Rail cutting operation goggles wearing detection method, device, system, equipment and medium

Abstract

The application relates to a rail cutting operation goggles wearing detection method, device, system, equipment and medium, and relates to the technical field of target detection. The method comprises the following steps: inputting a to-be-detected image into multiple different scale deep separable convolution modules of a trained goggles recognition model to obtain convolution features; inputting first fusion features obtained after fusing all the convolution features into a channel attention module and a spatial attention module of the goggles recognition model respectively to obtain channel features and spatial features output by the spatial attention module; and inputting second fusion features obtained after fusing the channel features and the spatial features into a position detection module and a state detection module of the goggles recognition model respectively to obtain a goggles wearing state of a target object and a goggles wearing position of the target object output by the position detection module. The method can improve the accuracy of the goggles wearing state detection result under actual working conditions.

Description

Technical Field

[0001] This application relates to the field of target detection technology, and in particular to methods, devices, systems, equipment and media for detecting the wearing of safety goggles in rail cutting operations. Background Technology

[0002] With the continuous expansion of the railway transportation network and the ongoing trends towards heavy-haul and high-speed operations, the frequency of rail maintenance and replacement work has increased significantly, leading to a rise in operational safety risks. Among various on-site rail construction operations, cutting work poses a serious threat to the eye safety of workers due to high temperatures, strong light, and flying metal particles. Therefore, properly wearing standard-compliant safety goggles has become a crucial measure to protect the occupational health and safety of frontline workers.

[0003] Currently, the detection of personal protective equipment is mostly based on general target detection models, which are difficult to adapt to actual working conditions such as complex lighting, occlusion, and posture changes. This results in low accuracy of the detection results of goggle wearing status under such actual working conditions.

[0004] Therefore, improving the accuracy of goggle wearing status detection results under actual working conditions is a problem that needs to be solved. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, device, system, computer equipment, computer-readable storage medium, and computer program product for detecting the wearing status of safety goggles in rail cutting operations, in order to improve the accuracy of the detection results of safety goggle wearing status under actual working conditions.

[0006] Firstly, this application provides a method for detecting the wearing of safety goggles during rail cutting operations, including:

[0007] Acquire the image to be detected, including the facial region of the target object;

[0008] The image to be detected is input into multiple depthwise separable convolutional modules of different scales of the trained goggle recognition model to obtain the convolutional features output by each depthwise separable convolutional module.

[0009] The first fused feature, obtained by fusing all the convolutional features, is input into the channel attention module and the spatial attention module of the goggle recognition model, respectively, to obtain the channel features output by the channel attention module and the spatial features output by the spatial attention module.

[0010] The second fused feature, obtained by fusing the channel features and the spatial features, is input into the position detection module and the state detection module of the goggle recognition model, respectively, to obtain the goggle wearing state of the target object output by the state detection module and the goggle wearing position of the target object output by the position detection module; the goggle wearing state and the goggle wearing position represent the goggle wearing detection result of the target object.

[0011] In one embodiment, the goggle recognition model is trained through the following steps:

[0012] Input the goggle sample image into the initial recognition model to obtain the goggle wearing training state and goggle wearing training position output by the initial recognition model;

[0013] Based on the goggle-wearing training state and the actual goggle-wearing state in the goggle sample images, the state loss is determined;

[0014] Based on the goggle wearing training position and the actual goggle wearing position in the goggle sample image, the position loss is determined;

[0015] Based on the state loss and the position loss, the model parameters of the initial recognition model are adjusted to obtain the goggle recognition model.

[0016] In one embodiment, the step of inputting the image to be detected into multiple depthwise separable convolutional modules of different scales in a trained goggle recognition model, and obtaining the convolutional features output by each depthwise separable convolutional module, includes:

[0017] The following method is performed in each of the depthwise separable convolutional modules:

[0018] Based on the number of channels in each depth-separable convolution module, the corresponding image to be detected is convolved channel by channel to obtain the feature map of the corresponding channel;

[0019] In each of the depthwise separable convolutional modules, the feature maps of all channels are convolved at spatial pixel points to obtain the convolutional features output by each depthwise separable convolutional module.

[0020] In one embodiment, the first fused feature, obtained by fusing all the convolutional features, is input into the channel attention module and the spatial attention module of the goggle recognition model, respectively, to obtain the channel features output by the channel attention module and the spatial features output by the spatial attention module, including:

[0021] All the convolutional features are fused to obtain the first fused feature;

[0022] The first fused feature is input into the channel attention module to perform the following steps:

[0023] The first fused feature is subjected to global average pooling and global max pooling in the spatial dimension to obtain the average pooling feature and max pooling feature respectively; the average pooling feature and the max pooling feature are fused across channels to obtain the channel attention feature; the channel attention feature and the first fused feature are multiplied channel by channel in the channel dimension to obtain the channel feature.

[0024] The first fused feature is input into the spatial attention module to perform the following steps:

[0025] The first fused feature is subjected to average pooling and max pooling in the channel dimension to obtain the corresponding first channel feature and second channel feature; the first channel feature and the second channel feature are convolved to obtain the spatial attention feature; the spatial attention feature and the first fused feature are multiplied pixel by pixel in the spatial dimension to obtain the spatial feature.

[0026] In one embodiment, after the second fused feature, obtained by fusing the channel features and the spatial features, is input into the position detection module and the state detection module of the goggle recognition model, respectively, to obtain the goggle-wearing state of the target object output by the position detection module, the method further includes:

[0027] Based on the detection results of the target object wearing goggles, the safety level of the target object wearing goggles is determined;

[0028] Based on the security level and the preset violation level, the alarm information is determined;

[0029] The alarm information is transmitted to the terminal in the scene where the target object is located, so that the terminal can warn the target object about the status of wearing goggles.

[0030] In one embodiment, acquiring the image to be detected, including the facial region of the target object, includes:

[0031] Obtain the environmental parameters of the environment in which the target object is located;

[0032] Based on the aforementioned environmental parameters, the interference factors for rail cutting operations are determined.

[0033] Adjust the acquisition parameters of the image acquisition device based on the interference factor;

[0034] Based on the acquisition parameters, the image acquisition device is controlled to acquire the image to be detected of the facial region of the target object.

[0035] Secondly, this application also provides a detection device for the wearing of safety goggles during rail cutting operations, comprising:

[0036] The image acquisition module is used to acquire an image to be detected, including the facial region of the target object;

[0037] The convolution module is used to input the image to be detected into multiple depth-separable convolution modules of different scales of the trained goggle recognition model to obtain the convolution features output by each depth-separable convolution module.

[0038] An attention module is used to input the first fused feature, which is the result of fusing all the convolutional features, into the channel attention module and the spatial attention module of the goggle recognition model, respectively, to obtain the channel features output by the channel attention module and the spatial features output by the spatial attention module.

[0039] The detection module is used to input the second fused feature, which is the result of fusing the channel features and the spatial features, into the position detection module and the state detection module of the goggle recognition model, respectively, to obtain the goggle wearing state of the target object output by the state detection module and the goggle wearing position of the target object output by the position detection module; the goggle wearing state and the goggle wearing position represent the goggle wearing detection result of the target object.

[0040] Thirdly, this application also provides a detection system for the wearing of safety goggles during rail cutting operations, the system comprising:

[0041] Image acquisition module, edge processing module, and alarm feedback module;

[0042] The image acquisition module is used to acquire the image to be detected of the facial region of the target object and transmit the image to be detected to the edge processing module;

[0043] The edge processing module is used to execute the goggle wearing detection method for rail cutting operations described in any of the embodiments of the first aspect above, so as to obtain the goggle wearing detection result of the target object, and transmit the goggle wearing detection result to the alarm feedback module;

[0044] The alarm feedback module is used to issue a warning about the goggles wearing status based on the goggles wearing detection results.

[0045] Fourthly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the rail cutting operation goggle wearing detection method in any of the embodiments of the first aspect above.

[0046] Fifthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for detecting the wearing of safety goggles for rail cutting operations in any of the embodiments of the first aspect described above.

[0047] Sixthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the rail cutting operation goggle wearing detection method in any of the embodiments of the first aspect described above.

[0048] In the above implementation process, the image to be detected is convolved through multiple depthwise separable convolutional modules at different scales of the goggle recognition model to obtain the convolutional features output by each depthwise separable convolutional module. Compared with traditional standard convolution operations, depthwise separable convolutional modules can reduce the number of parameters and computational complexity of the goggle recognition model. Furthermore, multi-scale feature extraction ensures that the goggle recognition model can identify both the minute features of distant goggles and the detailed textures of large nearby targets, improving the model's adaptability to multi-scale targets and facilitating higher accuracy in detection results under real-world conditions. All convolutional features are fused into a first fusion feature, effectively integrating multi-level semantic information. Furthermore, processing the first fused feature through the channel attention module enables it to automatically learn important information from different feature channels. Processing the first fused feature through the spatial attention module guides the model to focus on key spatial regions in the image containing goggles (such as around the eyes). Then, the outputs of the channel attention module and the spatial attention module are fused into the second fused feature, enabling the goggle recognition model to focus on specific targets and their locations. Finally, the second fused feature is input into the position detection module and the state detection module respectively, allowing the goggle recognition model to perform dual-task detection of both the goggle wearing status and the goggle wearing position of the target object, effectively improving the accuracy of goggle wearing status detection results under actual working conditions. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a schematic diagram illustrating the application environment of a method for detecting the wearing of safety goggles during rail cutting operations, as provided in an embodiment of this application.

[0051] Figure 2 This is a flowchart illustrating a method for detecting the wearing of safety goggles during rail cutting operations, as provided in an embodiment of this application.

[0052] Figure 3 This is a schematic diagram of the structure of a safety goggle wearing detection device for rail cutting operations provided in an embodiment of this application;

[0053] Figure 4 This is a schematic diagram of a safety goggle wearing detection system for rail cutting operations provided in an embodiment of this application;

[0054] Figure 5 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0056] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0057] The method for detecting the wearing of safety goggles during rail cutting operations provided in this application embodiment can be applied to, for example... Figure 1 The application environment shown. Figure 1 This is a schematic diagram illustrating the application environment of a method for detecting the wearing of safety goggles during rail cutting operations, as provided in an embodiment of this application. Figure 1As shown, the terminal communicates with the server via a network. The data storage system can store the data that the server needs to process. The data storage system can be integrated into the server, or it can be located in the cloud or on other network servers. The terminal can be an image acquisition device installed at the rail cutting construction site. The server can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, a cloud server providing cloud computing services, or a server located in an edge computing device.

[0058] In one exemplary embodiment, Figure 2 This is a flowchart illustrating a method for detecting the wearing of safety goggles during rail cutting operations, provided in an embodiment of this application. The method is applied to... Figure 1 Let's take the server in the example of this as an illustration. Figure 2 As shown, the method may include steps 201 to 204. Wherein:

[0059] Step 201: Obtain the image to be detected, which includes the facial region of the target object.

[0060] For example, when a target object is being cut into rails, an image acquisition device placed in the environment where the target object is located can acquire image information of the target object.

[0061] Specifically, industrial-grade cameras can be deployed at the rail cutting site, targeting the workers performing the rail cutting operation. The deployed industrial-grade cameras capture video streams containing the workers' facial areas. Dynamic interference factors are generated based on the environmental parameters of the rail cutting site, and the image imaging parameters are dynamically adjusted using these interference factors. The dynamically adjusted image imaging parameters are then subjected to image preprocessing operations for goggle recognition, outputting an image frame sequence adapted for subsequent recognition tasks. This image video frame is used as the image to be detected, including the facial area of ​​the target object.

[0062] As an example, preprocessing operations on images that can be acquired by image acquisition devices may include performing lightweight face detection on the acquired raw video stream, locking the region of interest on the worker's face, cropping to generate local image patches centered on the eyes, narrowing the processing range and focusing on the important areas of the goggles, and further performing targeted preprocessing operations such as noise reduction, contrast enhancement and detail sharpening on the local image patches to eliminate residual interference and highlight the edges and wearing features of the goggles, ultimately outputting a sequence of image frames with high signal-to-noise ratio, high recognition and suitable for downstream recognition tasks.

[0063] Furthermore, the image acquisition device transmits the acquired image to be detected to the server, so that the server can obtain the image to be detected, including the facial region of the target object.

[0064] Step 202: Input the image to be detected into multiple depth-separable convolutional modules of different scales of the trained goggle recognition model to obtain the convolutional features output by each depth-separable convolutional module.

[0065] For example, the wearing status of goggles during rail cutting operations can be detected based on a trained goggle recognition model and the image to be detected. As an example, the trained goggle recognition model may include multiple depthwise separable convolutional modules of different scales, channel attention modules, spatial attention modules, position detection modules, and state detection modules.

[0066] Specifically, when the goggles recognition model detects the goggles wearing status of the image to be detected, it can input the image to be detected into multiple depthwise separable convolutional modules of different scales of the trained goggles recognition model, so that each depthwise separable convolutional module convolves the image to be detected, and obtains the convolutional features output by each depthwise separable convolutional module.

[0067] As an example, multiple depthwise separable convolutional modules at different scales are run in parallel in the goggle recognition model. Each depthwise separable convolutional module at each scale is a convolutional branch used to capture local details and global contextual information in the image to be detected from the corresponding scale, thereby obtaining the convolutional features output by each depthwise separable convolutional module.

[0068] Step 203: Input the first fused feature, which is obtained by fusing all convolutional features, into the channel attention module and the spatial attention module of the goggle recognition model respectively, to obtain the channel features output by the channel attention module and the spatial features output by the spatial attention module.

[0069] Furthermore, the trained goggle recognition model may also include a feature fusion layer, which is used to fuse the convolutional features output by all depthwise separable convolutional modules to obtain the first fused feature.

[0070] Furthermore, the channel attention module and the spatial attention module run in parallel in the goggle recognition model. The first fused feature is input into the channel attention module and the spatial attention module respectively, so that the channel attention module and the spatial attention module dynamically weight the first fused feature to obtain the channel feature output by the channel attention module and the spatial feature output by the spatial attention module.

[0071] Step 204: Input the second fused feature, which is the fusion of channel features and spatial features, into the position detection module and state detection module of the goggle recognition model, respectively, to obtain the goggle wearing status of the target object output by the state detection module and the goggle wearing position of the target object output by the position detection module.

[0072] Among them, the goggle wearing status and the goggle wearing position represent the goggle wearing detection results of the target object.

[0073] For example, the goggle recognition model may further include an attention fusion layer for fusing channel features and spatial features to obtain a second fused feature. Further, the second fused feature is input into the position detection module and the state detection module, respectively, so that the position detection module determines the goggle wearing position of the target object based on the second fused feature, and the state detection module outputs the goggle wearing state of the target object.

[0074] Specifically, the position detection module and state detection module in the goggle recognition model are dual-task-head output structures sharing features. These modules are used to determine the goggle-wearing status and position of the target object. The goggle-wearing status can include one of the following: properly worn, not worn, or worn askew. The goggle-wearing position can include the bounding box coordinates of the goggle frame in the corresponding image to be detected.

[0075] Furthermore, the goggle-wearing status and position of the target object are determined as the goggle-wearing detection results for the target object.

[0076] In the above implementation process, the image to be detected is convolved through multiple depthwise separable convolutional modules at different scales of the goggle recognition model to obtain the convolutional features output by each depthwise separable convolutional module. Compared with traditional standard convolution operations, depthwise separable convolutional modules can reduce the number of parameters and computational complexity of the goggle recognition model. Furthermore, multi-scale feature extraction ensures that the goggle recognition model can identify both the minute features of distant goggles and the detailed textures of large nearby targets, improving the model's adaptability to multi-scale targets and facilitating higher accuracy in detection results under real-world conditions. All convolutional features are fused into a first fusion feature, effectively integrating multi-level semantic information. Furthermore, processing the first fused feature through the channel attention module enables it to automatically learn important information from different feature channels. Processing the first fused feature through the spatial attention module guides the model to focus on key spatial regions in the image containing goggles (such as around the eyes). Then, the outputs of the channel attention module and the spatial attention module are fused into the second fused feature, enabling the goggle recognition model to focus on specific targets and their locations. Finally, the second fused feature is input into the position detection module and the state detection module respectively, allowing the goggle recognition model to perform dual-task detection of both the goggle wearing status and the goggle wearing position of the target object, effectively improving the accuracy of goggle wearing status detection results under actual working conditions.

[0077] In one exemplary embodiment, the goggle recognition model is trained through the following steps:

[0078] Input the goggle sample images into the initial recognition model to obtain the goggle wearing training state and goggle wearing training position output by the initial recognition model; based on the goggle wearing training state and the actual goggle wearing state in the goggle sample images, determine the state loss; based on the goggle wearing training position and the actual goggle wearing position in the goggle sample images, determine the position loss; based on the state loss and position loss, adjust the model parameters of the initial recognition model to obtain the goggle recognition model.

[0079] For example, during the training of the goggle recognition model, goggle sample images can be collected at the rail cutting operation site and in a simulated environment to construct a goggle sample image dataset for rail cutting operation scenarios. Specifically, the collected goggle sample images can include goggle image samples under different lighting conditions, head postures, occlusion situations, and various wearing states (including standard wearing, not wearing, wearing at an angle, and being obscured by the helmet strap, etc.). A goggle sample image dataset that highly closely reflects actual working conditions is constructed using a few-shot learning strategy. This facilitates the training of the initial recognition model using the goggle sample image dataset.

[0080] Specifically, goggle sample images from the goggle sample image dataset are input into the initial recognition model. Multiple depthwise separable convolutional modules at different scales in the initial recognition model capture local details and global contextual information respectively. Goggle-related feature channels are dynamically weighted through channel attention modules and spatial attention modules. The outputs of the channel attention modules and spatial attention modules are then fused and input into the state detection module and position detection module respectively. The dual-task head output structure composed of shared features from the state detection module and position detection module outputs the goggle wearing training state and goggle wearing training position respectively.

[0081] As an example, a goggle recognition model may include an input layer, multiple depthwise separable convolutional modules of different scales, a feature fusion layer, a channel attention module, a spatial attention module, an attention fusion layer, a state detection module, and a position detection module.

[0082] The input layer can be a 1×1 convolutional layer, which is used to distribute the feature map of the input image after channel dimension adjustment to multiple parallel depth-separable convolutional modules of different scales.

[0083] Multiple depthwise separable convolutional modules of different scales can be three depthwise separable convolutional layers with scales of 3×3, 5×5, and 7×7. The 3×3 depthwise separable convolutional layer can extract high-frequency texture features such as the goggle frame and straps; the 5×5 depthwise separable convolutional layer, which can be a 5×5 dilated convolution (dilation rate d=2), can expand the receptive field without increasing the number of parameters, capturing the overall facial layout across the spark region; the 7×7 depthwise separable convolutional layer can preserve the original shallow layer features, preventing the deep network from losing the absolute positional information of the goggles. Thus, each depthwise separable convolutional layer outputs its corresponding convolutional features.

[0084] The feature fusion layer concatenates all convolutional features along the channel dimension, and then passes the concatenated features through a 1×1 standard convolutional layer. This layer performs cross-channel linear combination and information dimensionality reduction fusion on the concatenated features, enabling the network to adaptively learn the importance of features at different scales, thereby obtaining the first fused feature.

[0085] Furthermore, the channel attention module and the spatial attention module can be of the structure of CBAM (convolutional block attention module), which can dynamically generate weights on the first fused feature, strengthen the feature channels related to the goggles, and suppress background interference information such as sparks and dust.

[0086] The attention fusion layer fuses the outputs of the channel attention module and the spatial attention module to obtain a second fused feature, so that the second fused feature contains high-quality features.

[0087] Furthermore, the second fusion feature is input into the dual-task head output structure of the state detection module and the position detection module respectively, so that the state detection module and the position detection module output the goggle wearing training state and the goggle wearing training position respectively.

[0088] Furthermore, based on the training state of goggles wearing and the actual wearing state of goggles in the goggles sample images, a state loss is determined; based on the training position of goggles wearing and the actual wearing position of goggles in the goggles sample images, a position loss is determined. Then, based on the joint loss function, state loss, and position loss, the total loss is determined. The model parameters of the initial recognition model are then adjusted using the total loss to obtain the goggles recognition model.

[0089] As an example, the joint loss function can be determined by the following expression:

[0090] ;

[0091] in, This represents the total loss corresponding to the joint loss function; This indicates the loss of position of the goggle frame; This indicates a loss of the ability to wear protective goggles. The weighting coefficients representing the location loss; The weighting coefficients represent the state loss.

[0092] By calculating the total loss of different goggle sample images during the training process, and iteratively adjusting the model parameters of the initial recognition model based on the total loss until the number of iterations reaches a preset value, the corresponding adjusted model is determined as the trained goggle recognition model.

[0093] In the above implementation process, during the training of the initial recognition model, the state loss and position loss are used together for backpropagation to update the model parameters. This enables the trained goggle recognition model to take into account both the wearing state and position of the goggles, avoiding the overfitting problem that may be caused by single-task training, and significantly improving the model's comprehensive generalization ability and detection accuracy.

[0094] In an exemplary embodiment, the image to be detected is input into multiple depthwise separable convolutional modules of different scales in a trained goggle recognition model to obtain the convolutional features output by each depthwise separable convolutional module, including:

[0095] In each depthwise separable convolutional module, the following method is performed: based on the number of channels in each depthwise separable convolutional module, the corresponding image to be detected is convolved channel by channel to obtain the feature map of the corresponding channel; in each depthwise separable convolutional module, the feature maps of all channels are convolved at spatial pixels to obtain the convolutional features output by each depthwise separable convolutional module.

[0096] For example, in each depth-separable convolution module, the image to be detected is first convolved channel by channel to extract contours and textures, and then convolved point by point to achieve cross-channel fusion.

[0097] Specifically, if the input image to be detected in each depthwise separable convolutional module has 3 channels (red, green, and blue), a convolutional kernel with the same number of channels (e.g., 3×3×C) is used. in Spatial feature extraction is performed. The convolutional kernel is only responsible for sliding in space without changing the number of channels. The computational cost is minimal, and the feature maps of the corresponding channels are obtained.

[0098] Furthermore, in each depthwise separable convolutional module, a 1×1 convolution is performed on the feature maps of all channels at the same spatial pixel point to obtain the convolutional features output by the corresponding depthwise separable convolutional module, thereby realizing the cross-channel linear combination and information fusion of feature maps of all channels.

[0099] This method allows each depthwise separable convolutional module to output its corresponding convolutional features.

[0100] In the above implementation, each depthwise separable convolutional module first performs channel-wise convolution on the image to be detected, realizing the processing of spatial information in a single channel and avoiding the dense multiplication operations in the channel dimension of traditional convolution, thus greatly reducing the amount of computation. Then, the feature maps of all channels are convolved at spatial pixels, achieving cross-channel feature fusion to obtain the convolutional features output by each depthwise separable convolutional module. In other words, the channel-wise convolution and spatial pixel-wise convolution of each depthwise separable convolutional module effectively achieve a lightweight design, enabling the model to significantly reduce memory usage and computational overhead while maintaining feature extraction capabilities.

[0101] In an exemplary embodiment, the first fused feature, obtained by fusing all convolutional features, is input into the channel attention module and the spatial attention module of the goggle recognition model, respectively, to obtain the channel features output by the channel attention module and the spatial features output by the spatial attention module, including:

[0102] All convolutional features are fused together to obtain the first fused feature.

[0103] The first fused feature is input into the channel attention module and the following steps are performed: the first fused feature is subjected to global average pooling and global max pooling in the spatial dimension to obtain the average pooling feature and max pooling feature respectively; the average pooling feature and max pooling feature are fused across channels to obtain the channel attention feature; the channel attention feature and the first fused feature are multiplied channel by channel in the channel dimension to obtain the channel feature.

[0104] The first fused feature is input into the spatial attention module and the following steps are performed: the first fused feature is average pooled and max pooled in the channel dimension to obtain the first channel feature and the second channel feature respectively; the first channel feature and the second channel feature are convolved to obtain the spatial attention feature; the spatial attention feature and the first fused feature are multiplied pixel by pixel in the spatial dimension to obtain the spatial feature.

[0105] For example, the feature fusion layer fuses the convolutional features output by all depthwise separable convolutional modules to obtain a first fused feature. This first fused feature is then input into both the channel attention module and the spatial attention module.

[0106] As an example, if the dimensions of the first fused feature are C×H×W.

[0107] In the channel attention module, the first fused feature is subjected to global average pooling and global max pooling along the channel dimension (C) to compress spatial information, resulting in corresponding average pooling and max pooling features, each a one-dimensional vector of length C. Further, the average pooling and max pooling features are fed into a shared multilayer perceptron (MLP, typically containing a dimensionality reduction layer and an dimensionality increase layer) for cross-channel information fusion. The fused information is then processed by an activation function (Sigmoid function) to map the values ​​to between 0 and 1, yielding the channel attention feature, which is a channel weight vector of dimension C×1×1. The channel attention feature is then multiplied channel-by-channel along the channel dimension with the first fused feature to obtain the channel features. This amplifies the weights of feature channels related to goggles (such as edge detection channels) (approaching 1), while suppressing the weights of redundant channels related to sparks and dust (approaching 0).

[0108] In the spatial attention module, the first fused feature is subjected to average pooling and max pooling along the channel dimension, resulting in the first channel feature and the second channel feature, thus compressing C channels into two single-channel feature maps (size 2×H×W). The two channel feature maps are then concatenated along the channel direction and fed into a 7×7 convolutional layer to integrate spatial context information. This spatial context information is then activated by a sigmoid function to generate a spatial attention mask with the same spatial size as the original feature map, thus obtaining the spatial attention feature. Further, the spatial attention feature is multiplied pixel-by-pixel along the spatial dimension with the first fused feature to obtain the spatial feature. This significantly enhances the pixel weights of the area where the goggles are located, while forcibly reducing the weights of the surrounding background areas such as sparks, dust, and facial skin.

[0109] In the above implementation process, the channel attention module captures the overall response intensity and local salient response intensity of each channel through global average pooling and global max pooling, respectively, to obtain average pooling features and max pooling features. The channel attention features obtained by fusing these two features across channels enable the model to accurately extract information from important channels, thereby automatically enhancing the channel responses related to key semantic information such as goggle material, frame, and reflective strips, while suppressing the weights of channels irrelevant to background noise and spark interference. The spatial attention module performs average pooling and max pooling in the channel dimension, compressing multi-channel information into two two-dimensional spatial response maps, which are then convolved and fused to generate spatial attention features, enabling the model to accurately focus on important locations. This dual attention mechanism allows the goggle recognition model to select the correct feature dimension and lock onto key areas when facing complex backgrounds at rail cutting sites, significantly reducing the probability of false positives and false negatives and improving the accuracy of the recognition results.

[0110] In an exemplary embodiment, after the second fused feature, which combines the channel features and spatial features, is input into the position detection module and the state detection module of the goggle recognition model, respectively, to obtain the goggle-wearing state of the target object output by the position detection module, the method further includes:

[0111] Based on the detection results of the target object wearing goggles, determine the safety level of the target object wearing goggles; based on the safety level and the preset violation level, determine the alarm information; transmit the alarm information to the terminal in the scene where the target object is located, so that the terminal can warn the target object about the goggles wearing status.

[0112] For example, after determining the goggle-wearing detection results of the target object, the safety level of the target object wearing goggles can also be determined based on the goggle-wearing detection results. Specifically, safety level association information can be pre-set, which includes the safety level corresponding to different goggle-wearing detection results and the violation level corresponding to each safety level. Then, based on the detected goggle-wearing detection results of the target object, the corresponding safety level of the target object wearing goggles and the corresponding violation level can be matched from the safety level association information.

[0113] Furthermore, corresponding alarm messages can be pre-assigned to each violation level, thus determining the alarm message corresponding to the target object's goggle-wearing detection result. For example, if the target object's goggle-wearing detection result is "wearing correctly," the safety level is Level 1, the corresponding violation level is "none," and the corresponding alarm message is "none." If the target object's goggle-wearing detection result is "wearing at an angle," the safety level is Level 2, the corresponding violation level is "basic," and the corresponding alarm message is "wearing goggles correctly." If the target object's goggle-wearing detection result is "not wearing," the safety level is Level 3, the corresponding violation level is "high," and the corresponding alarm message is "wear goggles promptly."

[0114] Furthermore, the confirmed alarm information is transmitted to the terminal in the scene where the target object is located, so that the target object can be alerted about the wearing status of goggles through the terminal. Specifically, standardized alarm commands can be synchronously pushed to the corresponding multi-level safety response terminals through the industrial communication interface. When the field-level terminal triggers the audible and visual alarm, the operation-level terminal displays the personnel alarm screen, and the platform-level terminal records the event and initiates the supervision process.

[0115] In the above implementation process, the corresponding safety level is determined by detecting the wearing of goggles by the target object, and the violation level is determined based on the safety level. This allows for accurate alarm information to be generated based on the violation level and transmitted to the terminal in the target object's location, thus providing an alarm notification. This effectively achieves the grading and quantification of risk levels, facilitating differentiated response strategies for the terminal based on the actual degree of danger.

[0116] In one exemplary embodiment, acquiring a detectable image including the facial region of a target object includes:

[0117] The process involves: acquiring environmental parameters of the target object's environment; determining interference factors for the rail cutting operation based on these parameters; adjusting the acquisition parameters of the image acquisition device based on the interference factors; and controlling the image acquisition device based on these parameters to acquire images of the target object's facial region.

[0118] For example, in order to accurately acquire facial region images of the target object, the server can also obtain various environmental parameters such as ambient light intensity, spark flicker frequency, and dust concentration at the rail cutting site where the target object is located through various sensors, and convert the acquired environmental parameters into structured interference factors for the rail cutting operation. These interference factors can be vectors.

[0119] Furthermore, the server can adjust the acquisition parameters of the image acquisition device based on this interference factor. For example, it can adjust the exposure, white balance, contrast, and other acquisition parameters of the image acquisition device, and control the image acquisition device to acquire the image to be detected of the facial region of the target object with the adjusted acquisition parameters. This allows the image acquisition device to stably acquire original video frames containing clear facial information even under extreme conditions such as strong light overexposure, low light backlight, or dust obscuring.

[0120] In addition, lightweight face detection can be performed on the adjusted original video frames to quickly locate the facial area of ​​the workers and crop out local image blocks centered on the eyes, effectively narrowing the scope of subsequent processing and focusing on the key area where the goggles are located. Finally, targeted preprocessing operations such as noise reduction, contrast enhancement and detail sharpening are performed on the local image blocks to eliminate residual interference and highlight the edges and wearing features of the goggles, ultimately outputting an image frame sequence with high signal-to-noise ratio, high recognition and adaptability to downstream recognition tasks.

[0121] In the above implementation process, the environmental parameters of the rail cutting operation site are transformed into interference factors, and the acquisition parameters are dynamically adjusted through the interference factors. This can effectively suppress environmental noise and ensure that the acquired facial area images to be detected always remain clear and have high contrast. This enhances the generalization ability of the acquired images to the scene and facilitates the improvement of the accuracy of subsequent detection results.

[0122] In one exemplary embodiment, the method may further include:

[0123] The results of the goggle-wearing detection of the target object are used to generate a security audit data storage package, which is then encrypted and stored locally. When local network conditions permit, the security audit data storage package is encrypted and uploaded to the superior proactive security control platform via a secure transmission protocol. Upon receiving a confirmation signal from the superior proactive security control platform, the entire closed-loop management process is completed.

[0124] Specifically, based on the goggle-wearing detection results of the target object, the system can analyze the structured elements of the goggle-wearing event, including the wearing status category, bounding box position, confidence score, and timestamp. It can also automatically determine whether the target object has engaged in violations such as not wearing goggles or not wearing them correctly, according to preset safety rules. Once a violation is determined, standardized alarm information is immediately generated based on the violation level and simultaneously pushed accurately to the three-level safety response terminal via industrial communication interfaces such as RS485, 4G / 5G, or Wi-Fi. The field-level terminal immediately triggers an audible and visual alarm to alert the workers, while the work-level terminal simultaneously displays the image of the personnel involved for the work team. During long-term intervention, the platform-level terminal records the event and initiates a remote monitoring process. It can also automatically associate the original video clips that triggered the alarm with the corresponding structured judgment data, package them into a security audit data storage package containing a complete chain of evidence, and store it locally in an encrypted manner to ensure that the data cannot be tampered with. Finally, when network conditions permit, the evidence upload and closed-loop confirmation submodule encrypts and uploads the storage package to the superior proactive security control platform through a secure transmission protocol, and waits for the platform to return a confirmation signal. This completes the closed-loop management of the entire process from real-time detection and hierarchical alarms to evidence retention and remote confirmation, effectively supporting post-event traceability, responsibility determination, and security assessment.

[0125] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0126] Based on the same inventive concept, this application also provides a rail cutting operation goggle wearing detection device for implementing the above-mentioned rail cutting operation goggle wearing detection method. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more rail cutting operation goggle wearing detection device embodiments provided below can be found in the limitations of the rail cutting operation goggle wearing detection method above, and will not be repeated here.

[0127] In one exemplary embodiment, Figure 3 This is a schematic diagram of the structure of a safety goggle detection device for rail cutting operations provided in an embodiment of this application, as shown below. Figure 3 As shown, the device may include:

[0128] Image acquisition module 301 is used to acquire an image to be detected, including the facial region of the target object;

[0129] Convolution module 302 is used to input the image to be detected into multiple depth-separable convolution modules of different scales of the trained goggle recognition model to obtain the convolution features output by each depth-separable convolution module.

[0130] Attention module 303 is used to input the first fused feature after fusing all convolutional features into the channel attention module and spatial attention module of the goggle recognition model, respectively, to obtain the channel features output by the channel attention module and the spatial features output by the spatial attention module.

[0131] The detection module 304 is used to input the second fused feature, which is the result of fusing the channel features and spatial features, into the position detection module and the state detection module of the goggle recognition model, respectively, to obtain the goggle wearing status of the target object output by the state detection module and the goggle wearing position of the target object output by the position detection module; the goggle wearing status and the goggle wearing position represent the goggle wearing detection results of the target object.

[0132] In one exemplary embodiment, the goggle recognition model is trained through the following steps:

[0133] Input the goggle sample image into the initial recognition model to obtain the goggle wearing training state and goggle wearing training position output by the initial recognition model;

[0134] Based on the training state of goggles wearing and the actual state of goggles wearing in the goggles sample images, the state loss is determined.

[0135] The positional loss is determined based on the training position of the goggles and the actual position of the goggles in the goggles sample images;

[0136] Based on state loss and position loss, the model parameters of the initial recognition model are adjusted to obtain the goggle recognition model.

[0137] In one exemplary embodiment, the convolution module 302 is specifically used for:

[0138] Perform the following method in each depthwise separable convolutional module:

[0139] Based on the number of channels in each depth-separable convolutional module, the corresponding image to be detected is convolved channel by channel to obtain the feature map of the corresponding channel;

[0140] In each depthwise separable convolutional module, the feature maps of all channels are convolved at spatial pixels to obtain the convolutional features output by each depthwise separable convolutional module.

[0141] In one exemplary embodiment, the attention module 303 is specifically used for:

[0142] All convolutional features are fused to obtain the first fused feature;

[0143] The first fused feature is input into the channel attention module, and the following steps are performed:

[0144] The first fusion feature is subjected to global average pooling and global max pooling in the spatial dimension to obtain the average pooling feature and max pooling feature respectively; the average pooling feature and max pooling feature are fused across channels to obtain the channel attention feature; the channel attention feature and the first fusion feature are multiplied channel by channel in the channel dimension to obtain the channel feature.

[0145] The first fused feature is input into the spatial attention module to perform the following steps:

[0146] The first fused feature is subjected to average pooling and max pooling in the channel dimension to obtain the first channel feature and the second channel feature respectively; the first channel feature and the second channel feature are convolved to obtain the spatial attention feature; the spatial attention feature and the first fused feature are multiplied pixel by pixel in the spatial dimension to obtain the spatial feature.

[0147] In one exemplary embodiment, the detection module 304 is further configured to:

[0148] Based on the detection results of the target object wearing goggles, determine the safety level of the target object wearing goggles;

[0149] Based on the security level and the preset violation level, the alarm information is determined;

[0150] The alarm information is transmitted to the terminal in the scene where the target object is located, so that the terminal can alert the target object to the status of wearing goggles.

[0151] In one exemplary embodiment, the image acquisition module 301 is specifically used for:

[0152] Obtain environmental parameters of the target object's environment;

[0153] Based on environmental parameters, the interference factors of rail cutting operations are determined;

[0154] Adjusting the acquisition parameters of the image acquisition device based on the interference factor;

[0155] The image acquisition device is controlled based on the acquisition parameters to acquire the image of the facial region of the target object.

[0156] The various modules in the aforementioned safety goggle detection device for rail cutting operations can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0157] In one exemplary embodiment, this application also provides a detection system for the wearing of safety goggles during rail cutting operations. Figure 4 This is a schematic diagram of a safety goggle wearing detection system for rail cutting operations provided in an embodiment of this application. Figure 4 As shown, the system may include: an image acquisition module 401, an edge processing module 402, and an alarm feedback module 403.

[0158] Image acquisition module 401 is used to acquire the image to be detected of the facial region of the target object and transmit the image to be detected to the edge processing module 402;

[0159] The edge processing module 402 is used to execute the goggle wearing detection method for rail cutting operations in any of the above embodiments, so as to obtain the goggle wearing detection result of the target object, and transmit the goggle wearing detection result to the alarm feedback module 403;

[0160] The alarm feedback module 403 is used to issue warnings about the wearing status of goggles based on the goggles wearing detection results.

[0161] In one exemplary embodiment, a computer device is provided, which may be a server. Figure 5 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application, such as... Figure 5As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores the image data to be detected. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for detecting the wearing of safety goggles during rail cutting operations.

[0162] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0163] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for detecting the wearing of safety goggles in rail cutting operations as described in any of the above embodiments.

[0164] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method for detecting the wearing of safety goggles in rail cutting operations as described in any of the above embodiments.

[0165] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0166] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0167] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0168] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for detecting the wearing of safety goggles during rail cutting operations, characterized in that, The method includes: Acquire the image to be detected, including the facial region of the target object; The image to be detected is input into multiple depthwise separable convolutional modules of different scales of the trained goggle recognition model to obtain the convolutional features output by each depthwise separable convolutional module. The first fused feature, obtained by fusing all the convolutional features, is input into the channel attention module and the spatial attention module of the goggle recognition model, respectively, to obtain the channel features output by the channel attention module and the spatial features output by the spatial attention module. The second fused feature, obtained by fusing the channel features and the spatial features, is input into the position detection module and the state detection module of the goggle recognition model, respectively, to obtain the goggle wearing state of the target object output by the state detection module and the goggle wearing position of the target object output by the position detection module; the goggle wearing state and the goggle wearing position represent the goggle wearing detection result of the target object.

2. The method according to claim 1, characterized in that, The goggle recognition model was trained through the following steps: Input the goggle sample image into the initial recognition model to obtain the goggle wearing training state and goggle wearing training position output by the initial recognition model; Based on the goggle-wearing training state and the actual goggle-wearing state in the goggle sample images, the state loss is determined; Based on the goggle wearing training position and the actual goggle wearing position in the goggle sample image, the position loss is determined; Based on the state loss and the position loss, the model parameters of the initial recognition model are adjusted to obtain the goggle recognition model.

3. The method according to claim 1, characterized in that, The step of inputting the image to be detected into multiple depthwise separable convolutional modules of different scales in a trained goggle recognition model to obtain the convolutional features output by each depthwise separable convolutional module includes: The following method is performed in each of the depthwise separable convolutional modules: Based on the number of channels in each depth-separable convolution module, the corresponding image to be detected is convolved channel by channel to obtain the feature map of the corresponding channel; In each of the depthwise separable convolutional modules, the feature maps of all channels are convolved at spatial pixel points to obtain the convolutional features output by each depthwise separable convolutional module.

4. The method according to claim 1, characterized in that, The first fused feature, obtained by fusing all the convolutional features, is input into the channel attention module and the spatial attention module of the goggle recognition model, respectively, to obtain the channel features output by the channel attention module and the spatial features output by the spatial attention module, including: All the convolutional features are fused to obtain the first fused feature; The first fused feature is input into the channel attention module to perform the following steps: The first fused feature is subjected to global average pooling and global max pooling in the spatial dimension to obtain the average pooling feature and max pooling feature respectively; the average pooling feature and the max pooling feature are fused across channels to obtain the channel attention feature; the channel attention feature and the first fused feature are multiplied channel by channel in the channel dimension to obtain the channel feature. The first fused feature is input into the spatial attention module to perform the following steps: The first fused feature is subjected to average pooling and max pooling in the channel dimension to obtain the corresponding first channel feature and second channel feature; the first channel feature and the second channel feature are convolved to obtain the spatial attention feature; the spatial attention feature and the first fused feature are multiplied pixel by pixel in the spatial dimension to obtain the spatial feature.

5. The method according to claim 1, characterized in that, After the second fused feature, obtained by fusing the channel features and the spatial features, is input into the position detection module and the state detection module of the goggle recognition model, respectively, and the goggle wearing state of the target object is obtained from the position detection module, the method further includes: Based on the detection results of the target object wearing goggles, the safety level of the target object wearing goggles is determined; Based on the security level and the preset violation level, the alarm information is determined; The alarm information is transmitted to the terminal in the scene where the target object is located, so that the terminal can warn the target object about the status of wearing goggles.

6. The method according to any one of claims 1 to 5, characterized in that, The acquisition of the image to be detected, including the facial region of the target object, includes: Obtain the environmental parameters of the environment in which the target object is located; Based on the aforementioned environmental parameters, the interference factors for rail cutting operations are determined. Adjust the acquisition parameters of the image acquisition device based on the interference factor; Based on the acquisition parameters, the image acquisition device is controlled to acquire the image to be detected of the facial region of the target object.

7. A device for detecting the wearing of safety goggles during rail cutting operations, characterized in that, The device includes: The image acquisition module is used to acquire an image to be detected, including the facial region of the target object; The convolution module is used to input the image to be detected into multiple depth-separable convolution modules of different scales of the trained goggle recognition model to obtain the convolution features output by each depth-separable convolution module. An attention module is used to input the first fused feature, which is the result of fusing all the convolutional features, into the channel attention module and the spatial attention module of the goggle recognition model, respectively, to obtain the channel features output by the channel attention module and the spatial features output by the spatial attention module. The detection module is used to input the second fused feature, which is the result of fusing the channel features and the spatial features, into the position detection module and the state detection module of the goggle recognition model, respectively, to obtain the goggle wearing state of the target object output by the state detection module and the goggle wearing position of the target object output by the position detection module; the goggle wearing state and the goggle wearing position represent the goggle wearing detection result of the target object.

8. A system for detecting the wearing of safety goggles during rail cutting operations, characterized in that, The system includes: Image acquisition module, edge processing module, and alarm feedback module; The image acquisition module is used to acquire the image to be detected of the facial region of the target object and transmit the image to be detected to the edge processing module; The edge processing module is used to execute the goggle wearing detection method for rail cutting operations as described in any one of claims 1 to 6, so as to obtain the goggle wearing detection result of the target object, and transmit the goggle wearing detection result to the alarm feedback module; The alarm feedback module is used to issue a warning about the goggles wearing status based on the goggles wearing detection results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images