Tunnel anomaly detection model establishment and detection and early warning method

Abstract

The present invention relates to the field of tunnel anomaly detection, and particularly to the establishment of a tunnel anomaly detection model and a detection and early warning method. An ARDMADet model is established, which is an improved YOLOv10 model. The ARDMADet model includes a backbone network for extracting features from an input image, a neck network for extracting and fusing features from a feature map, and a head structure for detecting and classifying the fused feature map output by the neck network. The C2f module in the backbone network of the YOLOv10 model is replaced with a SADMA_AR module, a first SADMA module is added to the Concat module and the convolution module of the C2f module, both convolution modules in the C2f module are replaced with a first adaptive rectangular convolution module, and the Bottleneck module in the C2f module is replaced with a Mablock module. The adaptive rectangular convolution module can not only change the shape of the convolution kernel, but also dynamically adjust the number of sampling points according to the learned height and width, thereby solving the problem that the number of sampling points in previous convolutions is fixed and it is difficult to adapt to multi-scale targets.

Description

Technical Field

[0001] The present invention relates to the field of tunnel anomaly detection, and in particular to the establishment of a tunnel anomaly detection model and a detection and early warning method. Background Art

[0002] Personnel detection and alarming in tunnel environments is a critical safety technology, primarily used in rail transit, mine tunnels, and underground engineering. Traditional monitoring methods primarily include video surveillance systems, infrared sensors, and ultrasonic detection equipment. Video surveillance systems struggle to provide clear images in low-light or dusty environments and are easily affected by lighting fluctuations, resulting in reduced detection accuracy. Infrared sensors may not accurately detect personnel in high-temperature environments or when there is interference from heat sources, while ultrasonic sensors may misjudge due to tunnel echoes.

[0003] In addition, as a high-precision detection device, three-dimensional laser radar (LiDAR) can provide complete spatial information, but the traditional three-dimensional point cloud data processing requires large amount of computation and high hardware resources, making it difficult to run efficiently on edge computing devices. Summary of the Invention

[0004] The purpose of the present invention is to provide a tunnel anomaly detection model establishment and detection and early warning method that reduces computational complexity.

[0005] In order to achieve the above object, the present invention adopts such technical solution:

[0006] A tunnel anomaly detection model is established by establishing an ARDMADet model. The ARDMADet model is an improved YOLOv10 model. The ARDMADet model includes a backbone network for extracting features from input images, a neck network for extracting and fusing features from feature maps, and a head structure for detecting and classifying the fused feature maps output by the neck network.

[0007] Replace the C2f module in the backbone network of the YOLOv10 model with the feature extraction unit SADMA_AR module, which is improved based on the C2f module;

[0008] The feature extraction unit SADMA_AR module adds a first SADMA module to the Concat module and the convolution module of the C2f module, replaces both convolution modules in the C2f module with a first adaptive rectangular convolution module, and sequentially processes the feature map output by the first adaptive rectangular convolution module through normalization and a first activation function, and replaces the Bottleneck module in the C2f module with a Mablock module;

[0009] The Mablock module includes two second adaptive rectangular convolution modules, a second SADMA module and a connection module connected in sequence. The feature map output by the second adaptive rectangular convolution module is sequentially normalized and processed by the first activation function. The connection module is used to connect the feature map input to the Mablock module with the feature map output by the second SADMA module.

[0010] Preferably, the first SADMA module and the second SADMA module each include four branches:

[0011] The first branch performs a first rotation on the initial feature map to obtain a first rotated feature map;

[0012] The second branch performs maximum pooling and average pooling operations on the first rotated feature map, uses deformable convolution modules of different dimensions to extract features from the pooled feature maps respectively, and splices the feature maps of different dimensions and outputs them using the second activation function;

[0013] Performing an element-by-element multiplication operation on the first rotated feature map and the feature map output by the second branch, performing a first reverse rotation, and then outputting the result;

[0014] The third branch performs a second rotation on the initial feature map to obtain a second rotated feature map, performs maximum pooling and average pooling operations on the second rotated feature map, uses deformable convolution modules of different dimensions to extract features from the pooled feature maps, splices the feature maps of different dimensions and outputs them using a second activation function, and performs a second reverse rotation on the feature map output by the second activation function before outputting it;

[0015] The fourth branch performs maximum pooling and average pooling operations on the initial feature map, uses deformable convolution modules of different dimensions to extract features from the pooled feature maps, concatenates the feature maps of different dimensions, and outputs them using the second activation function. The feature map output by the second activation function is element-wise multiplied with the initial feature map and then output.

[0016] The feature map output after the first reverse rotation, the feature map after the second reverse rotation, and the feature map output by the fourth branch are added together to obtain the average value and then output.

[0017] Preferably, the first rotation is a 90° counterclockwise rotation along the H axis, the first reverse rotation is a 90° clockwise rotation along the H axis, the second rotation is a 90° counterclockwise rotation along the W axis, and the second reverse rotation is a 90° clockwise rotation along the W axis.

[0018] Preferably, the second activation function is a Sigmoid activation function.

[0019] Preferably, the first activation function is a SILU activation function.

[0020] The tunnel anomaly detection and early warning method includes the following steps performed in sequence:

[0021] S1: Acquire 3D point cloud data in the tunnel in real time, project the 3D point cloud data to obtain a 2D data set, preprocess the 2D data set, and divide the preprocessed 2D data set into a training set, a validation set, and a test set according to a preset ratio;

[0022] S2: inputting the training set into the ARDMADet model established by any of the tunnel anomaly detection model establishment methods described above to train the ARDMADet model;

[0023] S3: Use the WGIoU loss function to reversely adjust the trained ARDMADet model to obtain the adjusted ARDMADet model. The calculation formula of the WGIoU loss function is as follows:

[0024] ;

[0025] Among them, A represents the real box, B represents the preset bounding box, and C represents the smallest box containing the real box and the predicted box. represents the coordinates of the center of the anchor box, Indicates the coordinates of the center of the target box, and Indicates the size of the minimum bounding box. ,in , is the outlier degree, is the IoU threshold between the predicted bounding box and the true bounding box, for The sliding average of , small outlier means high quality of anchor box, and is a hyperparameter, To utilize A non-monotonic focusing coefficient is constructed;

[0026] S4: Input the test set into the adjusted ARDMADet model for testing and obtain the test results;

[0027] S5: If the test result shows that there are people in the preset area, an alarm prompt will be issued.

[0028] Preferably, the preprocessing of step S1 includes performing background filtering, ROI region cropping and noise suppression on the two-dimensional data set in sequence.

[0029] A computer product includes a computer program, which, when executed by a processor, implements any of the above-mentioned tunnel anomaly detection and early warning methods.

[0030] A computer storage medium stores a computer program, which, when executed by a processor, implements any of the above-mentioned tunnel anomaly detection and early warning methods.

[0031] By adopting the above-mentioned design scheme, the beneficial effect of the present invention is: the ARDMADet model established in this application embeds the adaptive rectangular convolution module into the C2f module. The adaptive rectangular convolution module can dynamically learn the height and width of the convolution kernel to generate a rectangular convolution kernel, thereby flexibly adjusting the shape of the convolution window according to the size of different objects in the image. The adaptive rectangular convolution module can not only change the shape of the convolution kernel, but also dynamically adjust the number of sampling points according to the learned height and width, solving the problem that the number of sampling points in the previous convolution is fixed and difficult to adapt to multi-scale targets. The adaptive rectangular convolution module only needs to learn two parameters, height and width, and has better convergence in small data set tasks, so as to improve the adaptability of the ARDMADet model to targets of different scales and shapes. The finally constructed model ARDMADet significantly reduces the number of network parameters and computational complexity while improving the accuracy of small target detection.

[0032] By adopting the above-mentioned design scheme, the beneficial effects of the present invention are as follows: the present invention significantly improves the model's perception of target direction uncertainty and scale diversity by introducing the SADMA module; the SADMA module utilizes a mechanism that combines spatial rotation with multi-scale deformable convolution to achieve robust modeling of targets at multiple angles and sizes, effectively enhancing the model's small target detection performance in environments with uneven lighting and complex structures, such as tunnels; experimental results show that in the case of occlusion, distortion, or partial missing contours, the SADMA module can still effectively extract discriminative features to avoid false detection and missed detection;

[0033] In addition, the adaptive rectangular convolution module can automatically adjust the height and width ratio of the convolution kernel to address the structural characteristics of human-shaped targets in tunnel scenes, which are often "slender, irregular, and have unfixed orientation", and generate a rectangular receptive field with adaptive aspect ratio, thereby more accurately extracting the structural features of fine-grained areas in the BEV image; especially in the bird's-eye view generated by the lidar, since the target is prone to deformation or stretching during the projection process, the adaptive rectangular convolution module can effectively enhance the model's adaptability to human-shaped targets of different scales and postures, while controlling the computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Schematic diagram of the structure of the ARDMADet model of the present invention;

[0035] Figure 2 Schematic diagram of the structure of the SADMA module of the present invention;

[0036] Figure 3 Schematic diagram of the structure of the deformable convolution module of the present invention;

[0037] Figure 4 This is a schematic structural diagram of the Mablock module of the present invention;

[0038] Figure 5 Schematic diagram of the structure of the SADMA_AR module of the present invention. DETAILED DESCRIPTION

[0039] To make the objectives, technical solutions, and advantages of the present invention more apparent, the present invention will be further described in detail below with reference to the accompanying drawings. It is apparent that the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort are intended to fall within the scope of protection of the present invention.

[0040] The terms "first," "second," "third," and so on, in the description and claims of the present invention and the accompanying drawings are used to distinguish between different items, not to describe a specific order. Furthermore, the terms "including," "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements but may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to the process, method, product, or apparatus.

[0041] Tunnel anomaly detection model establishment method, such as Figure 1 The ARDMADet (Adaptive RotatedDeformable Multi-scale Attention Detector) model shown is an improved YOLOv10 model. The ARDMADet model includes a backbone network for extracting features from input images, a neck network for extracting and fusing features from feature maps, and a head structure for detecting and classifying the fused feature maps output by the neck network.

[0042] Replace the C2f module in the backbone network of the YOLOv10 model with Figure 5 The feature extraction unit SADMA_AR module shown is improved based on the C2f module;

[0043] The feature extraction unit SADMA_AR module adds a first SADMA (Small-Aware Deformable Multi-scale Attention) module to the Concat module and the convolution module of the C2f module, replaces both convolution modules in the C2f module with a first adaptive rectangular convolution (Adaptive Rotated Convolution, ARConv) module, and the feature map output by the first adaptive rectangular convolution module is sequentially normalized and processed with the first activation function. The Bottleneck module in the C2f module is replaced by Figure 4 The Mablock module shown in the figure. This application embeds the adaptive rectangular convolution module into the C2f structure, replacing the standard convolution module therein, to improve the model's adaptability to targets of different scales and shapes. The SADMA module is injected into the C2f structure in parallel, which can further enhance the model's perception of spatial areas. The constructed feature extraction unit SADMA_AR module has stronger expressive power.

[0044] The Mablock module includes two second adaptive rectangular convolution modules, a second SADMA module and a connection module connected in sequence. The feature map output by the second adaptive rectangular convolution module is sequentially normalized and processed by the first activation function. The connection module is used to connect the feature map input to the Mablock module with the feature map output by the second SADMA module.

[0045] The core idea of ​​the adaptive rectangular convolution module in this embodiment is to adjust the size of the convolution kernel in both height and width directions in a learnable way, so that the receptive field is adaptive to adapt to the structural characteristics of targets of different scales and aspect ratios.

[0046] like Figure 2 As shown, the first SADMA module and the second SADMA module each include four branches:

[0047] The first branch performs a first rotation on the initial feature map to obtain a first rotated feature map. In this embodiment, the first branch rotates the input C×H×W feature map 90° counterclockwise along the H axis to obtain a W×H×C first rotated feature map.

[0048] The second branch performs maximum pooling and average pooling operations on the first rotated feature map in the W dimension through Z-pool, and uses deformable convolution modules of different dimensions to extract features from the pooled feature maps respectively, and splices the feature maps of different dimensions and outputs them using the second activation function; in this embodiment, Figure 3As shown in FIG, the deformable convolution modules of different dimensions refer to three deformable convolution modules including convolution kernels of 1×1, 3×3, and 7×7, and the three deformable convolution modules of different dimensions constitute a DMC_Block module.

[0049] Among them, Z-pool focuses on the maximum pooling and average pooling of the input. It can be summarized as the following formula:

[0050] ;

[0051] represents the input tensor C × H × W, Indicates dimension is 0, represents the maximum pooling operation, Represents an average pooling operation.

[0052] The resulting input tensor after Z-pooling is 2 × H × W, where 2 represents the dimension. We then introduce three parallel variable convolutions with kernel sizes of 1, 3, and 7. This provides intermediate outputs of varying dimensions, which are then used to generate attention weights through a second activation function. The operation and final output can be summarized in the following formula:

[0053]

[0054]

[0055] Indicates the processing process of the DMC_Block module, Represents the dimension, 、 and denote two-dimensional convolution operations with kernels of 1, 3, and 7 and a stride of 1, express , represents the Sigmoid activation function, and represents the rotated tensor, Represents the output after activation function processing.

[0056] Performing an element-by-element multiplication operation on the first rotated feature map and the feature map output by the second branch, performing a first reverse rotation on the feature map, and then outputting the result. In this embodiment, the first reverse rotation is a 90° clockwise rotation along the H axis.

[0057] The third branch performs a second rotation on the initial feature map to obtain a second rotated feature map. In this embodiment, the third branch rotates the input C×H×W feature map 90° counterclockwise along the W axis to obtain an H × C × W second rotated feature map, and performs maximum pooling and average pooling operations on the second rotated feature map through Z-pool in the H dimension. Deformable convolution modules of different dimensions are used to extract features from the pooled feature maps respectively, and the feature maps of different dimensions are spliced ​​and output using a second activation function. The feature map output by the second activation function is subjected to a second reverse rotation and then output. In this embodiment, the second reverse rotation is a 90° clockwise rotation along the W axis.

[0058] The fourth branch performs maximum pooling and average pooling operations on the initial feature map, uses deformable convolution modules of different dimensions to extract features from the pooled feature maps, concatenates the feature maps of different dimensions, and outputs them using the second activation function. The feature map output by the second activation function is element-wise multiplied with the initial feature map and then output.

[0059] The feature map output after the first reverse rotation, the feature map after the second reverse rotation, and the feature map output by the fourth branch are added together to obtain the average value and then output.

[0060] The SADMA module constructs multiple attention directions by rotating the input feature map in spatial dimensions to simulate situations where the target may appear in various directions, thereby establishing stronger directional perception capabilities. In each direction branch, the SADMA module uses deformable convolution modules of different dimensions to extract local and global attention features in parallel to enhance the response capabilities to targets of different sizes, especially the recognition stability of small targets in complex backgrounds. Moreover, the use of deformable convolution modules enables each attention branch to have adaptive sampling capabilities to adapt to the possible deformation, non-rigid edges and other features of the targets in the tunnel, effectively improving the modeling effect of the model for irregular targets and occluded targets.

[0061] In this embodiment, the first activation function is the SILU activation function, and the second activation function is the Sigmoid activation function. In this application, the SILU activation function has a smooth and non-zero negative semi-axis output, which can provide more continuous gradient information while maintaining nonlinear characteristics, helping to alleviate the vanishing gradient problem and improve the stability of model training. Furthermore, the SILU function has a self-gating property, which can dynamically adjust the output amplitude based on the input, enhancing the network's ability to express features of different scales. For the detection tasks of small targets and low-contrast targets involved in this invention, the SILU helps retain more effective information, improving detection accuracy and model robustness. Therefore, this application replaces the activation function from ReLU to SILU, which not only optimizes the network training process but also improves the model's detection performance for abnormal tunnel targets.

[0062] This embodiment also provides a method for performing tunnel anomaly detection and early warning using the ARDMADet model established by the above-mentioned tunnel anomaly detection model establishment method.

[0063] The tunnel anomaly detection and early warning method includes the following steps performed in sequence:

[0064] S1: Acquire 3D point cloud data in the tunnel in real time. This 3D point cloud data is projected and transformed into a 2D dataset, i.e., the 3D point cloud is mapped to a 2D BEV (Bird's Eye View) plane to form an equivalent grayscale image or depth map representation. This 2D dataset is then preprocessed and divided into a training set, a validation set, and a test set according to a preset ratio.

[0065] In this embodiment, a conventional laser radar device is used to collect three-dimensional point cloud data. The laser radar device can be installed in a tunnel or on a rail car. The preprocessing of step S1 includes background filtering, ROI area cropping and noise suppression for the two-dimensional data set in sequence.

[0066] S2: Input the training set into the ARDMADet model established by any of the tunnel anomaly detection model establishment methods described above to train the ARDMADet model; this application uses Jetson Nano as the edge computing platform, and converts the trained ARDMADet model into a TensorRT engine after structural optimization and quantization compression and deploys it on the Jetson Nano platform. Jetson Nano has strong edge reasoning capabilities and can run target detection models at a high frame rate under low power conditions to meet real-time detection needs.

[0067] S3: Use the WGIoU loss function to reversely adjust the trained ARDMADet model to obtain the adjusted ARDMADet model. The calculation formula of the WGIoU loss function is as follows:

[0068] ;

[0069] Among them, A represents the real box, B represents the preset bounding box, and C represents the smallest box containing the real box and the predicted box. represents the coordinates of the center of the anchor box, Indicates the coordinates of the center of the target box, and Represents the size of the minimum bounding box. The entire loss function consists of three parts: generalized IoU term , used to measure the overall overlap between the predicted box and the true box, especially when the IoU is 0, the effective gradient is still maintained. Center point penalty , the offset between the center of the predicted box and the center of the true box is penalized in the form of Gaussian attenuation to improve the accuracy and stability of bounding box positioning.

[0070] The non-monotonic focusing coefficient r is defined as: ,in , It is defined as the outlier degree, which is used to describe the quality of the anchor box. is the IoU threshold between the predicted bounding box and the true bounding box, for The sliding average of can dynamically adjust the overall gain during training. A low outlier score indicates a high-quality anchor box, and assigning a small gradient gain to it allows bounding box regression to focus on anchor boxes of average quality. Assigning a smaller gradient gain to anchor boxes with a large outlier score effectively prevents low-quality image data from generating large, harmful gradients that can affect detection quality. and It is a hyperparameter and can be adjusted appropriately according to different models. Is to use A non-monotonic focusing coefficient is constructed to increase the penalty for high IoU samples and reduce the attention to low IoU samples, so that the network can focus more on high-quality candidate boxes.

[0071] In complex tunnel environments, the ARDMADet model is reversely adjusted using the WGIoU loss function, which can significantly improve the robustness and accuracy of detection. This is mainly due to its three core mechanisms: 1. The generalized IoU term effectively solves the inherent target sparsity, boundary ambiguity and occlusion problems of BEV projection (especially at the far end of the tunnel or on a curve), and can provide effective gradients even when the predicted box and the true box have zero overlap, significantly reducing missed detections; 2. The center point penalty term based on Gaussian attenuation strengthens the precise fine-tuning and noise suppression of target center positioning, and improves the accuracy and inter-frame stability of position estimation of targets such as people under the influence of point cloud noise; 3. The key non-monotonic focusing coefficient r (based on the outlier β and hyperparameters and Dynamically adjust training weights to intelligently suppress noise and outlier samples, such as "ghost targets" generated by tunnel wall reflections, false alarms from fixed structures, and low-quality, sparse targets. While simultaneously focusing on high-quality, learnable samples, this significantly reduces the false detection rate. These mechanisms work together to achieve more reliable target detection in complex tunnel scenarios (such as those with occlusion, structural interference, and sparse point clouds), providing a more accurate environmental perception foundation for safety warning systems.

[0072] S4: The test dataset is fed into the adjusted ARDMADet model for testing, and test results are obtained. These results, encapsulated in JSON format and including fields such as object category, confidence level, center coordinates, and bounding box dimensions, are then sent to an edge gateway on the local network via the lightweight MQTT protocol. This system achieves a complete closed-loop inference process without relying on an external server, offering low latency, low power consumption, and high adaptability, making it suitable for deployment in resource-constrained industrial sites.

[0073] S5: If the test result shows that there is a person in the preset area, an alarm prompt is issued; the alarm prompt here includes but is not limited to issuing an alarm sound, sending an alarm message to the management personnel, voice alarm, etc.

[0074] This application also provides a computer product that can implement the above-mentioned tunnel anomaly detection and early warning method.

[0075] A computer product includes a computer program, which, when executed by a processor, implements any of the above-mentioned tunnel anomaly detection and early warning methods.

[0076] In this application, a storage medium is also provided that stores the information that can implement the above-mentioned tunnel anomaly detection and early warning method.

[0077] A computer storage medium stores a computer program, which, when executed by a processor, implements any of the above-mentioned tunnel anomaly detection and early warning methods.

[0078] In summary, the present invention significantly improves the model's ability to perceive target directional uncertainty and scale diversity by introducing the SADMA module. The SADMA module utilizes a mechanism that combines spatial rotation with multi-scale deformable convolution to achieve robust modeling of targets at multiple angles and sizes, effectively enhancing the model's small target detection performance in environments with uneven lighting and complex structures, such as tunnels. Experimental results show that even in the presence of occlusion, distortion, or partial missing contours, the SADMA module can still effectively extract discriminative features, avoiding false detections and missed detections.

[0079] Furthermore, the adaptive rectangular convolution module addresses the often slender, irregular, and erratically oriented structural characteristics of humanoid targets in tunnel scenarios by automatically adjusting the height and width ratio of the convolution kernel to generate a rectangular receptive field with adaptive aspect ratios, thereby more accurately extracting structural features in fine-grained regions of the BEV image. In particular, in bird's-eye-view images generated by lidar, where targets are prone to deformation or stretching during projection, the adaptive rectangular convolution module effectively enhances the model's adaptability to humanoid targets of varying scales and poses, while also controlling computational complexity and maintaining efficient operation on edge platforms such as the Jetson Nano. Furthermore, to address common issues such as low grayscale values, weak target contrast, and strong background interference in tunnel BEV images, the SILU activation function is adopted instead of the traditional ReLU. The SILU function retains a non-zero response even in negative input regions, exhibiting better information retention and gradient smoothing, helping to enhance the network's ability to discriminate against dim targets and mitigate the vanishing gradient problem. This mechanism, when applied to the BEV input used in this application, significantly improves the detection accuracy and robustness of small humanoid targets in low-contrast scenarios.

[0080] Finally, the WGIoU loss function is used to reversely adjust the ARDMADet model in a complex tunnel environment, which can significantly improve the robustness and accuracy of detection.

[0081] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for establishing a tunnel anomaly detection model, characterized by: Establish an ARDMADet model, which is an improved YOLOv10 model. The ARDMADet model includes a backbone network for feature extraction of input images, a neck network for feature extraction and fusion of feature maps, and a head structure for detection and classification of the fused feature maps output by the neck network. Replace the C2f module in the backbone network of the YOLOv10 model with the feature extraction unit SADMA_AR module, which is an improvement based on the C2f module; The feature extraction unit SADMA_AR module adds a first SADMA module to the Concat module and the convolution module of the C2f module, replaces both convolution modules in the C2f module with a first adaptive rectangular convolution module, and sequentially processes the feature map output by the first adaptive rectangular convolution module through normalization and a first activation function, and replaces the Bottleneck module in the C2f module with a Mablock module; The Mablock module includes two second adaptive rectangular convolution modules, a second SADMA module and a connection module connected in sequence. The feature map output by the second adaptive rectangular convolution module is sequentially normalized and processed by the first activation function. The connection module is used to connect the feature map input to the Mablock module with the feature map output by the second SADMA module; The first SADMA module and the second SADMA module each include four branches: The first branch performs a first rotation on the initial feature map to obtain a first rotated feature map; The second branch performs maximum pooling and average pooling operations on the first rotated feature map, uses deformable convolution modules of different dimensions to extract features from the pooled feature maps respectively, and splices the feature maps of different dimensions and outputs them using the second activation function; Performing an element-by-element multiplication operation on the first rotated feature map and the feature map output by the second branch, performing a first reverse rotation, and then outputting the result; The third branch performs a second rotation on the initial feature map to obtain a second rotated feature map, performs maximum pooling and average pooling operations on the second rotated feature map, uses deformable convolution modules of different dimensions to extract features from the pooled feature maps, splices the feature maps of different dimensions and outputs them using a second activation function, and performs a second reverse rotation on the feature map output by the second activation function before outputting it; The fourth branch performs maximum pooling and average pooling operations on the initial feature map, uses deformable convolution modules of different dimensions to extract features from the pooled feature maps, concatenates the feature maps of different dimensions, and outputs them using the second activation function. The feature map output by the second activation function is element-wise multiplied with the initial feature map and then output. The feature map output after the first reverse rotation, the feature map after the second reverse rotation, and the feature map output by the fourth branch are added together to obtain the average value and then output.

2. The method for establishing a tunnel anomaly detection model according to claim 1, wherein: The first rotation is a 90° counterclockwise rotation along the H axis, the first reverse rotation is a 90° clockwise rotation along the H axis, the second rotation is a 90° counterclockwise rotation along the W axis, and the second reverse rotation is a 90° clockwise rotation along the W axis.

3. The method for establishing a tunnel anomaly detection model according to claim 2, wherein: The second activation function is the Sigmoid activation function.

4. The method for establishing a tunnel anomaly detection model according to claim 1, wherein: The first activation function is the SILU activation function.

5. A tunnel anomaly detection and early warning method, characterized by: The process includes the following steps: S1: Acquire 3D point cloud data in the tunnel in real time, project the 3D point cloud data to obtain a 2D data set, preprocess the 2D data set, and divide the preprocessed 2D data set into a training set, a validation set, and a test set according to a preset ratio; S2: inputting the training set into the ARDMADet model established by the tunnel anomaly detection model establishment method according to any one of claims 1 to 4 to train the ARDMADet model; S3: Use the WGIoU loss function to reversely adjust the trained ARDMADet model to obtain the adjusted ARDMADet model. The calculation formula of the WGIoU loss function is as follows: ； Among them, A represents the real box, B represents the preset bounding box, and C represents the smallest box containing the real box and the predicted box. represents the coordinates of the center of the anchor box, Indicates the coordinates of the center of the target box, and Indicates the size of the minimum bounding box. ,in , is the outlier degree, is the IoU threshold between the predicted bounding box and the true bounding box, for The sliding average of , small outlier means high quality of anchor box, and is a hyperparameter, To utilize A non-monotonic focusing coefficient is constructed; S4: Input the test set into the adjusted ARDMADet model for testing and obtain the test results; S5: If the test result shows that there are people in the preset area, an alarm prompt will be issued.

6. The tunnel anomaly detection and early warning method according to claim 5, characterized in that: The preprocessing in step S1 includes performing background filtering, ROI region cropping and noise suppression on the two-dimensional data set in sequence.

7. A computer product, characterized in that: The invention comprises a computer program, which, when executed by a processor, implements the tunnel anomaly detection and early warning method according to any one of claims 5 to 6.

8. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the tunnel anomaly detection and early warning method described in any one of claims 5 to 6 is implemented.

Images