A method for identifying and detecting the wearing of safety helmets in low-light environments

By constructing a helmet-wearing image detection model and utilizing a backbone feature extraction network and a multi-scale semantic auxiliary feature pooling network, the problem of low accuracy in helmet-wearing detection under low illumination conditions was solved, achieving high-precision and fast helmet-wearing recognition.

CN118968400BActive Publication Date: 2026-11-17HUNAN ACADEMY OF BUILDING RESEARCH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410963020.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2026-11-17
Estimated Expiration
2044-07-18

AI Technical Summary

Technical Problem

Existing deep learning-based helmet detection algorithms struggle to quickly and accurately identify whether construction workers are wearing helmets in low-light environments, especially under low-light conditions such as day and night, fog, and rain, resulting in missed detections and false detections.

Method used

A method for detecting helmet wearing in low-light environments is proposed. By constructing a helmet wearing image detection model, a ten-layer backbone feature extraction network, a neck network, and a detection head network are used, combined with a multi-scale semantic auxiliary feature pooling network to improve feature extraction and fusion capabilities, thereby achieving accurate detection of low-light images.

Benefits of technology

It significantly improves the accuracy and speed of helmet wearing detection in low-light environments, and can effectively identify helmet wearing status during the day and night and in adverse weather conditions. It is suitable for open-air construction sites, tunnels, caves and mines.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968400B_ABST
    Figure CN118968400B_ABST
Patent Text Reader

Abstract

This invention discloses a method for detecting and identifying safety helmet wearing in low-light environments, primarily addressing the difficulty of identifying safety helmet wearing in low-light conditions such as day and night, foggy days, and rainy days. The method involves acquiring on-site images under the low-light conditions to be detected, and inputting the acquired images into a pre-trained safety helmet wearing image detection model to obtain the detection result of the safety helmet wearing. The safety helmet wearing image detection model consists of a feature extraction network, a neck network, and a detection head network. This invention improves the detection accuracy of safety helmet wearing in low-light environments and can be applied to the automatic detection of safety helmet wearing in open-air construction sites, tunnels, caves, and mines with poor lighting conditions, both day and night, and in inclement weather.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a method for detecting images of safety helmets worn, specifically a method for recognizing and detecting safety helmet wear in low-light environments. Background Technology

[0002] Wearing safety helmets is a crucial safety measure for workers in high-risk industries such as construction and coal mining. Currently, many construction sites are equipped with video surveillance systems, allowing safety supervisors to monitor workers through these screens. However, the high labor intensity of 24 / 7 manual monitoring significantly reduces its efficiency over time, making it ineffective in monitoring whether workers are wearing safety helmets. To address this issue, an increasing number of security systems are employing artificial intelligence technologies such as robots and drones for inspections to overcome the shortcomings of manual monitoring.

[0003] In recent years, deep learning, as a hot topic in machine learning, has been widely applied in object detection, image recognition, and semantic segmentation. Currently, scholars both domestically and internationally are also applying deep learning technology to helmet wearing detection. Compared to traditional methods of manually extracting features, deep learning can utilize convolutional neural networks to extract higher-level and more effective features, greatly improving the accuracy and speed of helmet wearing detection. Existing deep learning-based helmet wearing detection algorithms have high accuracy, but they still experience a significant probability of missed detections and false detections in complex day and night environments. Furthermore, existing deep learning-based helmet wearing detection algorithms cannot effectively identify helmets in low-light environments. Helmet wearing detection in low-light environments during both day and night still faces the following challenges:

[0004] (1) Images obtained under low light conditions have poor visual quality. The feature information obtained through deep learning will be greatly reduced, and noise is easily generated, further reducing the visibility of the image.

[0005] (2) Images acquired under low illumination conditions such as foggy days and rainy days are similar to those at night. Therefore, the images need to be dehazed. Common dehazing methods can make the images even more blurry and often require a large computational cost, resulting in a long dehazing time.

[0006] (3) Under different monitoring perspectives, the safety helmets worn by construction workers in the images are not fully displayed or they are mutually obscured, which leads to the algorithm missing detection problems.

[0007] (4) In different images, safety helmets may appear in different sizes. There are also large, medium and small targets and construction workers wearing other hats that look similar to safety helmets, which leads to false detection problems in the algorithm.

[0008] Existing methods, under the aforementioned challenges, are unable to quickly and accurately detect helmet wearing in images or videos under low-light conditions such as day and night. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a method for detecting and identifying the wearing of safety helmets in low-light environments, which addresses the above-mentioned problems of the prior art. The present invention can solve the problem that the existing methods have low accuracy in detecting images of wearing safety helmets in low-light environments, both day and night.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0011] A method for detecting helmet wearing in low-light environments includes acquiring on-site images under low-light conditions, inputting the acquired images into a pre-trained helmet wearing image detection model to obtain a detection result of whether the person being detected is wearing a helmet. The step of constructing the helmet wearing image detection model includes:

[0012] A backbone feature extraction network is used to extract features at different levels from the input helmet-wearing image. The backbone network has a ten-layer structure: the first layer is a Focus module; the second, fourth, sixth, and eighth layers are Conv_BN_SiLU; the third, fifth, seventh, and ninth layers are C3 modules; and the tenth layer is a multi-scale semantic auxiliary feature pooling network (MSAFPNet). The Conv_BN_SiLU module consists of cascaded 2D convolutional layers Conv, BatchNorm2d, and the activation function SiLU. The C3 module is composed of n CSP Bottlenecks and 3 Convs. Bottleneck is composed of the main branch and the residual branch. The main branch consists of a cascaded Conv of size 1×1 and a Conv of size 3×3. The feature maps extracted by the third, fifth, seventh and ninth layers of the backbone feature extraction network are denoted as F2, F3, F4 and F5, respectively. The input of the multi-scale semantic auxiliary feature pooling network MSAFPNet is the feature maps F2, F3, F4 and F5 extracted by the backbone feature extraction network, and the output of MSAFPNet is feature map F6. MSAFPNet includes a low-level feature mixing processing module and a semantic feature enhancement module.

[0013] The Neck network is used to fully fuse the feature maps at three scales (F3, F4, and F6) extracted by the backbone feature extraction network to obtain feature maps P3, P4, and P5.

[0014] The detection head network (Head) is used to detect features P3, P4, and P5 to obtain the detection results of the helmet-wearing image.

[0015] Optionally, the multi-scale semantic auxiliary feature pooling network MSAFPNet includes a low-level feature fusion processing module LFMPM and a semantic feature enhancement module SFEM. The input to the low-level feature fusion processing module is feature maps F2, F3, and F4, and the input to the semantic feature enhancement module is feature map F5. The low-level feature fusion processing module contains three data processing branches. The input feature map F2 passes through the first data processing branch of the low-level feature fusion processing module. This branch consists of a cascaded 1×7 spatially separable convolution, a 7×1 spatially separable convolution, and a 3×3 kernel with a dilation rate of 7. The input feature map F2' is obtained by passing dilated convolutions and max pooling layers. The input feature map F3 then passes through the second data processing branch of the low-level feature fusion processing module. This branch consists of cascaded 1×5 spatially separable convolutions, 5×1 spatially separable convolutions, dilated convolutions with a kernel size of 3×3 and a dilation rate of 5, and max pooling layers. The input feature map F4 then passes through the third data processing branch of the low-level feature fusion processing module. This branch consists of cascaded 1×3 spatially separable convolutions, 3×1 spatially separable convolutions, dilated convolutions with a kernel size of 3×3 and a dilation rate of 3, and max pooling layers. The feature map F4' is obtained by performing a concatenation operation on the feature maps F2', F3', and F4'. These feature maps are then processed sequentially through a 1×1 lightweight GSConv kernel and CA attention to obtain feature map f1. The semantic feature enhancement module contains three data processing branches that process the input feature map F5. The first data processing branch of the semantic feature enhancement module consists of a cascaded 1×1 lightweight GSConv kernel, a 3×3 lightweight GSConv kernel, a 1×1 lightweight GSConv kernel, and a dilated convolution with a 3×3 kernel and a dilation rate of 3, resulting in feature map F5'. The data processing branch of the semantic feature enhancement module... Path 2 consists of a cascaded adaptive pooling layer and a dilated convolution with a kernel size of 3×3 and a dilation rate of 3, resulting in feature map F5”. Path 3 of the semantic feature enhancement module consists of a cascaded 3×3 lightweight GSConv, a 1×1 lightweight GSConv, and a dilated convolution with a kernel size of 3×3 and a dilation rate of 3, resulting in feature map F5”'. F5’, F5”, and F5”' are then concatenated and passed through a 1×1 lightweight GSConv and CA attention to obtain feature map f2. Finally, f1 and f2 are concatenated and passed through a 1×1 convolution to obtain the final output.

[0016] Optionally, the detection head module network includes a first detection head YOLO_Head1, a second detection head YOLO_Head2, and a third detection head YOLO_Head3; the first detection head YOLO_Head1 detects an 80×80 feature map P3 to obtain the detection result of helmet wearing; the second detection head YOLO_Head2 detects a 40×40 feature map P4 to obtain the detection result of helmet wearing; and the third detection head YOLO_Head3 detects a 20×20 feature map P5 to obtain the detection result of helmet wearing.

[0017] Optionally, the step of training the helmet-wearing image detection model is included before inputting the helmet-wearing image into the pre-trained helmet-wearing image detection model:

[0018] S1) Load the helmet-wearing image detection model and preset training parameters;

[0019] S2) Collect helmet-wearing image samples under different illumination environments during the day and night and attach labels to construct a helmet-wearing image dataset;

[0020] S3) The helmet wearing image detection model is trained using preset training parameters, and the trained helmet wearing image detection model is obtained and the best training weights are saved.

[0021] S4) Input the test image into the trained detection model to obtain the detection result.

[0022] Optionally, the preset training parameters are: an initial learning rate of 0.01, a standard SGD optimizer training model with a momentum of 0.937, and weight decay always set to 5 × 10⁻⁶. -4 A total of 400 batches were trained.

[0023] Furthermore, the present invention also provides an electronic device including a microprocessor and a memory interconnected thereto, the microprocessor being programmed or configured to perform the steps of the helmet wearing recognition detection method in a low-light environment.

[0024] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program for execution by a computer device to implement the steps of the helmet wearing recognition and detection method in a low-light environment.

[0025] Compared with the prior art, the present invention has the following advantages: The present invention can solve the problem that the existing methods have low accuracy in detecting safety helmet wearing images in low light conditions such as day and night, rain and fog. It can be applied to the automatic detection of safety helmet wearing in open-air construction sites, tunnels, caves and mines with poor lighting conditions, day and night and in bad weather. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0027] Figure 2 This is a schematic diagram of the network structure of the helmet wearing image detection model in an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of the structure of the multi-scale semantic auxiliary feature pooling network in an embodiment of the present invention. Detailed Implementation

[0029] This example provides a method for detecting and recognizing the wearing of a safety helmet in low-light environments. The process is as follows: Figure 1 As shown, the process includes acquiring images of people wearing helmets, constructing a helmet-wearing image detection model, pre-training the helmet-wearing image detection model, and inputting the helmet-wearing images into the pre-trained model to obtain the detection result of whether the person in the image is wearing a helmet. The specific steps are as follows:

[0030] S1) Load the helmet wearing image detection model and preset training parameters.

[0031] S2) Collect helmet-wearing image samples and attach labels to construct a helmet-wearing image dataset. In this embodiment, the helmet-wearing image dataset includes two detection categories: wearing a helmet (hat) and not wearing a helmet (person). The dataset is then divided into a training / validation set and a test set at a 7:3 ratio, and the training / validation set is further divided into a training set and a validation set at an 8:2 ratio.

[0032] S3) The helmet-wearing image detection model is trained using preset training parameters. The training process includes the following steps: feature extraction and target location classification, loss function calculation, SGD optimizer optimization of training parameters, calculation of mAP and loss on the validation set, and judgment of whether the loss has reached the minimum. Finally, the trained helmet-wearing image detection model and the optimal training weights are obtained.

[0033] Furthermore, in this embodiment, the preset training parameters in step S1) are: an initial learning rate of 0.01, a momentum of 0.937 using the standard SGD optimizer, and a weight decay always set to 5 × 10⁻⁶. -4 A total of 400 batches were trained.

[0034] In this example, the loss function calculated in step S3) can be selected as a consistent loss function for each category as needed. The calculated loss function is: Classification Loss L cls Confidence loss L obj and positioning loss L CIoU In this example, the total loss of the detection model is defined as follows:

[0035] LOSS=λ1L cls +λ2L obj +λ3L CIoU

[0036] Where L cls The binary cross-entropy function (BCE Loss) is used to calculate the classification loss between positive samples in the predicted bounding box and the ground truth bounding box. obj We still use BCE Loss to calculate the confidence loss for all predicted bounding boxes containing the target. CIoU CIoU Loss (Complete Intersection over Union) is used to calculate the localization error between positive samples in the predicted bounding box and the ground truth bounding box.

[0037] The network structure of the helmet-wearing image detection model is as follows: Figure 2 As shown, it includes:

[0038] Backone, a backbone feature extraction network, is used to extract features at different levels from the input helmet-wearing image.

[0039] The neck network is used to fully fuse the feature maps extracted by the backbone feature extraction network, and the resulting feature maps are fed into the detection head network.

[0040] The detection head network is used to detect features from the feature map after multi-scale feature aggregation to obtain the detection results of the helmet wearing image.

[0041] The backbone feature extraction network is used to achieve feature extraction at multiple levels. As an optional implementation, such as... Figure 2As shown, in this embodiment, the backbone feature extraction network is a ten-layer structure. The first layer is a Focus module, the second, fourth, sixth, and eighth layers are Conv_BN_SiLU modules, the third, fifth, seventh, and ninth layers are C3 modules, and the tenth layer is a multi-scale semantic auxiliary feature pooling network (MSAFPNet). The feature maps obtained by the input image after passing through the third, fifth, seventh, and ninth layers of the backbone are denoted as F2, F3, F4, and F5, respectively. F2, F3, F4, and F5 are used as inputs to the multi-scale semantic auxiliary feature pooling network (MSAFPNet), and the output of MSAFPNet is feature map F6. Subsequently, F3, F4, and F6 are fed into the Neck network for further feature processing.

[0042] The multi-scale semantic auxiliary feature pooling network MSAFPNet in Backbone, such as Figure 3 As shown, the multi-scale semantic auxiliary feature pooling network MSAFPNet in this embodiment includes a low-level feature fusion processing module LFMPM and a semantic feature enhancement module SFEM. LFMPM contains three data processing branches. The input feature map F2 is used as the input to the first data processing branch of LFMPM, and the size of F2 is (160, 160, 64). The first data processing branch consists of cascaded 1×7 spatially separable convolutions. 7×1 spatially separable convolution Dilated convolution with a kernel size of 3×3 and an inflation rate of 7 and the maximum pooling layer F Maxpool The output feature F2' with size (20, 20, 64) is obtained.

[0043]

[0044] The input feature map F3 is used as the input to the second data processing branch of LFMPM. The size of F3 is (80, 80, 128). The first data processing branch consists of a series of cascaded 1×5 spatially separable convolutions. 5×1 spatially separable convolution dilated convolution with a kernel size of 3×3 and an inflation rate of 5 and the maximum pooling layer F Maxpool The output feature F3' with size (20, 20, 128) is obtained.

[0045]

[0046] The input feature map F4 is used as the input to the third data processing branch of LFMPM. The size of F4 is (40, 40, 256). The first data processing branch consists of a series of cascaded 1×3 spatially separable convolutions. 3×1 spatially separable convolution dilated convolution with a kernel size of 3×3 and an inflation rate of 3 and the maximum pooling layer F Maxpool This yields an output feature F4' with dimensions (20, 20, 256).

[0047]

[0048] Feature maps F2', F3', and F4' are concatenated and then sequentially passed through a lightweight GSConv convolution of size 1×1. and CA attention F CA The final output feature map f1 of LFMPM is obtained, and the size of f1 is (20, 20, 512).

[0049]

[0050] F5 serves as the input to the three data processing branches of SFEM, and its size is (20, 20, 512). The first data processing branch of SFEM consists of a series of cascaded 1×1 lightweight GSConv convolutions. 3×3 lightweight GSConv convolution 1×1 lightweight GSConv convolution And dilated convolution with a 3×3 kernel and an inflation rate of 3 The output feature F5' with size (20, 20, 512) is obtained.

[0051]

[0052] Branch 2 consists of sequentially cascaded adaptive pooling layers F AdaptivePool And dilated convolution with a 3×3 kernel and an inflation rate of 3 The output feature F5 with a size of (20, 20, 512) is obtained.

[0053]

[0054] Branch 3 consists of cascaded 3×3 lightweight GSConv convolutions. 1×1 lightweight GSConv convolution And dilated convolution with a 3×3 kernel and an inflation rate of 3 The output feature F5”' with size (20, 20, 512) is obtained.

[0055]

[0056] Finally, the outputs F5', F5”, and F5”' of the three branches are concatenated and then sequentially passed through a 1×1 lightweight GSConv convolution. and CA attention F CA The resulting output f2 has a size of (20, 20, 512). f1 and f2 are concatenated and then passed through a 1×1 lightweight GSConv convolution. The final output is F6(20, 20, 512).

[0057]

[0058] Finally, the three outputs P3, P4, and P5 of the neck network are fed into the detection head module network for final prediction. Figure 2 As shown, in this embodiment, the detection head module network includes a first detection head YOLO_Head1, YOLO_Head2, and YOLO_Head3. The first detection head YOLO_Head1 detects an 80×80 feature map P3 to obtain the detection result of helmet wearing; the second detection head YOLO_Head2 detects a 40×40 feature map P4 to obtain the detection result of helmet wearing; and the third detection head YOLO_Head3 detects a 20×20 feature map P5 to obtain the detection result of helmet wearing.

[0059] To validate the high-precision detection performance of the proposed method, its detection accuracy was compared with that of two-stage models (EfficientDet-D1 and Faster-rcnn), one-stage detection models (YOLOv8-s, YOLOv7-l, YOLOX-s, YOLOv5-s, ConvNeXt, CenterNet, RetinaNet, and SSD), anchorfree detection model (FCOS), and Transformer detection models (DETR and Swin). The mAP (detection accuracy) metric was used for evaluation. 50 mAP 50:95 The experimental results, including parameters, computational cost (FLOPs), and detection speed (FPS), are shown in Table 1.

[0060] Table 1 shows the experimental results of this invention compared with other algorithms on the public dataset "Safety helmet hard hat waer in detect dataset".

[0061]

[0062] In Table 1, the detection label categories in the public dataset "Safety helmet hard hat waer in detect dataset" are hat (wearing) and person (not wearing). 50(%) represents the average accuracy of the two label categories when the IoU threshold is 0.5, mAP 50:95 (%) represents the average accuracy of the two labels within the confidence interval of 0.5 to 0.95 for the IoU threshold. Detection speed (FPS) is the number of images the model can process per second when the batch size is 1. According to Table 1, the mAP of the method in this example is... 50 and mAP 50:95 The results were 95.50% and 66.20% respectively, which were the best among all methods, exceeding the second-best YOLOv7-l by 0.47% and 7.80% respectively. It also had the fastest detection speed for a single image, reaching 91.74 frames per second.

[0063] In addition, this example also provides an electronic device including a microprocessor and a memory interconnected thereto, the microprocessor being programmed or configured to perform the steps of the above-described method for detecting and recognizing the wearing of a safety helmet in a low-light environment.

[0064] Please refer to the explanation of the method above for the specific implementation process of each step.

[0065] In addition, this example also provides a computer-readable storage medium storing a computer program for execution by a computer device to implement the steps of the above-described method for detecting and recognizing the wearing of a safety helmet in a low-light environment.

[0066] Please refer to the explanation of the method above for the specific implementation process of each step.

[0067] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0068] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A safety helmet wearing recognition detection method in a low-illumination environment, characterized in that, The process includes acquiring on-site images under low-light conditions, inputting the acquired images into a pre-trained helmet-wearing image detection model to obtain the detection result of whether the person being detected is wearing a helmet. The steps for constructing the helmet-wearing image detection model include: The network consists of a backbone feature extraction network, a neck network, and a head network. The backbone network employs a ten-layer structure: the first layer is a focus module; the second, fourth, sixth, and eighth layers are Conv_BN_SiLU modules; the third, fifth, seventh, and ninth layers are C3 modules; and the tenth layer is a multi-scale semantic auxiliary feature pooling network (MSAFPNet). The Conv_BN_SiLU module comprises cascaded 2D convolutional layers (Conv, BatchNorm2d, and SiLU activation function). The C3 module utilizes n CSP (Cost-Simplified Pooling) Bottleneck layers and three Conv layers. Bottleneck is composed of a main branch and a residual branch. The main branch consists of a cascaded Conv of size 1×1 and a Conv of size 3×3. The feature maps extracted by the third, fifth, seventh and ninth layers of the backbone feature extraction network are denoted as F2, F3, F4 and F5, respectively. The input of the multi-scale semantic auxiliary feature pooling network MSAFPNet is the feature maps F2, F3, F4 and F5 extracted by the backbone feature extraction network, and the output of MSAFPNet is feature map F6. MSAFPNet includes a low-level feature fusion processing module LFMPM and a semantic feature enhancement module SFEM. The Neck network is used to fully fuse the feature maps at three scales (F3, F4, and F6) extracted by the backbone feature extraction network to obtain feature maps P3, P4, and P5. The detection head network is used to detect features P3, P4, and P5 to obtain the detection results of the helmet-wearing image; The multi-scale semantic auxiliary feature pooling network MSAFPNet includes a low-level feature fusion processing module (LFMPM) and a semantic feature enhancement module (SFEM). The input to the low-level feature fusion processing module is feature maps F2, F3, and F4, while the input to the semantic feature enhancement module is feature map F5. The low-level feature fusion processing module contains three data processing branches. The input feature map F2 passes through the first data processing branch of the low-level feature fusion processing module, which consists of a cascaded 1×7 spatially separable convolution, a 7×1 spatially separable convolution, a dilated convolution with a kernel size of 3×3 and a dilation rate of 7, and a max pooling layer to obtain feature map F2'. The input feature map F3 passes through the low-level feature fusion processing module (SFEM). The second data processing branch of the hybrid processing module consists of a cascaded 1×5 spatially separable convolution, a 5×1 spatially separable convolution, a dilated convolution with a kernel size of 3×3 and a dilation rate of 5, and a max pooling layer, resulting in feature map F3'. The input feature map F4 passes through the third data processing branch of the low-level feature hybrid processing module, which also consists of a cascaded 1×3 spatially separable convolution, a 3×1 spatially separable convolution, a dilated convolution with a kernel size of 3×3 and a dilation rate of 3, and a max pooling layer, resulting in feature map F4'. Feature maps F2', F3', and F4' are then concatenated and sequentially processed through a lightweight 1×1 GSConv and CA attention layer to obtain the final feature map. The semantic feature enhancement module comprises three data processing branches that process the input feature map F5. Branch one involves a cascaded sequence of a 1×1 lightweight GSConv, a 3×3 lightweight GSConv, a 1×1 lightweight GSConv, and a dilated convolution with a 3×3 kernel and a dilation rate of 3, yielding feature map F5'. Branch two involves a cascaded sequence of an adaptive pooling layer and a dilated convolution with a 3×3 kernel and a dilation rate of 3, yielding feature map F5''. Branch three involves a cascaded sequence of a 3×3 lightweight GSConv, a 1×1 lightweight GSConv, and a dilated convolution with a 3×3 kernel and a dilation rate of 3, yielding feature map F5'''. F5', F5'', and F5''' are then concatenated and processed sequentially through a 1×1 lightweight GSConv and CA attention to obtain the final feature map. ; feature map and The concatenation operation, followed by a 1×1 convolution, yields feature map F6.

2. The method for detecting and identifying the wearing of a safety helmet in a low-light environment according to claim 1, characterized in that, The detection head network includes a first detection head YOLO_Head1, a second detection head YOLO_Head2, and a third detection head YOLO_Head3. The first detection head YOLO_Head1 detects an 80×80 feature map P3 to obtain the detection result of helmet wearing. The second detection head YOLO_Head2 detects a 40×40 feature map P4 to obtain the detection result of helmet wearing. The third detection head YOLO_Head3 detects a 20×20 feature map P5 to obtain the detection result of helmet wearing.

3. An electronic device comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to perform the steps of the helmet wearing recognition and detection method in low-light environments as described in any one of claims 1 to 2.

4. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is executed by a computer device to implement the steps of the helmet wearing recognition and detection method in low-light environments as described in any one of claims 1 to 2.

Citation Information

Patent Citations

  • Construction site safety helmet wearing detection method based on lightweight convolutional neural network

    CN113468992A

  • Lightweight safety helmet wearing detection method based on YOLOv5

    CN115761645A