Intelligent detection method and device for wearable power safety equipment based on hierarchical path feature fusion

Abstract

The present invention discloses a method and device for intelligent detection of electric safety equipment wearing based on hierarchical path feature fusion, including: obtaining a data set of electric power workers wearing safety equipment at work, dividing the data set into a training set, a validation set, and a test set, and preprocessing the data set using a data enhancement algorithm; replacing the convolution of the Backbone layer in YOLOv8 with a HybridVGG‑Rep module; using a hierarchical path feature fusion network to fuse features, and adding a weighted fusion mechanism during fusion, while H‑PFANet uses BiFPN fusion in the intermediate feature layer of the PANet network, and adds a feature fusion structure to the bottom C2f and Conv layers; and replacing the CIoU loss function in YOLOv8 with Inner‑EIoU. The improved YOLOv8 model is trained using the training set, and the YOLOv8 model maps the input data to the output space to generate a predicted result. The present invention can efficiently and reliably perform intelligent detection on the wearing of safety equipment for electric power workers by improving YOLOv8.

Description

Technical Field

[0001] The present invention relates to the field of target detection, and in particular to an intelligent detection method for wearable power safety equipment based on hierarchical path feature fusion. Background Art

[0002] In the power industry, ensuring that workers wear safety equipment correctly is an important measure to prevent work accidents and ensure safety. At present, the detection methods for the wearing of power safety equipment mainly include manual visual inspection and automated detection technology based on simple image processing. However, these existing technologies have many shortcomings and urgently need to be improved through more advanced technologies.

[0003] First, traditional manual visual inspection relies on manual operation, which is not only inefficient, but also easily affected by subjective factors such as fatigue and distraction of operators, resulting in instability and inaccuracy of inspection results. In addition, manual inspection often requires a lot of human resources, increasing the cost of enterprise operation.

[0004] Secondly, although automated detection technologies based on image processing have improved detection efficiency, these technologies are usually limited to simple image recognition and pattern matching, making it difficult to cope with detection needs in complex environments and having a great impact on small target detection. For example, when there are large changes in lighting, complex backgrounds, or the color of safety equipment is close to that of the environment, existing image processing technologies often find it difficult to accurately identify safety equipment, resulting in high rates of false alarms and missed alarms.

[0005] Furthermore, existing automated detection systems are usually unable to provide real-time feedback and adjustments, which means that once a problem occurs during detection, the system cannot promptly alert the operator or automatically adjust the detection strategy, thereby reducing the timeliness of problem handling.

[0006] Therefore, in order to overcome the above-mentioned shortcomings, there is an urgent need for an intelligent method that can detect the wearing of power safety equipment in complex scenarios. Summary of the invention

[0007] Purpose of the invention: In view of the problems of low efficiency, low accuracy and weak ability to cope with complex environments in the current detection of electric safety equipment wearing, the present invention provides an intelligent detection method and device for electric safety equipment wearing based on hierarchical path feature fusion, by replacing the convolution module of the Backbone layer of YOLOv8 with the HybridVGG-Rep module, introducing the hierarchical path feature fusion network H-FPANet, and replacing the original CIoU loss function in YOLOv8 with the Inner-EIoU loss function, which effectively improves the accuracy and robustness of the model in complex environment detection.

[0008] Technical solution: The present invention discloses an intelligent detection method for wearing power safety equipment based on hierarchical path feature fusion, comprising the following steps:

[0009] Step 1: Obtain a data set D1 of safety equipment worn by power workers at work, divide the data set into training set, validation set, and test set, and preprocess the data set using a data enhancement algorithm to obtain a data set D2;

[0010] Step 2: Extract the input layer, two 3×3 convolutional layers and the maximum downsampling layer from the VGG network, use RepViT to replace the two 3×3 convolutions to form a HybridVGG-Rep network, and use the formed HybridVGG-Rep network structure to replace the convolution module of the Backbone layer in YOLOv8;

[0011] Step 3: In the feature fusion stage of the model, the hierarchical path feature fusion network (H-PFANet) is used to fuse the features, and a weighted fusion mechanism is introduced;

[0012] Step 4: Use the Inner-EIoU loss function to replace the CIoU loss function in YOLOv8;

[0013] Step 5: Use the training set to train the improved YOLOv8 model. The YOLOv8 model maps the input data to the output space and generates predicted results.

[0014] Step 6: Evaluate the performance of the trained model on the test set.

[0015] Furthermore, the specific method of step 1 is:

[0016] Step 1.1: Obtain videos of power workers at work through factory monitoring, cut the videos into pictures, and create a data set;

[0017] Step 1.2: Use LabelImg software to manually annotate the collected images. For the target, use a rectangular label box to mark the position of the target and mark the category of the target. The labeled images are randomly divided into three groups in proportion: training set, validation set, and test set to form data set D1.

[0018] Step 1.3: The images in the formed data set D1 are translated and scaled, inverted, and data enhanced using the MosaicAugmentation method to obtain the data set D2.

[0019] Furthermore, the specific method of step 2 is:

[0020] Step 2.1: Extract the input layer, two 3×3 convolutional layers and a maximum pooling downsampling layer from the VGG network;

[0021] Step 2.2: Replace the two 3×3 convolutional layers with the RepViT module to form the HybridVGG-Rep module;

[0022] Step 2.3: Replace the convolution module in the Backbone layer of YOLOv8 with the fused HybridVGG-Rep module.

[0023] Furthermore, in the feature fusion stage of the model, step 3 adopts a hierarchical path feature fusion network H-PFANet to fuse features, and introduces a weighted fusion mechanism, specifically:

[0024] The H-PFANet improves the PANet network structure in YOLOv8. While the intermediate feature layer in the PANet network adopts BiFPN fusion, a feature fusion structure is added to the bottom C2f and Conv layers. It includes two connection strategies. Based on the original PANet structure of the YOLOv8 model, the third C2f module from the top to the bottom of the Backbone layer is added to the BiFPN feature fusion structure to be fused with the Concat module in the Neck layer.

[0025] Then, a feature fusion structure is added between the C2f module and the Conv module in the Neck layer to fuse the second C2f module from top to bottom of the Backbone layer with the C2f module of the Neck layer after being processed by the Concat layer.

[0026] The fused feature map is input into the convolution layer, and the weighted fusion mechanism is introduced in its newly added fusion structure. The second C2f module of the Backbone layer is defined as F1, and a weight α is defined to control the contribution of F1. The C2f of the Neck layer is defined as F2, and 1-α is used to represent the contribution of F2. The weighted fusion F can be expressed as F=α·F1+(1-α)·F2.

[0027] Furthermore, the specific method of step 4 is:

[0028] Step 4.1: First calculate the upper, lower, left and right sides of the auxiliary calculation box of the annotation box:

[0029]

[0030]

[0031] in is the center x coordinate of the ground truth box, is the center y coordinate of the ground truth box, h gt is the height of the groundtruth box, w gt is the width of the ground truth box and ratio is the scaling factor.

[0032] Step 4.2: Calculate the boundary of the auxiliary calculation box of the prediction box:

[0033]

[0034]

[0035] x c is the coordinate of the center x of the prediction box, y c is the y coordinate of the center of the prediction box, w is the width of the prediction box, h is the height of the prediction box, and ratio is the scaling factor.

[0036] Step 4.3: Calculate the intersection area S-inter and pathological area S-union of the auxiliary box of the prediction box and the ground truth box respectively:

[0037]

[0038] S-union=(w gt ×h gt )×(ratio) 2 +(w×h)×(ratio) 2 -S-inter,

[0039] Step 4.4: Calculate Inner-EIoU using the intersection area S-inter and the pathological area S-union:

[0040]

[0041] Step 4.5: Calculate the YOLOv8 network bounding box regression loss function L by Inner-EIoU Inner-EIoU :

[0042]

[0043] Where ρ represents the Euclidean distance between the center point of the predicted box and the real box, w c and h c It is the width and height of the minimum enclosing rectangle formed by the prediction box and the true value box.

[0044] Furthermore, the specific method of step 5 is:

[0045] Step 5.1: Forward propagation is performed on the training set obtained by dividing the data set, and the image is input into the improved YOLOv8 model to obtain the predicted result;

[0046] Step 5.2: Use the replaced Inner-EIoU cross entropy loss to calculate the loss between the model prediction and the true label;

[0047] Step 5.3: Use the back-propagation algorithm to calculate the gradient of the loss function with respect to the model parameters;

[0048] Step 5.4: Update the model weights using the gradient descent optimization algorithm;

[0049] Step 5.5: Repeat steps 5.1-5.4 for multiple iterations until the model converges or reaches the previously predetermined training rounds.

[0050] The present invention also discloses an intelligent detection device for wearable electric safety equipment based on hierarchical path feature fusion, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the above-mentioned intelligent detection method for wearable electric safety equipment based on hierarchical path feature fusion is implemented.

[0051] Beneficial effects:

[0052] 1. The present invention focuses on improving the accuracy and robustness of small target detection in complex scenarios for intelligent detection of power workers' power safety equipment. The working scenarios of power workers are diverse and complex. The use of hierarchical path feature fusion network (H-PFANet) to fuse the extracted features can improve the accuracy and robustness of detection of power workers' safety equipment.

[0053] 2. The present invention uses an improved YOLOv8 model and replaces the convolution module of the Backbone layer in YOLOV8 with a HybridVGG-Rep module, thereby enhancing the feature information of the safety equipment worn by power workers.

[0054] 3. The present invention introduces a hierarchical path feature fusion network (H-PFANet) to fuse the extracted features. After the HybridVGG-Rep module of the Backbone layer is convolved, the feature information obtained by C2f is more accurate and complete. Then, the hierarchical path feature fusion network (H-PFANet) is used to fuse the C2f of the Backbone layer with the C2f of the Neck layer, thereby improving the accuracy of small target detection. That is, the hierarchical path feature fusion network (H-FPANet) is used to fuse the feature graph, and a weighted fusion mechanism is added, so that the model can more efficiently integrate feature information from different network layers. This method not only strengthens the use of low-level detail features, but also effectively integrates high-level semantic features, greatly improving the accuracy of the model for the detection of safety equipment worn by power workers, which is of great significance for ensuring the safety of workers and effectively reducing safety accidents caused by improper wearing of safety equipment.

[0055] 4. The present invention replaces the original CIoU loss function in YOLOv8 with the Inner-EIoU loss function. By introducing Inner-EIoU, the core concept of this method is to use an adjustable auxiliary bounding box instead of the annotation box of the actual bounding box to calculate the IoU loss. Since the annotation box of the image is fixed, the strategy of Inner-IoU is to generate an auxiliary bounding box and a prediction box, and calculate the IoU value between them based on the two. The size of the auxiliary box is adjustable, so for those crack images with different IoU sizes, Inner-IoU can automatically adjust its size, so as to more effectively perform bounding box regression and reduce the loss value. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is the overall flow chart of the present invention;

[0057] Figure 2 It is a structural block diagram of the present invention;

[0058] Figure 3 Schematic diagram of the RepViT structure. DETAILED DESCRIPTION

[0059] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0060] The present invention discloses an intelligent detection method and device for wearing of electric safety equipment based on hierarchical path feature fusion, which is suitable for the problem that the detection of wearing of safety equipment by electric workers at work is affected by the complex surrounding environment. The convolution module of the Backbone layer in YOLOv8 is replaced by the HybridVGG-Rep module to enhance the extracted feature information, and a hierarchical path feature fusion network (H-PFANet) is introduced to enhance the fusion effect during feature fusion, thereby improving the detection accuracy. The model first obtains the data set D1 of the safety equipment worn by power workers at work, divides the data set into training set, validation set, and test set, and uses the data enhancement algorithm to preprocess the data set to obtain the data set D2; then replaces the convolution module of the Backbone layer in YOLOv8 with the HybridVGG-Rep module, and inputs the feature information obtained by convolution into SPPF; then, in the feature fusion stage of the model, the hierarchical path feature fusion network (H-PFANet) is used to fuse the features of the C2f of the Backbone layer and the C2f of the Neck layer; and the CIoU loss function in YOLOv8 is replaced by the Inner-EIoU loss function; finally, the training set is used to train the improved YOLOv8 model, and the YOLOv8 model maps the input data to the output space to produce the predicted results. It specifically includes the following steps:

[0061] Step 1: Obtain a dataset D1 of safety equipment worn by power workers at work, divide the dataset into a training set, a validation set, and a test set, and use a data enhancement algorithm to preprocess the dataset to obtain a dataset D2.

[0062] Step 1.1: Obtain videos of power workers at work through factory monitoring, cut the videos into images, and create a data set.

[0063] Step 1.2: Use LabelImg software to manually annotate the collected images. For the target, use a rectangular label box to mark the position of the target and mark the category of the target. The labeled images are randomly divided into three groups in proportion: training set, validation set, and test set to form data set D1.

[0064] Step 1.3: The images in the formed data set D1 are translated and scaled, inverted, and data enhanced using the MosaicAugmentation method to obtain the data set D2.

[0065] Step 2: Extract the input layer, two 3×3 convolutional layers and the maximum downsampling layer from the VGG network, use RepViT to replace the two 3×3 convolutions to form a HybridVGG-Rep network, and use the formed HybridVGG-Rep network structure to replace the convolution module of the Backbone layer in YOLOv8. RepViT is inspired by MobileNetV3 and applies a structural reparameterization method. This method separates the token mixer and channel mixer in MobileNetV3. The original MobileNetV3 block uses 1×1 dilated convolution and 1×1 projection layer to realize the interaction between channels (i.e., channel mixer), and configures 3×3 deep convolution for the fusion of spatial information (i.e., token mixer). RepViT moves the 3×3 deep convolution upward, adds a 1×1 deep convolution, and also moves the optional squeeze and excitation (SE) layer up to after the deep convolution.

[0066] Its structure is as Figure 3 As shown:

[0067] Step 2.1: Extract the input layer, two 3×3 convolutional layers and a maximum pooling downsampling layer from the VGG network.

[0068] Step 2.2: Replace the two 3×3 convolutional layers with the RepViT module to form the HybridVGG-Rep module.

[0069] Step 2.3: Adjust the input layer of the HybridVGG-Rep module so that it can receive an input 640×640×3 feature map.

[0070] Step 2.4: Replace the convolution module in the Backbone layer of YOLOv8 with the fused HybridVGG-Rep module.

[0071] Step 2.5: Input the 640×640×3 feature map. In the HybridVGG-Rep module, after two convolutions and one maximum pooling, a 320×320×64 feature map Feature-Map1 is obtained.

[0072] Step 2.6: Adjust the input layer of the HybridVGG-Rep module, input the feature map Feature-Map1 into the HybridVGG-Rep module, and after two convolutions and one maximum pooling, obtain the 160×160×128 feature map Feature-Map2, and input the feature map Feature-Map2 into the C2f module.

[0073] Step 2.7: Adjust the input layer of the HybridVGG-Rep module, input the feature map output by the C2f module into the HybridVGG-Rep module, and after two convolutions and one maximum pooling, obtain the 80×80×256 feature map Feature-Map3, and input the feature map Feature-Map3 into the C2f module.

[0074] Step 2.8: Adjust the input layer of the HybridVGG-Rep module, input the feature map output by the C2f module into the HybridVGG-Rep module, and after two convolutions and one maximum pooling, obtain a 40×40×512 feature map Feature-Map4, which is then input into the C2f module.

[0075] Step 2.9: Adjust the input layer of the HybridVGG-Rep module, input the feature map output by the C2f module into the HybridVGG-Rep module, and after two convolutions and one maximum pooling, obtain a 20×20×512 feature map Feature-Map4, which is then input into the C2f module.

[0076] Step 2.10: Input the feature map Feature-Map4 into the C2f module. After processing, the output feature map is passed into the SPPF module.

[0077] Step 3: In the feature fusion stage of the model, a hierarchical path feature fusion network (H-PFANet) is used to fuse the features. Figure 2 As shown:

[0078] H-PFANet is improved with reference to the BiFPN network. It contains two connection strategies. While maintaining the fusion of the same-layer feature layer in BiFPN, a feature fusion structure is added to the bottom feature layer. Based on the original PANet structure of YOLOv8, the third C2f module in the Backbone layer is added to the BiFPN feature fusion structure to fuse it with the Concat module in the Neck layer. Then a feature fusion structure is added between C2f and Conv in the Neck layer. The second C2f module in the Backbone layer is fused with the C2f module in the Neck layer after the Concat layer, thereby enhancing the feature information in the feature map. The fused feature map is input into the convolution layer, and a weighted fusion mechanism is introduced in its newly added fusion structure. The second C2f module of the Backbone layer is defined as F1, and a weight α is defined to control the contribution of F1. The C2f of the Neck layer is defined as F2, and 1-α is used to represent the contribution of F2. The weighted fusion F can be expressed as F=α·F1+(1-α)·F2.

[0079] Step 4: Use the Inner-EIoU loss function to replace the CIoU loss function in YOLOv8, as follows:

[0080] Step 4.1: First calculate the upper, lower, left and right sides of the auxiliary calculation box of the annotation box:

[0081]

[0082] in is the center x coordinate of the ground truth box, is the center y coordinate of the ground truth box, h gt is the height of the groundtruth box, w gt is the width of the ground truth box and ratio is the scaling factor.

[0083] Step 4.2: Calculate the boundary of the auxiliary calculation box of the prediction box:

[0084]

[0085] x c is the coordinate of the center x of the prediction box, y c is the y coordinate of the center of the prediction box, w is the width of the prediction box, h is the height of the prediction box, and ratio is the scaling factor.

[0086] Step 4.3: Calculate the intersection area S-inter and pathological area S-union of the auxiliary box of the prediction box and the ground truth box respectively:

[0087]

[0088] S-union=(w gt ×h gt )×(ratio) 2 +(w×h)×(ratio) 2 -S-inter

[0089] Step 4.4: Calculate Inner-EIoU using the intersection area S-inter and the pathological area S-union:

[0090]

[0091] Step 4.5: Calculate the YOLOv8 network bounding box regression loss function L by Inner-EIoU Inner-EIoU :

[0092]

[0093] Where ρ represents the Euclidean distance between the center point of the predicted box and the real box, w c and h c It is the width and height of the minimum enclosing rectangle formed by the prediction box and the true value box.

[0094] Step 5: Use the training set to train the improved YOLOv8 model. The YOLOv8 model maps the input data to the output space and produces the predicted results:

[0095] Step 5.1: Forward propagation is performed on the training set obtained by dividing the data set, and the image is input into the improved YOLOv8 model to obtain the predicted result.

[0096] Step 5.2: Use the replaced Inner-EIoU cross entropy loss to calculate the loss between the model prediction and the true label.

[0097] Step 5.3: Use the back-propagation algorithm to calculate the gradient of the loss function with respect to the model parameters.

[0098] Step 5.4: Update the model weights using the gradient descent optimization algorithm.

[0099] Step 5.5: Repeat steps 5.1-5.4 for multiple iterations until the model converges or reaches the previously predetermined training rounds.

[0100] The present invention can be combined with a computer system to become an intelligent auxiliary device based on the wearing of power safety equipment. The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the above-mentioned intelligent detection method for wearing power safety equipment based on hierarchical path feature fusion is implemented.

[0101] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent detection method for wearing power safety equipment based on hierarchical path feature fusion, characterized in that: The steps include: Step 1: Obtain a data set D1 of safety equipment worn by power workers at work, divide the data set into training set, validation set, and test set, and preprocess the data set using a data enhancement algorithm to obtain a data set D2; Step 2: Extract the input layer, two 3×3 convolutional layers and the maximum downsampling layer from the VGG network, use the RepViT module to replace the two 3×3 convolutions to form a HybridVGG-Rep module, and use the formed HybridVGG-Rep module to replace the convolution module of the Backbone layer in YOLOv8; Step 2.1: Extract the input layer, two 3×3 convolutional layers and a maximum pooling downsampling layer from the VGG network; Step 2.2: Replace the two 3×3 convolutional layers with the RepViT module to form the HybridVGG-Rep module; Step 2.3: Adjust the input layer of the HybridVGG-Rep module so that it can receive the input 640×640×3 feature map; Step 2.4: Replace the convolution module in the Backbone layer of YOLOv8 with the fused HybridVGG-Rep module; Step 2.5: Input the 640×640×3 feature map. In the HybridVGG-Rep module, after two convolutions and one maximum pooling, a 320×320×64 feature map Feature-Map1 is obtained. Step 2.6: Adjust the input layer of the HybridVGG-Rep module, input the feature map Feature-Map1 into the HybridVGG-Rep module, and after two convolutions and one maximum pooling, obtain the 160×160×128 feature map Feature-Map2, and input the feature map Feature-Map2 into the C2f module; Step 2.7: Adjust the input layer of the HybridVGG-Rep module, input the feature map output by the C2f module into the HybridVGG-Rep module, and after two convolutions and one maximum pooling, obtain the 80×80×256 feature map Feature-Map3, and input the feature map Feature-Map3 into the C2f module; Step 2.8: Adjust the input layer of the HybridVGG-Rep module, input the feature map output by the C2f module into the HybridVGG-Rep module, and after two convolutions and one maximum pooling, obtain a 40×40×512 feature map Feature-Map4, which is input into the C2f module; Step 2.9: Adjust the input layer of the HybridVGG-Rep module, input the feature map output by the C2f module into the HybridVGG-Rep module, and after two convolutions and one maximum pooling, obtain a 20×20×512 feature map Feature-Map5, which is input into the C2f module; Step 2.10: Input the feature map Feature-Map5 into the C2f module. After processing, the output feature map is passed into the SPPF module. Step 3: In the feature fusion stage of the model, the hierarchical path feature fusion network H-PFANet is used to fuse the features, and a weighted fusion mechanism is introduced; Step 4: Use the Inner-EIoU loss function to replace the CIoU loss function in YOLOv8; Step 5: Use the training set to train the improved YOLOv8 model. The YOLOv8 model maps the input data to the output space and generates predicted results. Step 6: Evaluate the performance of the trained model on the test set.

2. According to claim 1, the intelligent detection method for electric power safety equipment wear based on hierarchical path feature fusion is characterized in that: The specific method of step 1 is: Step 1.1: Obtain videos of power workers at work through factory monitoring, cut the videos into pictures, and create a data set; Step 1.2: Use LabelImg software to manually annotate the collected images. For the target, use a rectangular label frame to mark the position of the target and mark the category of the target. The labeled images are randomly divided into three groups according to proportion: training set, validation set, and test set to form data set D1; Step 1.3: The images in the formed data set D1 are translated and scaled, inverted, and enhanced using the Mosaic Augmentation method to obtain the data set D2.

3. According to claim 1, the intelligent detection method for electric power safety equipment wear based on hierarchical path feature fusion is characterized in that: In the feature fusion stage of the model, step 3 uses a hierarchical path feature fusion network H-PFANet to fuse features and introduces a weighted fusion mechanism, specifically: The H-PFANet improves the PANet network structure in YOLOv8. While the intermediate feature layer in the PANet network adopts BiFPN fusion, a feature fusion structure is added to the bottom C2f and Conv layers. It includes two connection strategies. Based on the original PANet structure of the YOLOv8 model, the third C2f module from the top to the bottom of the Backbone layer is added to the BiFPN feature fusion structure to be fused with the Concat module in the Neck layer. Then, a feature fusion structure is added between the C2f module and the Conv module in the Neck layer to fuse the second C2f module from top to bottom of the Backbone layer with the C2f module of the Neck layer after being processed by the Concat layer. The fused feature map is input into the convolution layer, and the weighted fusion mechanism is introduced in its newly added fusion structure. The second C2f module of the Backbone layer is defined as F1, and a weight α is defined to control the contribution of F1. The C2f of the Neck layer is defined as F2, and 1-α is used to represent the contribution of F2. The weighted fusion F can be expressed as F=α·F1+(1-α)·F2.

4. According to the method and device for intelligent detection of electric safety equipment wear based on hierarchical path feature fusion according to claim 1, the calculation method of Inner-EIoU in step 4 is specifically as follows: Step 4.1: First calculate the upper, lower, left and right sides of the auxiliary calculation box of the annotation box: in is the center x coordinate of the ground truth box, is the center y coordinate of the ground truth box, h gt is the height of the ground truth box, w gt is the width of the ground truth box, and ratio is the scaling factor; Step 4.2: Calculate the boundary of the auxiliary calculation box of the prediction box: in, x c is the coordinate of the center x of the prediction box, y c is the y coordinate of the center of the prediction box, w is the width of the prediction box, h is the height of the prediction box, and ratio is the scaling factor; Step 4.3: Calculate the intersection area S-inter and pathological area S-union of the auxiliary box of the prediction box and the ground truth box respectively: S-union=(w gt ×h gt )×(ratio) 2 +(w×h)×(ratio) 2 -S-inter, Step 4.4: Calculate Inner-EIoU using the intersection area S-inter and the pathological area S-union: Step 4.5: Calculate the YOLOv8 network bounding box regression loss function L by Inner-EIoU Inner-EIoU : Where ρ represents the Euclidean distance between the center point of the predicted box and the real box, w c and h c It is the width and height of the minimum enclosing rectangle formed by the prediction box and the true value box.

5. According to claim 1, the intelligent detection method for electric power safety equipment wear based on hierarchical path feature fusion is characterized in that: The specific method of step 5 is: Step 5.1: Forward propagation is performed on the training set obtained by dividing the data set, and the image is input into the improved YOLOv8 model to obtain the predicted result; Step 5.2: Use the replaced Inner-EIoU cross entropy loss to calculate the loss between the model prediction and the true label; Step 5.3: Use the back-propagation algorithm to calculate the gradient of the loss function with respect to the model parameters; Step 5.4: Update the model weights using the gradient descent optimization algorithm; Step 5.5: Repeat steps 5.1-5.4 for multiple iterations until the model converges or reaches the previously predetermined training rounds.

6. A wearable intelligent detection device for power safety equipment based on hierarchical path feature fusion, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the intelligent detection method for wearable electric power safety equipment based on hierarchical path feature fusion as described in any one of claims 1 to 5 is implemented.

Images