Unmanned aerial vehicle construction site safety helmet intelligent inspection method based on improved YOLOv11

Abstract

According to the invention, an unmanned aerial vehicle platform is creatively combined with the improved YOLOv11, and the unmanned aerial vehicle construction site safety helmet intelligent inspection method based on the improved YOLOv11 is provided. The method comprises the following steps: firstly, carrying out digital planning on an inspection route, and generating an executable flight path file; the unmanned aerial vehicle carries high-definition camera equipment, and carries out omnibearing video acquisition on a construction area according to a preset route; and the ground control center receives the real-time video stream data and inputs the real-time video stream data into a specially-designed safety helmet identification model to complete intelligent analysis of the state of personnel safety equipment and record of violation events. The detection model is subjected to triple optimization on the basis of a YOLOv11 architecture: a C3k2SPD module is adopted to replace a C3k2 module in an original backbone network, an ASPP module is introduced to replace an original SPPF module, and meanwhile, an NWD loss function is fused at a detection head part, so that the small target recognition capability and the complex scene fitness are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and particularly relates to an unmanned aerial vehicle construction site safety helmet intelligent inspection method based on an improved YOLOv11. BACKGROUND

[0002] As a necessary personal protective equipment in the construction site scene, the correct wearing of safety helmets is one of the key measures to protect personnel safety. Traditionally, the supervision of safety helmet wearing mainly relies on the manual on-site inspection of safety officers. This method is not only inefficient and limited in coverage, but also greatly affected by personnel subjective factors and fatigue levels, making it difficult to achieve 24-hour uninterrupted effective monitoring.

[0003] With the development of technology, automatic detection technology based on video images has been introduced into this field. Initially, some scenes used fixed cameras to monitor. Although this method reduces the human burden to some extent, it has obvious limitations: the monitoring field of view is limited by the installation location and angle of the camera, and there are many blind spots; the infrastructure construction cost is high, and it lacks flexibility and cannot adapt to the dynamic changes of the work site.

[0004] In recent years, the unique aerial mobility and wide field of view of unmanned aerial vehicle technology have provided a new solution for safety inspection. Unmanned aerial vehicles equipped with cameras can flexibly fly to any area that needs to be detected, capturing on-site images from the best angle, completely eliminating monitoring blind spots, and significantly improving inspection efficiency. However, the core of converting the massive video data collected by the unmanned aerial vehicle into effective safety warnings relies on efficient and accurate target detection algorithms.

[0005] Currently, the mainstream research and application tends to use deep learning-based target detection models. Among them, the YOLO series algorithm is favored due to its good balance between speed and accuracy. However, directly applying the general YOLO model pre-trained on standard datasets to the unmanned aerial vehicle safety helmet inspection scene exposes a series of serious technical bottlenecks: the small size of the safety helmet target caused by high-altitude photography of the unmanned aerial vehicle leads to high model miss detection rate; the complex background and many interference objects in the work site easily cause false detection; the contradiction between the computational complexity of the model and the limited hardware resources of the unmanned aerial vehicle on-board platform makes it difficult to balance real-time performance and detection accuracy; at the same time, the dynamic range of target size caused by the change of height and angle of the unmanned aerial vehicle in flight makes it difficult for the general model to adaptively and effectively detect targets of different sizes. These factors together result in poor performance of existing technologies in actual applications.

[0006] To address these challenges, this invention proposes an intelligent inspection method for safety helmets on construction sites using drones based on an improved YOLOv11 model. By optimizing the YOLOv11 model, the method enhances small target detection capabilities, effectively suppresses interference from complex backgrounds, ensures high-speed inference efficiency on edge computing devices, and exhibits good multi-scale adaptability, thereby truly meeting the practical requirements for automated intelligent inspection of safety helmets using drones. Summary of the Invention

[0007] This invention aims to overcome the shortcomings of existing construction site drone safety helmet recognition accuracy and proposes an intelligent inspection method for construction site safety helmets based on improved YOLOv11. This method improves the detection accuracy of small targets and adaptability to complex environments through deep learning model structure optimization and algorithm innovation, while ensuring inspection efficiency and system reliability, providing an efficient and accurate intelligent solution for construction site safety supervision.

[0008] A method for intelligent inspection of safety helmets at construction sites using drones based on an improved YOLOv11 includes the following steps:

[0009] Step 1: The drone initializes the safety helmet inspection task, loads the preset inspection route and takes off automatically, flying to the predetermined inspection area above the construction site according to the planned path;

[0010] Step 2: During the flight, the drone continuously collects video data of the construction site through a high-definition gimbal camera and transmits the video stream to the ground control center in real time through a wireless transmission system.

[0011] Step 3: After receiving the video stream, the ground control center inputs it into the pre-trained helmet detection model to identify people in the video frames in real time and determine their helmet wearing status, and automatically identify and mark behaviors of not wearing helmets.

[0012] Step 4: The detection results are displayed in real time on the ground monitoring interface, including the recognition box, category label and confidence level; the system automatically records the time, location and image evidence of the violation, and generates an inspection report which is uploaded to the cloud storage platform;

[0013] Step 5: After completing the full-area inspection, the drone will automatically return to base and land, ending the inspection operation.

[0014] The inspection task proposed in step 1 is constructed as follows:

[0015] Step 1.1: Import the 3D digital model or high-precision map of the construction site into the flight control software to identify key detection areas;

[0016] Step 1.2: Based on the visual coverage requirements of safety helmet inspection, patrol points are set up above the construction site in a grid pattern or around a path. The spacing between the points is dynamically set according to the camera's field of view and recognition accuracy requirements, and is kept at 5-10 meters to ensure that there are no blind spots for the detection of personnel on the ground.

[0017] Step 1.3: Set the takeoff point, landing point and hovering detection node, and set the video acquisition resolution to at least 1080p and the frame rate to no less than 30fps to ensure image clarity and recognition stability under personnel movement.

[0018] Step 1.4: Generate a complete inspection route using flight control software, perform collision detection and coverage integrity verification, and finally export the KML format route file for the UAV to load.

[0019] Furthermore, the key detection areas in step 1.3 include: construction work surfaces, high-altitude operating platforms, material storage areas, entrances and exits, passageways, and areas around temporary facilities where personnel activities are frequent.

[0020] The safety helmet detection model proposed in step 3 is constructed as follows:

[0021] Step 2.1: Construct a safety helmet dataset. Use a DJI Matrice 300RTK drone equipped with a Zenmuse H20T gimbal camera to collect images of construction workers at a construction site. Use the LabelImg tool to label the safety helmet targets in PASCAL VOC format. The labeling information includes two types of targets: people wearing safety helmets and people not wearing safety helmets. The training set, validation set, and test set are divided in an 8:1:1 ratio.

[0022] Step 2.2: Construct an improved YOLOv11 network, such as... Figure 1 As shown, the network is based on YOLOv11, replacing the two C3k2 modules at the end of the sequence in its backbone network with C3k2_SPD modules as feature extraction modules, and replacing the SPPF module in the backbone network with the ASPP module; the NWD loss function is introduced into the head network.

[0023] Step 2.3: Train the helmet detection model on the helmet dataset using the improved YOLOv11 network.

[0024] Furthermore, the safety helmet dataset proposed in step 2.1 needs to cover a variety of actual work scenarios, such as typical construction scenarios like rebar processing areas, scaffolding work areas, and material stacking areas. The image acquisition angles should include vertical overhead shots of 70 to 90 degrees, oblique shots of 30 to 60 degrees, and eye-level shots of 0 to 30 degrees. The acquisition time periods must cover 6:00 to 9:00 AM, 11:00 to 2:00 PM, 4:00 to 6:00 PM, and nighttime lighting work periods.

[0025] Furthermore, the C3k2_SPD module proposed in step 2.2, such as Figure 2 As shown, its construction structure is as follows: The C3k2 module structure includes a CBS module, a Split module, n C3 modules, a Concat module, and a CBS module connected in sequence. Each C3 module contains n Bottleneck modules stacked together. The second 3x3 convolution in the Bottleneck module is replaced by SPD to form the C3k2_SPD module.

[0026] SPD scales the input feature map X∈R by an operation with a scaling factor s. H×W×C Transform into output feature map The calculation formula is as follows:

[0027] Yx,y,c·s2+i·s+j=Xs·x+i,s·y+j,c

[0028] Where (x, y) are the spatial coordinates of the output feature map y, c is the channel index of the input feature map, and (i, j) is the internal offset within the s×s local window. This operation rearranges all pixels in each s×s local spatial block along the channel dimension, thereby achieving downsampling without information loss.

[0029] Furthermore, the ASPP module proposed in step 2.2, such as Figure 3 As shown, its construction structure is as follows: ASPP consists of two parallel parts. The first part performs dimensionality reduction through a 1×1 convolution and constructs a pooling pyramid by stacking the corresponding 3×3 spatial convolutional layers together to extract features at different scales. The second part first applies global average pooling to simplify the feature map into a numerical value that captures the global contextual information of the image. Then, it connects the output of a 1×1 convolutional layer with a kernel of 256 to output a single mean feature, which is then upsampled back to the size of the original feature map through bilinear interpolation. Finally, the outputs of the two zero convolutional layers and the global pooling layer are connected, and a 1×1 convolutional layer is used to fuse the compressed multi-scale features to generate the final output feature map.

[0030] The SPPF module extracts features through multi-scale feature fusion, while ASPP can capture information at different scales through zero convolutions with different sampling rates and through global context fusion, which can more efficiently handle helmet detection tasks under different distances and background conditions. Especially when accurately locating small targets, this improvement enables the network to achieve higher accuracy and robustness when handling helmet detection in diverse and complex field environments.

[0031] Furthermore, the NWD loss function proposed in step 2.2 encodes the center position and scale information of the bounding box in the form of a two-dimensional Gaussian distribution, thereby uniformly considering the differences in the position and size of the target. The NWD loss function captures the spatial distribution deviation under the fuzzy boundary by calculating the second-order Wasserstein distance between the predicted box and the ground truth box, thus making up for the lack of ability of traditional IoU-type loss functions to perceive uncertain regions. The formula for calculating the second-order Wasserstein distance is as follows:

[0032] Where is the mean, and represents the center positions of the predicted box and the ground truth box, respectively;

[0033] Where: [c x ] p ,[c y ] p and [c x ] g ,[c y ] g , where w is the mean, representing the center positions of the predicted and ground truth bounding boxes, respectively; p ,h p ,w g ,h g The width and height of the bounding box and the actual bounding box are given.

[0034] To improve the numerical stability of this distance metric and transform it into an effective similarity metric, the NWD loss function adopts an exponentially normalized form, and its calculation formula is as follows:

[0035]

[0036] Here, C is a constant related to the dataset, usually taken as the mean of the distribution of target sizes (such as width and height). The introduction of this constant helps to balance the weight differences of targets of different scales in the distance metric.

[0037] The final formula for calculating the NWD loss function is as follows:

[0038] L_NWD = 1 - NWD(N_p, N_g)

[0039] The final localization regression loss function is constructed from the weighted sum of the CIoU loss function and the NWD loss function, and its calculation formula is as follows:

[0040] L_reg=β·L_CIoU+(1-β)·L_NWD

[0041] Where β∈[0,1] represents the weighting coefficients of the CIoU loss function and the NWD loss function, used to balance the contributions of traditional geometric metrics and distribution metrics.

[0042] The detection results proposed in step 4 are displayed in real time, including the following steps:

[0043] Step 4.1: The ground monitoring system continuously receives the H.264 encoded video stream and detection metadata transmitted back by the UAV via a wireless transmission link;

[0044] Step 4.2: The system calls the OpenGL interface to synchronize and overlay each frame of image with the corresponding detection metadata in the video rendering area opened in the monitoring interface, and draws the recognition results on the video frame in real time according to the preset drawing specifications.

[0045] Step 4.3: Based on the recognition results, an alarm mechanism is automatically triggered to generate and store an inspection log containing time, location and image evidence. At the same time, the drone status, flight trajectory and statistical information are displayed on the integrated monitoring interface.

[0046] Furthermore, the detection metadata proposed in step 4.1 includes target bounding box coordinates, category label, confidence level, timestamp, and GPS location information.

[0047] Furthermore, the drawing specifications proposed in step 4.2 include marking personnel wearing safety helmets with green boxes and highlighting personnel not wearing them with red boxes, with the category and confidence level displayed at the top of the boxes. The beneficial effects of this invention are as follows:

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] This invention achieves comprehensive monitoring of the safe working environment of construction sites through intelligent route planning and multi-view image acquisition technology, effectively solving the problem of blind spots in traditional monitoring methods and greatly improving the coverage and inspection efficiency of safety supervision.

[0050] This invention makes targeted improvements to the YOLOv11 network structure. By introducing multi-scale feature fusion and fusion loss function, it significantly improves the detection accuracy of small-sized safety helmet targets from the perspective of UAVs. At the same time, it effectively overcomes interference factors such as complex construction environment, lighting changes and target occlusion, and greatly reduces the false detection rate and missed detection rate. Attached Figure Description

[0051] Figure 1 It is an improved YOLOv11 network architecture diagram;

[0052] Figure 2 This is the structure diagram of the C3k2_SPD module;

[0053] Figure 3 This is a structural diagram of the ASPP module. Detailed implementation method:

[0054] The specific implementation of the present invention is as follows:

[0055] A method for intelligent inspection of safety helmets at construction sites using drones based on an improved YOLOv11 includes the following steps:

[0056] Step 1: The drone initializes the safety helmet inspection task, loads the preset inspection route and takes off automatically, flying to the predetermined inspection area above the construction site according to the planned path;

[0057] Step 2: During the flight, the drone continuously collects video data of the construction site through a high-definition gimbal camera and transmits the video stream to the ground control center in real time through a wireless transmission system.

[0058] Step 3: After receiving the video stream, the ground control center inputs it into the pre-trained helmet detection model to identify people in the video frames in real time and determine their helmet wearing status, and automatically identify and mark behaviors of not wearing helmets.

[0059] Step 4: The detection results are displayed in real time on the ground monitoring interface, including the recognition box, category label and confidence level; the system automatically records the time, location and image evidence of the violation, and generates an inspection report which is uploaded to the cloud storage platform;

[0060] Step 5: After completing the full-area inspection, the drone will automatically return to base and land, ending the inspection operation.

[0061] The inspection task proposed in step 1 is constructed as follows:

[0062] Step 1.1: Import the 3D digital model or high-precision map of the construction site into the flight control software to identify key detection areas;

[0063] Step 1.2: Based on the visual coverage requirements of safety helmet inspection, patrol points are set up above the construction site in a grid pattern or around a path. The spacing between the points is dynamically set according to the camera's field of view and recognition accuracy requirements, and is kept at 5-10 meters to ensure that there are no blind spots for the detection of personnel on the ground.

[0064] Step 1.3: Set the takeoff point, landing point and hovering detection node, and set the video acquisition resolution to at least 1080p and the frame rate to no less than 30fps to ensure image clarity and recognition stability under personnel movement.

[0065] Step 1.4: Generate a complete inspection route using flight control software, perform collision detection and coverage integrity verification, and finally export the KML format route file for the UAV to load.

[0066] Furthermore, the key detection areas in step 1.3 include: construction work surfaces, high-altitude operating platforms, material storage areas, entrances and exits, passageways, and areas around temporary facilities where personnel activities are frequent.

[0067] The safety helmet detection model proposed in step 3 is constructed as follows:

[0068] Step 2.1: Construct a safety helmet dataset. Use a DJI Matrice 300RTK drone equipped with a Zenmuse H20T gimbal camera to collect images of construction workers at a construction site. Use the LabelImg tool to label the safety helmet targets in PASCAL VOC format. The labeling information includes two types of targets: people wearing safety helmets and people not wearing safety helmets. The training set, validation set, and test set are divided in an 8:1:1 ratio.

[0069] Step 2.2: Construct an improved YOLOv11 network, such as... Figure 1 As shown, the network is based on YOLOv11, replacing the two C3k2 modules at the end of the sequence in its backbone network with C3k2_SPD modules as feature extraction modules, and replacing the SPPF module in the backbone network with the ASPP module; the NWD loss function is introduced into the head network.

[0070] Step 2.3: Train the helmet detection model on the helmet dataset using the improved YOLOv11 network.

[0071] Furthermore, the safety helmet dataset proposed in step 2.1 needs to cover a variety of actual work scenarios, such as typical construction scenarios like rebar processing areas, scaffolding work areas, and material stacking areas. The image acquisition angles should include vertical overhead shots of 70 to 90 degrees, oblique shots of 30 to 60 degrees, and eye-level shots of 0 to 30 degrees. The acquisition time periods must cover 6:00 to 9:00 AM, 11:00 to 2:00 PM, 4:00 to 6:00 PM, and nighttime lighting work periods.

[0072] Furthermore, the C3k2_SPD module proposed in step 2.2, such as Figure 2 As shown, its construction structure is as follows: The C3k2 module structure includes a CBS module, a Split module, n C3 modules, a Concat module, and a CBS module connected in sequence. Each C3 module contains n Bottleneck modules stacked together. The second 3x3 convolution in the Bottleneck module is replaced by SPD to form the C3k2_SPD module.

[0073] SPD scales the input feature map X∈R by an operation with a scaling factor s. H×W×C Transform into output feature map The calculation formula is as follows:

[0074] Yx,y,c·s2+i·s+j=Xs·x+i,s·y+j,c

[0075] Where (x, y) are the spatial coordinates of the output feature map y, c is the channel index of the input feature map, and (i, j) is the internal offset within the s×s local window. This operation rearranges all pixels in each s×s local spatial block along the channel dimension, thereby achieving downsampling without information loss.

[0076] Furthermore, the ASPP module proposed in step 2.2, such as Figure 3 As shown, its construction structure is as follows: ASPP consists of two parallel parts. The first part performs dimensionality reduction through a 1×1 convolution and constructs a pooling pyramid by stacking the corresponding 3×3 spatial convolutional layers together to extract features at different scales. The second part first applies global average pooling to simplify the feature map into a numerical value that captures the global contextual information of the image. Then, it connects the output of a 1×1 convolutional layer with a kernel of 256 to output a single mean feature, which is then upsampled back to the size of the original feature map through bilinear interpolation. Finally, the outputs of the two zero convolutional layers and the global pooling layer are connected, and a 1×1 convolutional layer is used to fuse the compressed multi-scale features to generate the final output feature map.

[0077] The SPPF module extracts features through multi-scale feature fusion, while ASPP can capture information at different scales through zero convolutions with different sampling rates and through global context fusion, which can more efficiently handle helmet detection tasks under different distances and background conditions. Especially when accurately locating small targets, this improvement enables the network to achieve higher accuracy and robustness when handling helmet detection in diverse and complex field environments.

[0078] Furthermore, the NWD loss function proposed in step 2.2 encodes the center position and scale information of the bounding box in the form of a two-dimensional Gaussian distribution, thereby uniformly considering the differences in the position and size of the target. The NWD loss function captures the spatial distribution deviation under the fuzzy boundary by calculating the second-order Wasserstein distance between the predicted box and the ground truth box, thus making up for the lack of ability of traditional IoU-type loss functions to perceive uncertain regions. The formula for calculating the second-order Wasserstein distance is as follows:

[0079] Where is the mean, and represents the center positions of the predicted box and the ground truth box, respectively;

[0080] Where: [c x ] p ,[c y ] p and [c x] g ,[c y ] g , where w is the mean, representing the center positions of the predicted and ground truth bounding boxes, respectively; p ,h p ,w g ,h g The width and height of the bounding box and the actual bounding box are given.

[0081] To improve the numerical stability of this distance metric and transform it into an effective similarity metric, the NWD loss function adopts an exponentially normalized form, and its calculation formula is as follows:

[0082]

[0083] Here, C is a constant related to the dataset, usually taken as the mean of the distribution of target sizes (such as width and height). The introduction of this constant helps to balance the weight differences of targets of different scales in the distance metric.

[0084] The final formula for calculating the NWD loss function is as follows:

[0085] L_NWD = 1 - NWD(N_p, N_g)

[0086] The final localization regression loss function is constructed from the weighted sum of the CIoU loss function and the NWD loss function, and its calculation formula is as follows:

[0087] L_reg=β·L_CIoU+(1-β)·L_NWD

[0088] Where β∈[0,1] represents the weighting coefficients of the CIoU loss function and the NWD loss function, used to balance the contributions of traditional geometric metrics and distribution metrics.

[0089] The detection results proposed in step 4 are displayed in real time, including the following steps:

[0090] Step 4.1: The ground monitoring system continuously receives the H.264 encoded video stream and detection metadata transmitted back by the UAV via a wireless transmission link;

[0091] Step 4.2: The system calls the OpenGL interface to synchronize and overlay each frame of image with the corresponding detection metadata in the video rendering area opened in the monitoring interface, and draws the recognition results on the video frame in real time according to the preset drawing specifications.

[0092] Step 4.3: Based on the recognition results, an alarm mechanism is automatically triggered to generate and store an inspection log containing time, location and image evidence. At the same time, the drone status, flight trajectory and statistical information are displayed on the integrated monitoring interface.

[0093] Furthermore, the detection metadata proposed in step 4.1 includes target bounding box coordinates, category label, confidence level, timestamp, and GPS location information.

[0094] Furthermore, the drawing specifications proposed in step 4.2 include marking personnel wearing safety helmets with green boxes and highlighting personnel not wearing them with red boxes, with the category and confidence level displayed at the top of the boxes.

[0095] The above are preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions made by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent inspection of safety helmets at construction sites using unmanned aerial vehicles (UAVs) based on an improved YOLOv11, characterized in that, Includes the following steps: Step 1: The drone initializes the safety helmet inspection task, loads the preset inspection route and takes off automatically, flying to the predetermined inspection area above the construction site according to the planned path; Step 2: During the flight, the drone continuously collects video data of the construction site through a high-definition gimbal camera and transmits the video stream to the ground control center in real time through a wireless transmission system. Step 3: After receiving the video stream, the ground control center inputs it into the pre-trained helmet detection model to identify people in the video frames in real time and determine their helmet wearing status, and automatically identify and mark behaviors of not wearing helmets. Step 4: The detection results are displayed in real time on the ground monitoring interface, including the recognition box, category label and confidence level; the system automatically records the time, location and image evidence of the violation, and generates an inspection report which is uploaded to the cloud storage platform; Step 5: After completing the full-area inspection, the drone will automatically return to base and land, ending the inspection operation.

2. The intelligent inspection method for safety helmets on construction sites using unmanned aerial vehicles based on an improved YOLOv11, as described in claim 1, is characterized in that... The inspection task proposed in step 1 includes the following steps: Step 1: Import the 3D digital model or high-precision map of the construction site into the flight control software, focusing on key detection areas; Step 2: Based on the visual coverage requirements of safety helmet inspection, set up inspection points in a grid or circular pattern above the construction site. The spacing between the points is dynamically set according to the camera's field of view and recognition accuracy requirements, and is kept at 5-10 meters to ensure that there are no blind spots for the detection of personnel on the ground. Step 3: Set the takeoff point, landing point and hovering detection node, and set the video acquisition resolution to at least 1080p and the frame rate to no less than 30fps to ensure image clarity and recognition stability under personnel movement. Step 4: Generate a complete inspection route using flight control software, perform collision detection and coverage integrity verification, and finally export the KML format route file for the UAV to load.

3. The inspection task according to claim 2, characterized in that... The key inspection areas mentioned in step 2 include: construction work surfaces, high-altitude operating platforms, material storage areas, entrances and exits, passageways, and areas around temporary facilities where personnel activities are frequent.

4. The intelligent inspection method for safety helmets at construction sites using drones based on an improved YOLOv11, as described in claim 1, is characterized in that... The helmet detection model proposed in step 3 includes the following steps: Step 1: Construct a safety helmet dataset. Use a DJI Matrice 300RTK drone equipped with a Zenmuse H20T gimbal camera to collect images of construction workers at a construction site. Use the LabelImg tool to label the safety helmet targets in PASCALVOC format. The labeling information includes two types of targets: people wearing safety helmets and people not wearing them. Divide the target into training set, validation set and test set in an 8:1:1 ratio. Step 2: Construct an improved YOLOv11 network. The network is based on YOLOv11, but the C3k2 module in its backbone network is replaced with the C3k2_SPD module as the feature extraction module, and the SPPF module in the backbone network is replaced with the ASPP module. The NWD loss function is introduced into the head network. Step 3: Train the helmet detection model on the helmet dataset using the improved YOLOv11 network.

5. The safety helmet detection model according to claim 4, characterized in that, In step 2, the original YOLOv11 network backbone has four C3k2 modules. The two C3k2 modules located at the end of the sequence are replaced with C3k2_SPD modules to enhance the feature extraction capability.

6. The C3k2_SPD module according to claim 5, characterized in that... The second 3x3 convolutional layer of the Bottleneck module in the C3 module is replaced with SPD to improve the model's ability to extract complex image features.

7. The safety helmet detection model according to claim 4, characterized in that, In step 2, the NWD loss function is introduced into the head network of the original YOLOv11 network, and the final localization regression loss function is constructed by weighting and summing it with the original CIoU loss function to enhance the network's ability to detect safety helmets.

8. A method for intelligent inspection of safety helmets at construction sites using unmanned aerial vehicles (UAVs) based on an improved YOLOv11, as described in claim 1, is characterized in that... The detection results proposed in step 4 are displayed in real time, including the following steps: Step 1: The ground monitoring system continuously receives the H.264 encoded video stream and detection metadata transmitted back by the UAV via a wireless transmission link; Step 2: The system calls the OpenGL interface to synchronize and overlay each frame of image with the corresponding detection metadata in the video rendering area opened in the monitoring interface, and draws the recognition results on the video frame in real time according to the preset drawing specifications. Step 3: Based on the recognition results, an alarm mechanism is automatically triggered to generate and store an inspection log containing time, location and image evidence. At the same time, the drone status, flight trajectory and statistical information are displayed on the integrated monitoring interface.

9. The real-time display of detection results according to claim 8, characterized in that, The metadata detected in step 1 includes: target bounding box coordinates, category label, confidence level, timestamp, and GPS location information.

10. The real-time display of detection results according to claim 8, characterized in that, The drawing specifications in step 2 include: marking personnel wearing safety helmets with green boxes, highlighting personnel not wearing safety helmets with red boxes, and displaying the category and confidence level at the top of the boxes.

Images