combine 2, 4, 6, or up to 9 images, randomly crop and scale them, and then randomly arrange and stitch them together to form an image as input training data. Images containing up to nine different semantic information will include smaller smoke targets, which is beneficial for detecting small smoke targets in the early stages of a fire. Furthermore, the increased total number of bounding boxes in the batch sample also enhances the effect of the batch processing normalization layer.

[0070] In this embodiment, YOLOv5s is used as the base network to construct a fire smoke detection network model. Specifically, a Transformer Encoder module is introduced into the backbone network; a lightweight convolution method GSConv module is used in the neck network; a lightweight and efficient channel attention ECA module is added to the multi-scale feature branch; and the CIoU loss function and SiLU activation function are used.

[0071] The original YOLOv5 object detection network architecture is mainly divided into four parts: Input, Backbone, Neck, and Prediction. Figure 2 As shown;

[0072] The input layer employs the Mosaic method to randomly scale, crop, and arrange four images to enrich the sample dataset. Adaptive anchor box calculation obtains the optimal anchor box, and adaptive image scaling unifies the image size, reducing information redundancy. The Backbone and Neck layers form a feature extraction network, responsible for image feature extraction and fusion, respectively. This network mainly includes the Focus, CSP, Feature Pyramid (FPN), and Path Aggregation Network (PAN) structures. The Focus module performs image slicing, similar to downsampling, replacing three convolutional layers, reducing computation and accelerating the model. The CSP module uses a Cross-Stage Local (CSP) network, referencing the ResNet network, to merge feature maps across multiple convolutional layers, enhancing the information in deep feature maps, reducing computation while ensuring detection accuracy. The FPN and PAN structures achieve complementary fusion of high and low-level features, jointly enhancing the model's feature extraction and fusion capabilities. The Prediction layer consists of a loss function and Non-Maximum Suppression (NMS) operation. The loss function combines the traditional cross-entropy loss function with the GIoU loss function. In post-processing, non-maximum suppression is used to filter the output of the best target detection box with IoU as the threshold to complete the prediction of smoke targets at different scales.

[0073] In this embodiment, a Transformer Encoder module is introduced at the end of the backbone network and combined with the standard convolutional module of that layer.

[0074] The Multi-Head Attention module in the Transformer Encoder helps extract richer global and contextual information from the image. The Transformer Encoder is integrated with the C3 module in the last layer of the backbone network. The input to the Transformer Encoder module comes from the feature map obtained from the previous Conv convolution operation, which is divided into patches with location information by the Embedded Patches layer. After linear transformation, these patches are passed to the Multi-Head Attention and MLP layers, and finally, the feature map with rich contextual information is passed to the next layer of the network.

[0075] We utilize the state-of-the-art lightweight convolution method GSConv in the neck network.

[0076] To mitigate high computational costs and ensure timely detection, a lightweight design is considered. While using depthwise separable convolution (DSC) reduces the number of parameters and FLOPS to improve detection speed, DSC completely severs the channel information links of the input image, resulting in poorer feature extraction and fusion capabilities compared to standard convolution (SC). The lightweight convolution method GSConv combines DSC and SC, using a shuffle mechanism to permeate information generated by SC into each part of the information generated by DSC.

[0077] Add a lightweight and efficient channel attention ECA module to the multi-scale feature branch.

[0078] To improve false negatives, an attention module is added to the multi-scale feature branch of the Neck network. The ECA module first takes a feature map as input, with dimensions H1*W1*C1, where H1 is the feature map height, W1 is the feature map width, and C1 is the number of feature map channels. Then, global average pooling (GlobalAvgPool) is used to compress the feature map spatially, resulting in a 1*1*C1 feature map. Finally, a 1*1 convolution is used to learn the channel features of the compressed feature map, studying the importance of different channels. The output feature map has dimensions of 1*1*C1, representing the channel attention features. Figure 1 *1*C1, the original input feature map H1*W1*C1, is multiplied channel by channel to finally output a feature map with channel attention.

[0079] Use a combination of CIoU loss function and SiLU activation function.

[0080] The localization loss is calculated using an IoU-based loss function. Previously, the regression of predicted bounding boxes was mainly calculated using coordinate regression losses, such as L1 Loss and L2 Loss. However, the IoU loss treats the bounding box as a whole, which has scale invariance and makes the location regression of the predicted bounding box more accurate.

[0081] IoU represents the ratio of the intersection to the union of the predicted bounding box and the ground truth bounding box. The loss is defined as follows:

[0082]

[0083]

[0084] In the formula: J represents the area of ​​the predicted bounding box, and K represents the area of ​​the ground truth bounding box. According to the formula, there exists an IoU (Intersection over Union) that cannot be optimized because the predicted bounding box and the ground truth bounding box do not intersect; the IoU value is 0. A value of 1 will cause gradient vanishing, failing to reflect the distance between the two boxes; furthermore, even when the IoU is equal, it cannot reflect the degree of overlap between the two boxes, making the effect of predicted box regression uncertain. To address this, a penalty term is added to the minimum bounding rectangle of the predicted and ground truth boxes to obtain the proportion of the predicted and ground truth boxes within the closure region. The definition is as shown in the equation:

[0085]

[0086] And again The penalty term was modified to regress the Euclidean distance between the center points of the two frames and then proposed.

[0087] Although While it can directly minimize the distance between the center points of the two frames, accelerating convergence, it doesn't consider another important factor: aspect ratio. Therefore, a further improvement is proposed... function.

[0088]

[0089]

[0090]

[0091] In the above formula, G represents the bounding rectangle of the predicted bounding box and the ground truth bounding box; a,a gt ρ represents the center point of the two frames, and ρ represents the Euclidean distance between the two center points. 2 (a,a gt ) represents the square of the Euclidean distance between the center points of the two bounding boxes, c represents the diagonal distance of the smallest closure region that can simultaneously contain both boxes, ζ is a parameter that measures aspect ratio consistency, and θ is a weighting coefficient; i and w represent the length and width of the predicted box, respectively. gt ,w gt Let represent the length and width of the ground truth bounding box, respectively. CIoU loss takes into account the distance between the center points of the predicted box and the ground truth bounding box, the overlap, and the aspect ratio, making the regression of candidate boxes more stable and making up for many shortcomings of early IoU loss methods.

[0092] The SiLU activation function possesses the characteristics of being unbounded at the upper limit but bounded at the lower limit, smooth, and non-monotonic. SiLU outperforms ReLU in deep models. The SiLU activation function formula is defined as follows:

[0093] SiLU(x) = x·sigmoid(x)

[0094] Where, the sigmoid(x) function is equal to x is the input value.

[0095] In addition, the total loss of YOLOv5 consists of classification loss cls_loss, confidence loss obj_loss, and localization loss box_loss. The classification loss cls_loss, confidence loss obj_loss, and the binary cross-entropy loss function BCEWithLogitsLoss are calculated.

[0096] Precision, Recall, and mAP@0.5 were selected as metrics to measure model accuracy; FPS was selected as a metric to measure model timeliness.

[0097]

[0098]

[0099]

[0100]

[0101] In the formula, TP (True Positives) is the number of positive samples in the correct predictions, FP (False Positives) is the number of positive samples in the incorrect predictions, and FN (False Negatives) is the number of negative samples in the incorrect predictions.

[0102] AP@0.5 means that when the IoU threshold of the confusion matrix is ​​0.5, for a certain class of samples with n positive examples, all detection results are sorted in descending order of confidence, and each additional positive example corresponds to a Precision value (P). i For n P i The average value is the AP@0.5 for that class, and mAP@0.5 is the average of AP@0.5 for all classes.

[0103] In this embodiment, a ranging model is established based on computer monocular vision positioning technology to output the longitudinal distance between the smoke front area and the current monitoring camera, thereby enabling the monitoring of the subsequent smoke spread range and enriching the post-disaster perception information of tunnel fires.

[0104] By calibrating a monocular camera, the camera's intrinsic parameters, rotation matrix, translation matrix, tangential distortion, and radial distortion are obtained. By correcting the monocular camera, the original image is input, and the calibrated intrinsic and extrinsic parameters are obtained to perform distortion elimination, thereby reducing the ranging error caused by the selection of reference points.

[0105] This invention uses the Zhang Zhengyou calibration method to obtain camera intrinsic parameters. By taking no fewer than three images of chessboard corner points in different poses, the homography relationship between the imaging plane and the chessboard plane is calculated using their world coordinate information, and the camera parameters are obtained through calibration.

[0106] Based on the deployment of surveillance cameras in tunnel scenarios, and using computer monocular vision positioning technology and a monocular ranging model based on geometric relationships, this invention establishes a ranging model. It analyzes the geometric relationship between the camera and the target during the imaging process, and then calculates the distance using parameters obtained from camera calibration. Figure 10 As shown.

[0107] The fire smoke detection network model can detect smoke areas in videos or images and return the pixel coordinates of the smoke detection box area. Combining the smoke diffusion law of tunnel fires (analysis of smoke diffusion law shows that when smoke spreads in the tunnel after a fire, the upstream and downstream front smoke can be regarded as spreading closely along the tunnel arch towards the two side openings), the smoke front position of the fire smoke target area is determined. Specifically, it includes: taking the point closest to the tunnel opening in the target area as the target corner point, taking the target corner point as the starting point, selecting target points along the tunnel length direction, and taking the target points as the smoke front position; wherein, the distance between the target points and the target corner points is 1 / 4 of the width of the target area along the tunnel length direction.

[0108] like Figure 11 As shown, the point (x) located 1 / 4 of the distance from the top left corner of the detection box is selected. c ,y c The position of the top left corner of the detection box is approximately taken as the leading edge of the smoke at the top of the dome. min ,y min The coordinates of the lower right corner of the detection box are (x max ,y max Initially, it is assumed that the smoke front is on the same horizontal plane, and the reference point is at the same height as the tunnel arch. However, since the detection frame may fluctuate, the coordinate information of the reference point can be corrected later using actual detection images.

[0109] The monitoring equipment uses surveillance cameras or cameras, such as... Figure 12 As shown, the surveillance camera is positioned at point O on the tunnel sidewall. O′ is the center point of the imaging plane. Point A is the perpendicular point of point B on the optical axis plane. Point B is the reference point (the position of the smoke front). Point C is the perpendicular point of point B on the tunnel wall to one side of the camera. Point D is the intersection of the optical axis and the tunnel surface. Point E is the coordinate point of point A on the imaging plane. Point F is the coordinate point of reference point B on the imaging plane. The length d of O″C represents the longitudinal distance between the reference point and the surveillance camera, and the length l of BC is the lateral distance.

[0110] Given that the tunnel arch height is H and the camera distance from the ground is h′, the intrinsic parameter f can be obtained through camera calibration. y f x Let v0, u0, and O′O have a focal length f. The pitch angle α = arctan(h′ / l3), where l3 is the distance from the optical axis intersection point D on the ground to directly below the camera. The deflection angle δ = arctan(l2 / l1), where l2 is the vertical distance from the optical axis intersection point D on the ground to the tunnel wall, and l1 is the horizontal distance from the optical axis intersection point D on the ground to directly below the camera. In the case of the optical axis plane, let O″A = d′ and (x...) c ,y c The pixel height of pixel F is h = v - v0, the width is b, and the vertical distance d is calculated as follows:

[0111] a. In ΔO″OA, In ΔO′EF

[0112] b. ΔOEF is similar to ΔOBA, therefore, let AB = l′

[0113]

[0114] c. In ΔO″AB, ∠O″AB = 90°. In ΔO″CB, ∠O″CB = 90°, ∠CO″B = δ - ε.

[0115]

[0116]

[0117]

[0118] Substituting the obtained parameters, we get the longitudinal distance d:

[0119]

[0120] in,

[0121] It should be noted that the above model formula considers the position of the smoke front reference point to be on the right side of the optical axis plane. When it is on the left side of the optical axis plane, ∠CO″B=δ+ε; the calculation principle of the longitudinal distance d is the same as when ∠CO″B=δ-ε, and will not be repeated here.

[0122] This invention presents a method for detecting and locating smoke fronts in highway tunnel fires. It enriches tunnel fire smoke image data using an enhanced Mosaic method. At the end of the backbone network, a Transformer Encoder module is introduced, combined with the standard convolutional module of that layer. This leverages the advantages of both multi-head self-attention and convolutional modules in feature extraction, enhancing the network's global feature extraction capabilities and improving the difficulty in extracting features from smaller smoke targets, thus improving network performance. Simultaneously, in the neck network, the latest lightweight convolutional method GSConv replaces some convolutional modules (Conv), effectively reducing the number of network parameters without sacrificing performance, achieving network compression. Furthermore, a lightweight and efficient channel attention (ECA) module is added to the multi-scale feature branch, mitigating missed detections of distant targets and small smoke targets by employing a local cross-channel interaction strategy, improving network performance with a small increase in parameters. Additionally, a combination of CIoU loss and SiLU activation functions allows the network model to converge faster and better. Based on computer monocular vision localization technology, a ranging model is established to output the longitudinal distance between the smoke front region and the current monitoring camera, enabling monitoring of the subsequent smoke spread range and enriching post-disaster perception information for tunnel fires. Actual effect as Figure 13 As shown.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. 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 be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting and locating the smoke front in a highway tunnel fire, characterized in that: Includes the following steps: S1. Perform data augmentation processing on the collected fire smoke images to obtain an image dataset; S2. Input the image dataset into the fire smoke detection network model to train the network model and obtain the trained detection network model; S3. Input the fire smoke image to be tested into the trained detection network model and output the fire smoke target area; S4. Determine the smoke front position of the target area of ​​the fire smoke, specifically including: taking the point closest to the tunnel entrance in the target area as the target corner point, taking the target corner point as the starting point, selecting a target point along the tunnel length direction, and taking the target point as the smoke front position; wherein, the distance between the target point and the target corner point is 1 / n of the width of the target area along the tunnel length direction; S5. Determine the longitudinal distance between the smoke front location and the monitoring location; The longitudinal distance d between the smoke front position and the monitoring position is determined using the following formula: Among them, the location of the smoke front is taken as reference point B, and the location of the monitoring equipment is taken as the monitoring location; H is the height of the tunnel arch, h′ is the distance between the monitoring equipment and the ground, v is the longitudinal coordinate value of the reference point B, and v0 is the longitudinal coordinate value of the center point on the imaging plane of the monitoring equipment; f y The focal length in the vertical direction of the monitoring device is α, and the angle between the optical axis of the monitoring device and the horizontal plane is α. f is the focal length of the monitoring device, b is the pixel width of coordinate point F, h is the pixel height of coordinate point F, and coordinate point F is the coordinate point of reference point B on the imaging plane. δ is the deflection angle, which is defined as arctan(l2 / l1), where l2 is the vertical distance from point D to the tunnel wall, l1 is the horizontal distance from point D to the bottom of the monitoring device, and point D is the intersection of the optical axis of the monitoring device and the tunnel surface.

2. The method for detecting and locating the smoke front image of a highway tunnel fire according to claim 1, characterized in that: Data augmentation processing was performed on the collected fire smoke images to obtain an image dataset, which specifically includes: S11. Use rectangles to mark the locations of smoke areas in the fire smoke image to obtain the marked data image; S12. Randomly select several labeled data images to form a subset of data images; S13. Perform random cropping and scaling on the data images in the data image subset to obtain the processed data image subset; S14. Randomly arrange the data images in the processed data image subset and stitch them together to form a data image; S15. Repeat steps S12-S14 to obtain several data images.

3. The method for detecting and locating the smoke front image of a highway tunnel fire according to claim 1, characterized in that: A fire smoke detection network model was constructed using YOLOv5s as the base network. The TransformerEncoder module was introduced into the backbone network; the lightweight convolution method GSConv module was used in the neck network; a lightweight and efficient channel attention ECA module was added to the multi-scale feature branch; and the CIoU loss function and SiLU activation function were used.

4. The method for detecting and locating the smoke front image of a highway tunnel fire according to claim 3, characterized in that: The Transformer Encoder module is introduced into the backbone network, specifically including: In the last layer of the backbone network, a Transformer Encoder is introduced and integrated with the C3 module. The input of the Transformer Encoder module comes from the feature map obtained by the previous Conv convolution operation, which is divided into Patches with positional information by the Embedded Patches layer, then linearly transformed and passed to the Multi-HeadAttention and MLP layers. Finally, the feature map with rich contextual information is passed to the next layer of the network.

5. The method for detecting and locating smoke front images in highway tunnel fires according to claim 3, characterized in that: The GSConv module, a lightweight convolutional method, is used in the neck network, specifically including: The lightweight convolutional method GSConv combines DSC with SC, using shuffle to permeate the information generated by SC into each part of the information generated by DSC.

6. The method for detecting and locating smoke front images in highway tunnel fires according to claim 3, characterized in that: A lightweight and efficient channel attention ECA module is added to the multi-scale feature branch, specifically including: An efficient channel attention (ECA) module is added to the multi-scale feature branch of the Neck network. The ECA module first takes a feature map as input, with dimensions H1*W1*C1, where H1 is the feature map height, W1 is the feature map width, and C1 is the number of channels. Then, global average pooling is used to compress the spatial features of the feature map, resulting in a 1*1*C1 feature map. Next, a 1*1 convolution is used to learn the channel features of the compressed feature map, studying the importance of different channels, and outputting a 1*1*C1 feature map. Finally, the channel-attention feature map 1*1*C1 and the original input feature map H1*W1*C1 are multiplied channel-wise to output a feature map with channel attention.

7. The method for detecting and locating the smoke front image of a highway tunnel fire according to claim 3, characterized in that: The CIoU loss function is determined according to the following formula. in, J represents the predicted bounding box area, K represents the ground truth bounding box area, ρ(·) represents the Euclidean distance function between the two center points, and a represents the center point of the predicted bounding box. gt denoted by , c represents the diagonal distance of the smallest enclosing region that can simultaneously contain two boxes, θ is a weighting coefficient, and ζ is a parameter that measures the consistency of aspect ratio. w gt Indicates the width of the actual bounding box, i gt represents the length of the ground truth bounding box, w represents the width of the predicted bounding box, and i represents the length of the predicted bounding box.

8. The method for detecting and locating smoke front images in highway tunnel fires according to claim 3, characterized in that: The SiLU activation function SiLU(x) is determined according to the following formula: SiLU(x) = xsigmoid(x); Where x is the input value, and sigmoid(·) is a sigmoid function.

Images