Construction site hazard identification method and terminal

Abstract

This application provides a method and terminal for identifying potential hazards at construction sites. The method includes: generating an inspection route based on preset inspection locations; collecting images of the target construction site at each inspection location along the inspection route using a drone; and inputting the target construction site images into a target hazard identification model to obtain hazard identification results corresponding to the target construction site images. This application achieves standardized, repeatable, and comparable collection of inspection paths by generating inspection routes based on preset inspection locations; it achieves unattended automatic inspection by collecting target construction site images at each inspection location along the inspection route using a drone, thus improving inspection efficiency; and it achieves automatic identification and structured output of construction site hazards by inputting the target construction site images into a target hazard identification model, replacing manual visual inspection and improving the efficiency and accuracy of discovering construction site safety hazards.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and terminal for identifying potential hazards at construction sites. Background Technology

[0002] During construction site inspections, the drone flight paths only support broad, extensive patrols and lack targeted inspections at fixed locations. Furthermore, the diverse perspectives captured during inspections make cross-sectional comparisons of footage from different periods difficult. Additionally, the hazard identification algorithms for construction site images use general detection models, not optimized for specific construction scenarios, and can only perform simple anomaly marking, resulting in low accuracy in identifying typical construction hazards such as missing safety helmets and bare soil coverage. After hazard identification, there is a lack of alarm, dispatch, and follow-up inspection processes to resolve the hazard, making it difficult to effectively trace and archive inspection data and meet the compliance requirements for construction site safety supervision. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and terminal for identifying potential hazards on construction sites, which can improve the efficiency and accuracy of identifying potential hazards on construction sites.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for identifying potential hazards at construction sites, the method comprising: The inspection route is generated based on the preset inspection locations; Images of the target construction site are collected by drones at each of the inspection locations along the inspection route. The target construction site image is input into the target hazard identification model to obtain the hazard identification result corresponding to the target construction site image.

[0005] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows: A construction site hazard identification terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the construction site hazard identification method.

[0006] The beneficial effects of this invention are as follows: By generating inspection routes based on preset inspection locations, standardized and repeatable collection of inspection paths is achieved, solving the problems of traditional random flight that cannot perform periodic comparisons and result in missed or incorrect inspections; by collecting images of the target construction site at various inspection locations along the inspection route using drones, unattended automatic inspection is achieved, significantly reducing manual operation costs and safety risks; by inputting the target construction site images into the target hazard identification model, the hazard identification results corresponding to the target construction site images are obtained, realizing automatic identification and structured output of construction site hazards, replacing manual visual inspection, and improving the efficiency and accuracy of hazard discovery. Attached Figure Description

[0007] Figure 1 This is a flowchart of a method for identifying potential hazards at a construction site according to an embodiment of the present invention; Figure 2 This is a system architecture diagram of a method for identifying potential hazards at construction sites according to an embodiment of the present invention; Figure 3 This is another flowchart of a method for identifying potential hazards at a construction site according to an embodiment of the present invention; Figure 4 This is a timing diagram of a method for identifying potential hazards at a construction site according to an embodiment of the present invention. Figure 5 This is a schematic diagram of a construction site hazard identification terminal according to an embodiment of the present invention. Detailed Implementation

[0008] Definitions:

[0009] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0010] In the relevant technologies, the following problems exist in the construction site hazard identification scenario: only wide-range flight path patrol is supported, there is no site-specific fixed-point planning, and precise settings such as latitude, longitude, altitude, and heading are lacking; there is a lack of standardized multi-angle shooting capabilities, the shooting angles of hazards are inconsistent, and long-term comparison cannot be achieved; there is no construction site identification model, and only simple anomaly marking can be achieved, which cannot accurately identify typical hazards on the construction site; there is no complete closed-loop management link, the inspection data cannot be traced, and the construction site safety compliance requirements are not met.

[0011] To address at least the aforementioned issues, an inspection route is first generated based on pre-defined inspection locations. Next, drones collect images of the target construction site at each inspection location along the route. Then, these images are input into a target hazard identification model to obtain the corresponding hazard identification results. This approach improves the efficiency and accuracy of hazard detection at construction sites.

[0012] The following details a method for identifying potential hazards at construction sites according to the present invention. Please refer to [link / reference]. Figure 1 The method 100 includes steps 110 to 130.

[0013] Step 110: Generate inspection routes based on preset inspection locations.

[0014] In one optional implementation, inspection locations and corresponding inspection parameters are set in a preset construction site map. These parameters include flight altitude, flight speed, flight heading, and gimbal pitch angle. An initial inspection flight path is generated based on all inspection locations and their corresponding parameters. The locations of obstacles and no-fly zone boundaries in the initial inspection flight path are identified. When the initial inspection flight path conflicts with obstacles or no-fly zones, the position or altitude of the corresponding inspection location is adjusted to obtain the final inspection flight path, ensuring that the generated inspection flight path maintains a preset safe distance from obstacles and no-fly zones.

[0015] Step 120: Collect images of the target construction site at each of the inspection locations along the inspection route using a drone.

[0016] In one optional implementation, the system establishes a two-way encrypted communication link with the UAV through a preset interface and uploads the inspection route to the UAV via the communication link. A takeoff command is issued, controlling the UAV to take off autonomously and fly sequentially to each inspection location according to the inspection route. In response to the UAV arriving at the current inspection location, the system performs waypoint actions according to the inspection parameters corresponding to that location. Waypoint actions include at least one of the following: initiating photography, initiating video recording, hovering, adjusting the gimbal angle, and camera zooming, to acquire images of the target construction site. After completing the acquisition of images of all inspection locations, the UAV automatically returns and lands.

[0017] Step 130: Input the target construction site image into the target hazard identification model to obtain the hazard identification result corresponding to the target construction site image.

[0018] In one optional implementation, the target construction site image transmitted back by the UAV is preprocessed by adjusting its resolution to 640×640 pixels and normalizing the pixel values ​​to the 0-1 range. The mean is subtracted and the result is divided by the standard deviation to obtain the preprocessed construction site image. The preprocessed target construction site image is then input into a trained target hazard identification model for forward inference calculation to obtain the structured hazard identification result output by the target hazard identification model. The hazard identification result includes the hazard category, hazard bounding box coordinates, identification confidence score, image number, and shooting time. The identification confidence score is compared with a preset confidence threshold. When the identification confidence score is greater than or equal to the preset confidence threshold, a valid hazard is determined to exist in the target construction site image, and the hazard identification result is output. When the identification confidence score is less than the preset confidence threshold, a valid hazard is determined to not exist, no hazard identification result is output, and an inspection record is generated and written to the inspection log.

[0019] As described above, by generating inspection routes based on preset inspection locations, standardized and repeatable collection of inspection paths is achieved, solving the problems of traditional random flight that cannot perform periodic comparisons and result in missed or incorrect inspections. By collecting images of the target construction site at various inspection locations along the inspection route using drones, unattended automatic inspection is achieved, significantly reducing the cost and safety risks of manual operation. By inputting the target construction site images into the target hazard identification model, the hazard identification results corresponding to the target construction site images are obtained, realizing the automatic identification and structured output of construction site hazards, replacing manual visual inspection, and improving the efficiency and accuracy of hazard discovery.

[0020] In one alternative implementation, step 110 includes steps 111 to 113.

[0021] Step 111: Set the inspection locations and inspection parameters for the inspection locations in the preset construction site area map.

[0022] In one optional implementation, an electronic construction site map covering the entire construction site area is imported. Several fixed inspection points are set on the construction site map, and inspection parameters are configured for each inspection point. The inspection parameters include parameters such as UAV flight altitude, gimbal shooting angle, hovering time, image shooting resolution, and video shooting frame rate.

[0023] Step 112: Generate an initial inspection route based on all the inspection locations and the inspection parameters corresponding to the inspection locations.

[0024] In one optional implementation, all inspection locations and corresponding inspection parameters are retrieved, and the waypoint and route generation interface is called to generate an initial inspection route containing flight trajectory and inspection parameters.

[0025] Step 113: Identify obstacles and no-fly zones in the initial inspection route, and generate the inspection route by avoiding the obstacles and no-fly zones.

[0026] In one optional implementation, the initial inspection route is detected based on the coordinate data of obstacles such as buildings and construction equipment in the construction site area map and the no-fly zone data, combined with the visual obstacle avoidance and radar obstacle avoidance capabilities of the UAV. The overlapping sections of the initial inspection route with obstacles and no-fly zones are identified, and the flight trajectory of the initial inspection route is finely adjusted based on the obstacles and no-fly zones to avoid all obstacles and no-fly zones, thereby obtaining the final inspection route.

[0027] As described above, setting inspection locations and corresponding inspection parameters for each patrol location in the construction site map ensures consistency in the image acquisition process and methods during UAV inspections across different inspection cycles, providing comparability for long-term comparative analysis of potential hazards. The automatic generation of initial inspection routes based on all inspection locations and their parameters improves the efficiency of initial route generation and reduces the cost of manual route planning. By identifying obstacles and no-fly zones in the initial route and automatically generating obstacle-avoidance routes, the safety risks of flight collisions and unauthorized entry into no-fly zones are effectively mitigated, ensuring UAV flight safety in complex construction site environments.

[0028] In one alternative implementation, step 120 includes steps 121 to 122.

[0029] Step 121: Establish a communication link with the drone.

[0030] In one optional implementation, a bidirectional encrypted communication link between the system and the UAV flight control system is established based on TCP (Transmission Control Protocol) / IP (Internet Protocol) by calling the communication interface of the DJI software development kit. Data such as the UAV's battery status and flight status are collected in real time through this established communication link. After confirming that the UAV equipment is in normal condition and the communication link is functioning correctly, an inspection command is issued to the UAV.

[0031] Step 122: The inspection route is sent to the UAV via the communication link to instruct the UAV to collect images of the target construction site at each inspection location along the inspection route according to the inspection parameters corresponding to the inspection location.

[0032] In one optional implementation, the inspection route is sent to the drone via a communication link. Upon receiving the inspection operation instruction, the drone takes off autonomously, hovers automatically upon reaching each inspection point along the route, and collects high-definition images or videos of the construction site according to the inspection parameters for that point. The drone then associates the images or videos with information such as the shooting time and location's latitude and longitude. After completing the image acquisition for all inspection points, the drone automatically returns to base and lands.

[0033] As described above, by establishing a communication link with the UAV and issuing inspection routes, a reliable data interaction channel is built for the UAV's automatic inspection, ensuring that control commands can be reliably issued between the ground system and the UAV, and that image data can be stably transmitted back. By issuing inspection routes carrying inspection locations and parameters to the UAV, the UAV can autonomously perform site image acquisition offline according to the inspection parameters, ensuring the consistency of shooting methods in different inspection cycles and solving the problems of inconsistent shooting points and incomparable images caused by control errors in traditional manual remote control flight.

[0034] In an optional implementation, the site hazard identification method 100 further includes steps 140 to 160.

[0035] Step 140: Construct the initial hazard identification model.

[0036] In one alternative implementation, the initial hazard identification model is built on the target detection model of Volcano Ark. Redundant convolutional layers in the original model that are not suitable for construction site scenarios are removed, and the lightweight inference requirements of UAV inspection images are adapted to construct the initial hazard identification model.

[0037] Step 150: Construct a construction site image training set, which includes construction site images and the corresponding real labels of the construction site images.

[0038] In one optional implementation, an original construction site image dataset is constructed by collecting real-time images of the construction site and publicly available smart construction site inspection images. Each original construction site image in the dataset undergoes standardized preprocessing with uniform resolution and format to obtain a construction site image. A preset annotation format is then used to annotate each construction site image with potential hazards, generating a real label containing the hazard category and bounding box coordinates. An initial construction site image dataset is constructed based on the construction site images and their corresponding real labels. This initial dataset is then divided into a training set, a validation set, and a test set according to a preset ratio to obtain a training set suitable for training the initial hazard identification model.

[0039] Step 160: Input each of the construction site images into the initial hazard identification model to obtain a predicted label, and train the model based on the difference between the predicted label and the real label to obtain the trained target hazard identification model.

[0040] In one optional implementation, construction site images from the training set are batch-input into an initial hazard identification model, which outputs hazard identification results including hazard category, target bounding box coordinates, and identification confidence. The difference between the classification loss and the regression loss between the predicted label and the true label are calculated separately, and the total loss difference is obtained by weighted fusion. Based on the stochastic gradient descent optimizer, the model parameters of the initial hazard identification model are iteratively updated by backpropagation according to the total loss difference. Combined with optimization strategies such as learning rate decay, the model is continuously trained until the preset training termination condition is met. The optimal model parameters or the final model parameters are saved as the model parameters of the target hazard identification model to obtain the trained target hazard identification model.

[0041] As described above, by constructing an initial hazard identification model, a trainable basic network architecture is built for inspection tasks in construction site scenarios. By constructing a training set of construction site images containing their corresponding real-world labels, standardized supervision signals are provided for the model to learn the mapping relationship between hazard features and real-world labels, providing a reliable training dataset for accurate model identification. Each construction site image is input into the initial hazard identification model to obtain a predicted label, and iterative training is performed based on the difference between the predicted and real labels. This allows the model parameters to be continuously optimized and updated during backpropagation, gradually approaching the optimal mapping function for construction site hazard identification. Finally, the trained target hazard identification model is obtained, enabling it to automatically output structured hazard identification results such as hazard category, hazard bounding box coordinates, and identification confidence level, replacing the traditional manual visual interpretation method and significantly improving the efficiency and accuracy of construction site hazard identification.

[0042] In one alternative implementation, step 140 includes steps 141 to 145.

[0043] Step 141: Obtain the initial backbone network, initial neck network, and initial head network from the preset target detection model.

[0044] In one alternative implementation, a pre-trained target detection model is loaded from the official open-source framework of Volcano Ark (a platform developed for detecting target objects) as a preset target detection model. The initial backbone network, initial neck network, and initial head network in the target detection model are extracted, and redundant convolutional layers in the initial backbone network, initial neck network, and initial head network that are not suitable for the construction site scenario are removed to adapt to the lightweight inference requirements of UAV inspection images.

[0045] Step 142: Add an attention mechanism to the initial backbone network to obtain the target backbone network.

[0046] In one alternative implementation, an attention mechanism is added to the initial backbone network to enhance the model's ability to extract deep features of small-sized construction site hazards such as safety helmets and bare soil, while suppressing interference from complex background noises such as building materials and construction equipment at the construction site. This improves the accuracy and efficiency of the backbone network in extracting features of construction site hazards, resulting in an optimized target backbone network.

[0047] Step 143: Set up a feature pyramid network and a path aggregation network in the initial neck network to obtain the target neck network.

[0048] In one alternative implementation, the initial neck network is structurally optimized by adopting an optimized FPN (Feature Pyramid Network) and PAN (Path Aggregation Network) structure, and the feature fusion dimension of the initial neck network is uniformly adjusted to 320 dimensions to obtain a target neck network adapted for multi-scale hazard detection at construction sites.

[0049] Step 144: Set the output layer of the initial head network to a classification branch structure and a regression branch structure to obtain the target head network.

[0050] In one optional implementation, the output layer of the initial head network is configured with a classification branch structure and a regression branch structure. The classification branch structure is used to output the probability distribution of hazard categories in the construction site image. The regression branch structure is used to output the bounding box coordinates of the hazard. Furthermore, based on the size statistics of typical hazards at actual construction sites, the anchor box parameters of the head output layer are modified, and three sets of adapted anchor boxes are re-clustered and generated. The parameters of the three sets of anchor boxes are 12×16 pixels, 24×32 pixels, and 48×64 pixels, respectively, to adapt to the preset anchor box generation rules of the target detection model.

[0051] Step 145: Use the output of the target backbone network as the input of the target neck network, and use the output of the target neck network as the input of the target head network to obtain the initial hazard identification model.

[0052] In one optional implementation, the hazard features extracted from the target backbone network are input into the target neck network to complete multi-scale feature fusion, and then the fused hazard features are input into the target head network to complete hazard classification and coordinate regression. Furthermore, to prevent overfitting of the initial hazard identification model, a batch normalization layer and a random deactivation layer are added to the initial hazard identification model. The batch normalization layer is a BN (Batch Normalization) layer, using the output of each convolutional feature extraction operation in the target backbone network as the input to the first BN layer. The output of the convolution after fusing the features of the neck network's FPN and PAN is used as the input to the second BN layer. Additionally, the output of the convolutional layer of the head network's classification or regression branch structure is used as the input to the third BN layer. The outputs of the first, second, and third BN layers are standardized feature tensors with dimensions identical to the inputs, with no changes to channels, feature map dimensions, or batch size. The random deactivation layer, or Dropout layer (a regularization technique for neural networks), processes the feature tensors output from the first, second, or third BN layers through a preset activation function before inputting them into the Dropout layer. The Dropout layer then outputs a feature tensor with randomly discarded features. For example, the feature tensor output from the Dropout layer in the target backbone network can be used as the input to the target neck network, and the feature tensor output from the Dropout layer in the target neck network can be used as the input to the target head network. The random deactivation rate of the random deactivation layer is 0.2, ultimately constructing an initial hazard identification model adapted to complex construction site scenarios.

[0053] As described above, adding an attention mechanism to the initial backbone network enables the network to adaptively enhance the feature responses of hazard-related channels and suppress interference from the complex background of the construction site, thereby improving the backbone network's ability to extract deep features of hazards. Setting up a feature pyramid network and a path aggregation network in the initial neck network constructs a dual-path fusion structure that transmits high-level semantic information from top to bottom and low-level localization information from bottom to top, effectively improving the efficiency of multi-scale feature fusion and the recognition accuracy of small-target hazards. Setting the output layer of the initial head network as a dual-branch structure with classification and regression branches allows the network to simultaneously output the probability distribution of hazard categories and the bounding box coordinate offset. Concatenating the target backbone network, target neck network, and target head network sequentially to construct the initial hazard identification model achieves structural adaptation from general target detection capabilities to construction site-specific hazard detection capabilities.

[0054] In one alternative implementation, step 150 includes steps 151 to 153.

[0055] Step 151: Obtain the original construction site image and perform standardization processing on the original construction site image. Construct an initial construction site image dataset based on the standardized original construction site image.

[0056] In one optional implementation, raw construction site images are acquired from a set of on-site photographs and a publicly available dataset. The on-site photographs cover building construction, municipal engineering, and rail transit construction sites, including sunny, cloudy, and nighttime environments, and were taken by drones according to inspection standards, totaling 8,000 images. The publicly available dataset contains images from a smart construction site safety inspection dataset, supplemented with images of potential hazards from different regions and construction stages, totaling 4,000 images. The resolution of the raw construction site images is uniformly adjusted to 640×640 pixels, and the image format is standardized to JPG, resulting in standardized raw construction site images. An initial construction site image dataset is constructed based on all standardized raw construction site images, with a total data size of 12,000 images.

[0057] Step 152: Label each construction site image in the initial construction site image dataset with the actual hazard category and the actual bounding box coordinates, and use the actual hazard category and the actual bounding box coordinates as the actual label of the construction site image.

[0058] In one optional implementation, for each construction site image in the initial construction site image dataset, the true hazard category, true bounding box coordinates, and initial confidence level are labeled according to the VOC format (a data annotation format for neural networks). True hazard categories include not wearing a safety helmet, lack of edge protection, illegal scaffolding, uncovered bare soil, and illegal electrical and open flame hazards. True bounding box coordinates include the x-coordinate of the top-left corner, the y-coordinate of the top-left corner, the x-coordinate of the bottom-right corner, and the y-coordinate of the bottom-right corner. The true hazard category and true bounding box coordinates are used as the true labels for the construction site images.

[0059] Step 153: Perform image augmentation operation on each construction site image in the initial construction site image dataset to generate the final construction site image training set.

[0060] In one optional implementation, the initial construction site image dataset is divided into an initial construction site image training set, a construction site image validation set, and a construction site image test set according to a preset partitioning ratio, for example, a partitioning ratio of 8:1:1. Image augmentation processing is performed on each construction site image in the initial construction site image training set, and the bounding box coordinates corresponding to the augmented construction site image are simultaneously corrected to generate expanded augmented image samples, i.e., the augmented construction site image training set. The initial construction site image training set and the augmented construction site image training set are then merged to obtain the final construction site image training set.

[0061] As described above, acquiring and standardizing the original construction site images ensures that all images are uniformly processed to a preset resolution and format, eliminating input specification differences caused by different shooting devices and resolutions, and ensuring the consistency and standardization of input data during model training. Each construction site image in the initial image dataset is labeled with its true hazard category and true bounding box coordinates, transforming the hidden hazard information in the images into quantifiable supervisory signals. This provides the model with a clear training objective, enabling it to accurately learn the mapping relationship between image pixel features and hazard categories and locations during training. Image augmentation is performed on each construction site image in the initial image dataset, expanding the diversity of training samples without altering the semantic content of the images. This significantly improves the model's generalization ability under different lighting conditions, shooting angles, and construction site environments, effectively reducing the risk of overfitting in the initial hazard identification model.

[0062] In one alternative implementation, step 153 includes steps 1531 to 1534.

[0063] Step 1531: Perform at least one image enhancement operation on each construction site image in the initial construction site image dataset, including random cropping, horizontal flipping, brightness adjustment, Gaussian blurring, and Gaussian noise addition, to obtain an enhanced construction site image.

[0064] In one optional implementation, for each construction site image in the initial construction site image dataset, at least one image enhancement operation is performed, including random cropping, horizontal flipping, brightness adjustment, Gaussian blurring, and Gaussian noise addition, to obtain diverse enhanced construction site images. The horizontal flipping probability is set to 0.5, the brightness adjustment range is set to ±15%, and the standard deviation of the Gaussian blur is set to 0.1 to 0.3.

[0065] Step 1532: Perform a coordinate transformation on the real bounding box coordinates in the real label of the construction site image corresponding to the image enhancement operation to obtain the real label corresponding to the enhanced construction site image.

[0066] In one optional implementation, for image enhancement operations involving spatial transformations such as random cropping and horizontal flipping, the coordinates of the hazard bounding boxes in the real labels corresponding to the construction site image are proportionally and in the same direction for coordinate conversion and correction, ensuring that the bounding box coordinates of the enhanced image and the hazard match, so as to avoid label misalignment and failure, and generating real labels corresponding to the enhanced construction site image based on the real hazard category of the construction site image corresponding to the enhanced construction site image.

[0067] Step 1533: Construct an augmented construction site image training set based on the augmented construction site image and the corresponding real label.

[0068] In one alternative implementation, all enhanced construction site images and their corresponding corrected ground truth labels are integrated to construct a training set of enhanced construction site images with rich samples and diverse scenes.

[0069] Step 1534: Merge the initial construction site image dataset and the enhanced construction site image training set to obtain the final construction site image training set.

[0070] In one alternative implementation, the initial construction site image training set and the enhanced construction site image training set are merged, and duplicate image samples are removed to obtain the final construction site image training set.

[0071] As described above, each construction site image in the initial construction site image dataset undergoes at least one enhancement operation: random cropping, horizontal flipping, brightness adjustment, Gaussian blurring, and Gaussian noise addition. By introducing a combination of geometric and photometric transformations, the different shooting angles, lighting conditions, and image noise scenarios that may be encountered during UAV inspections are simulated, enriching the distribution space of the training samples and effectively expanding the scene diversity covered by the limited initial construction site image dataset. A coordinate transformation corresponding to the image enhancement operation is performed on the real bounding box coordinates in the real labels of the construction site images, ensuring that the hazard area annotations in the enhanced image remain aligned with the transformed image pixel space. This solves the technical problem of coordinate failure after geometric transformation and guarantees the integrity and accuracy of the enhanced construction site image sample supervision signal. An enhanced construction site image training set is constructed based on the enhanced construction site images and their corresponding real labels. The initial construction site image dataset is then merged with the enhanced construction site image training set to obtain the final construction site image training set. This process retains the real distribution characteristics of the original images while introducing the diverse features of the enhanced samples. This allows the final construction site image training set to have a stronger generalization ability while covering the original scene. The diverse construction site image training set effectively reduces the risk of overfitting of the initial hazard identification model.

[0072] In one alternative implementation, step 160 includes steps 161 to 166.

[0073] Step 161: Input each of the construction site images into the initial hazard identification model for model training, and output the predicted label of the construction site image through the initial hazard identification model; the predicted label includes the predicted hazard category and the predicted bounding box coordinates.

[0074] In one optional implementation, the batch size for each training round is set to 32. Construction site images from the training set are batch-input into the initial hazard identification model according to the batch size. The initial hazard identification model extracts hazard features through the target backbone network, then inputs these features into the target neck network for feature fusion to obtain fused features. Finally, the fused features are input into the target head network to output the hazard identification result, which includes the predicted hazard category, predicted hazard bounding box coordinates, and identification confidence level. Furthermore, a transfer learning method is employed to transfer the pre-trained weights of the preset target detection model to the initial hazard identification model. The parameters of the first 10 layers of the backbone network are frozen, and only subsequent network layers adapted to the construction site scenario are trained.

[0075] Step 162: Calculate the classification loss difference between the predicted hazard category in the predicted label and the actual hazard category in the actual label using a preset first loss function.

[0076] In one optional implementation, a preset classification loss function is used as the first loss function to calculate the error between the predicted hazard category and the actual hazard category, thereby obtaining the classification loss difference.

[0077] Step 163: Calculate the regression loss difference between the predicted bounding box coordinates in the predicted labels and the true bounding box coordinates in the true labels using a preset second loss function.

[0078] In one alternative implementation, a regression loss function is used as the second loss function, for example, the cross-union loss function. The error between the predicted bounding box and the true bounding box is calculated to obtain the regression loss difference.

[0079] Step 164: Calculate the total loss difference between the predicted label and the true label by weighting the difference between the classification loss and the regression loss.

[0080] In one optional implementation, for each batch of iterative training, the weight of the classification loss difference is set to 0.3 and the weight of the regression loss difference is set to 0.7. The classification loss difference and regression loss difference are then weighted and fused to obtain the total loss difference for a single training iteration. Furthermore, for small sample categories such as open flame hazards, a loss weight coefficient of 1.5 is set to balance the training weights of samples from each category, improving the identification accuracy of rare hazard categories, and thus obtaining the total loss difference.

[0081] Step 165: Based on the total loss difference, perform backpropagation through a preset optimizer to update the model parameters of the initial hazard identification model.

[0082] In one optional implementation, a stochastic gradient descent optimizer is selected, configured with a momentum parameter of 0.9, a weight decay coefficient of 0.0005, and an initial learning rate of 0.001. Backpropagation is performed based on the calculated total loss difference to update the model parameters of the initial hazard identification model layer by layer.

[0083] Step 166: Repeat until the preset training termination condition is met to obtain the trained target hazard identification model.

[0084] In one optional implementation, steps 161 to 165 are executed repeatedly. Every 10 training epochs, the performance of the current initial hazard identification model is verified using a construction site image validation set, and the recognition accuracy and recall of the construction site image validation set are recorded. Every 20 training epochs, the learning rate is decayed to 0.1 of its original value. When the recognition accuracy of the construction site image validation set shows no improvement for 10 consecutive training epochs, or when the preset 100 training epochs are reached, training is stopped, and the current model parameters are saved as the model parameters for the trained target hazard identification model.

[0085] As described above, each construction site image is input into the initial hazard identification model for forward computation. The predicted labels output by the model convert the image pixel features into quantifiable prediction results. A preset first loss function is used to calculate the classification loss difference between the predicted hazard category and the actual hazard category, and a preset second loss function is used to calculate the regression loss difference between the predicted bounding box coordinates and the actual bounding box coordinates. The total loss difference is obtained by weighting the classification and regression loss differences. By adjusting the weight ratio of the classification and regression tasks, the model training maintains a relative balance between the two objectives of hazard category determination and hazard location. Based on the total loss difference, backpropagation is performed through a preset optimizer, progressively feeding back the error signal and updating the model parameters. In each iteration, the model optimizes the weights along the direction of loss descent, gradually approaching the optimal model parameters for construction site hazard identification. This process is repeated until a preset training termination condition is met, ensuring that training terminates promptly when the model performance reaches saturation on the validation set. Finally, the trained target hazard identification model is obtained, guaranteeing that the target hazard identification model has a high accuracy in identifying typical construction site hazards.

[0086] In an alternative implementation, steps 131 to 132 are included after step 130.

[0087] Step 131: Determine the alarm level based on the hazard category in the hazard identification results.

[0088] In one optional implementation, the hazard category and identification confidence level are obtained from the hazard identification results. When the hazard category is an open flame hazard or illegal use of electricity, the alarm level is determined to be a Level 1 alarm; when the hazard category is not wearing a safety helmet or lack of edge protection, the alarm level is determined to be a Level 2 alarm; when the hazard category is illegal scaffolding or uncovered bare soil, the alarm level is determined to be a Level 3 alarm.

[0089] Step 132: Perform the corresponding alarm operation according to the alarm level.

[0090] In one optional implementation, when the alarm level is Level 1, alarm information is simultaneously pushed to the project manager and safety officer via voice broadcast, SMS, and in-system notification; when the alarm level is Level 2, a rectification work order is generated and pushed to the safety officer via SMS and in-system notification; when the alarm level is Level 3, a rectification work order is generated and pushed to the safety officer via in-system notification. The rectification work order includes the location of the hazard, the type of hazard, the image of the hazard, the discovery time, and the rectification requirements. The rectification work order is dispatched to the responsible person, and the rectification timer is started; in response to the rectification completion information submitted by the responsible person, the drone is controlled to perform a follow-up flight mission according to the inspection location corresponding to the hazard location, and collect follow-up images; the follow-up images are input into the target hazard identification model to obtain the follow-up identification results; based on the follow-up identification results, it is determined whether the hazard has been eliminated; if so, the status of the rectification work order is updated to "rectified"; if not, the rectification work order is re-dispatched and the rectification requirements are updated.

[0091] As described above, determining alarm levels based on the hazard categories identified in the hazard identification results, and differentiating different types of hazards according to their safety risk levels, solves the technical problem of traditional methods that treat all hazards the same, leading to untimely responses to major hazards. Alarm actions are executed according to alarm levels. By configuring differentiated alarm channels and push targets for different levels, it ensures that major hazards triggering Level 1 alarms reach project managers and safety officers immediately through multiple methods such as voice broadcasts, SMS, and in-app notifications. This allows high-risk hazards to enter the handling process in the shortest possible time, improving the response efficiency and accuracy of site safety management, and reducing the risk of safety accidents escalating due to untimely or inadequate alarms. Matching push strategies are configured for general hazards triggering Level 2 alarms and minor hazards triggering Level 3 alarms, ensuring high-priority handling of emergencies and avoiding excessive interference from high-frequency irrelevant information to management personnel.

[0092] Figure 2 This is an architectural diagram illustrating a site hazard identification system according to an embodiment of the present invention. A site hazard identification method of this application is implemented based on a site hazard identification system. (Refer to...) Figure 2The system comprises a terminal hardware layer, an intelligent sensing and recognition layer, a business scheduling layer, and an application display layer. Through this four-layer architecture, the system achieves a complete closed loop from data collection at the hardware layer, AI (Artificial Intelligence) recognition at the intelligence layer, business coordination at the scheduling layer, to visualized management and control at the display layer. This ensures automated execution of inspection tasks, intelligent identification of hidden dangers, accurate alarm delivery, and full-process tracking of rectification, effectively improving the efficiency of construction site safety management.

[0093] The terminal hardware layer is configured using DJI industrial drones as the carrier, equipped with an onboard high-definition zoom camera, a high-precision GPS (Global Positioning System) positioning module, an intelligent flight control module, and a real-time image transmission module. It is responsible for performing inspection tasks, collecting construction site images, and transmitting these images back in real-time via a communication link. The intelligent perception and recognition layer is configured to receive the construction site images transmitted from the terminal hardware layer. A defect image receiving module identifies and deletes defective images, while an image preprocessing module performs resolution adjustment and pixel normalization. The images are then input into a construction site hazard identification module for inference and identification of safety hazards in the construction site images. A perception annotation and positioning module correlates the hazard boundary coordinates in the identification results with data such as the shooting location. Finally, a data storage module persistently stores the structured hazard identification results and the original construction site images.

[0094] The business scheduling layer is configured to communicate with the drone and issue control commands through the DJI SDK (Software Development Kit) interface module; the trajectory generation and obstacle avoidance module is responsible for generating inspection routes based on preset inspection locations and automatically avoiding obstacles and no-fly zones; the unified task scheduling module is responsible for the arrangement, issuance and status monitoring of inspection tasks; the maintenance point video module manages the site image acquisition tasks for each inspection location; the message alarm push module triggers tiered alarms based on the hazard identification results; and the log recording and anomaly handling module records the system operation log throughout the process and handles abnormal situations.

[0095] The application presentation layer is configured to provide users with interactive functions such as inspection task configuration, route planning, viewing of hazard identification results, export of inspection reports, and processing of rectification work orders through the smart construction site management backend, visualization screen terminal, Web (World Wide Web) management terminal, and mobile terminal, realizing the visualization of data and the entry point for business operations.

[0096] Figure 3 This is another flowchart illustrating a method for identifying potential hazards at a construction site according to an embodiment of the present invention. (Refer to...) Figure 3This embodiment includes steps 201 to 216.

[0097] Step 201: System initialization and memory card verification.

[0098] In one optional implementation, the system establishes a communication connection with the drone by calling the DJI official SDK, and establishes bidirectional encrypted communication with the drone's flight control system using the TCP / IP protocol. Device status information of the drone, including battery status and flight status, is obtained through the communication link. The status information of the drone's storage medium, including an onboard SD (Secure Digital Memory Card), is also obtained through the communication link. The remaining storage capacity and file system status of the SD card are read to determine whether the SD card meets the image storage requirements. This corresponds to step 121.

[0099] Step 202: Import the site area map and set up inspection locations.

[0100] In one optional implementation, in response to the user's import operation of the inspection location file, the inspection locations are set up in a preset construction site area map according to the latitude and longitude coordinates of the inspection locations. Alternatively, in the preset construction site area map, in response to the user's click operation on the construction site area map interface, the latitude and longitude coordinates corresponding to the clicked location are recorded as the inspection location. This corresponds to step 111.

[0101] Step 203: Configure inspection parameters.

[0102] In one optional implementation, for each inspection location, corresponding inspection parameters are configured. These parameters include flight altitude, flight speed, flight heading, gimbal pitch angle, and waypoint actions. Waypoint actions include at least one of the following: initiating photography, initiating video recording, hovering, adjusting gimbal angle, and camera zooming. This corresponds to step 111.

[0103] Step 204: Generate inspection routes.

[0104] In one optional implementation, an initial inspection route is generated by calling a route generation interface based on all inspection locations and their corresponding inspection parameters. Geographic information data of the areas traversed by the initial inspection route is acquired, and the locations of obstacles and no-fly zone boundaries within the initial route are identified. When the initial inspection route conflicts with an obstacle or no-fly zone, the route is adjusted to maintain a preset safe distance from the obstacle and no-fly zone, generating an obstacle-avoided inspection route. Alternatively, the UAV's automatic obstacle avoidance function can be used to control the UAV to automatically avoid obstacles or no-fly zones during flight. This corresponds to steps 112 and 113.

[0105] Step 205: Issue the inspection task to the drone.

[0106] In an optional implementation, the inspection task command, including the inspection route, is transmitted to the UAV via a communication link. This corresponds to step 122.

[0107] Step 206: The drone takes off autonomously and flies along the inspection route.

[0108] In one alternative implementation, the UAV responds to the inspection mission command, performs a takeoff operation, and flies sequentially to each inspection location according to the cruise parameters set in the inspection route. This corresponds to step 120.

[0109] Step 207: Determine whether to arrive at the inspection location and perform the inspection task.

[0110] In one optional implementation, the UAV acquires its current location information in real time during flight, compares this information with the location information of each inspection point along the inspection route, and determines whether it has reached the inspection point. If it has reached the inspection point, it performs the corresponding inspection task at the inspection point. The inspection task includes collecting site images at the inspection point according to corresponding inspection parameters. This corresponds to step 120.

[0111] Step 208: The drone automatically captures images of the construction site and transmits the data back.

[0112] In one optional implementation, when the drone arrives at the current inspection location, it performs site image acquisition according to the inspection parameters corresponding to that location. Site image acquisition includes initiating photo taking or video recording. During the site image acquisition process, data such as the shooting time and latitude / longitude of the captured site images are recorded. The drone transmits the acquired target site images back to the system in real time via a communication link. This corresponds to step 120.

[0113] Step 209: Identify safety hazards in construction site images.

[0114] In one optional implementation, the resolution of the returned target construction site image is adjusted to 640×640 pixels, and the pixel values ​​are normalized to obtain a preprocessed target construction site image. The preprocessed target construction site image is then input into a target hazard identification model for inference and identification. The hazard identification result output by the target hazard identification model is obtained. The hazard identification result is in JSON (a data format) format and includes the hazard category, hazard bounding box coordinates, identification confidence level, image number, and shooting time. Hazard categories include not wearing a safety helmet, lack of edge protection, illegal scaffolding erection, uncovered bare soil, illegal use of electricity, and open flame hazards. The hazard bounding box coordinates can be represented as (x1, y1, x2, y2), where (x1, y1) are the horizontal and vertical coordinates of the upper left corner of the bounding box, and (x2, y2) are the horizontal and vertical coordinates of the lower right corner of the bounding box. The identification confidence level is a floating-point number, ranging from 0 to 1. When the confidence level is ≥ 0.7, it is determined to be a valid hidden danger; when the identification confidence level is lower than 0.7, it is considered a false detection and no hidden danger is output. This corresponds to step 130.

[0115] Step 210: Determine if there are any safety hazards.

[0116] In one optional implementation, the identification confidence level output by the target hazard identification model is obtained; the identification confidence level is compared with a preset confidence threshold, which is 0.7; when the identification confidence level is greater than or equal to the preset confidence threshold, it is determined that there is a valid hazard in the target construction site image, and steps 215 and 216 are executed; when the identification confidence level is less than the preset confidence threshold, it is determined that there is no valid hazard in the target construction site image, the structured identification result is not output, and steps 211, 212, 213, and 214 are executed sequentially. This corresponds to step 130.

[0117] Step 211: Write the inspection log normally.

[0118] In one optional implementation, when it is determined that there are no valid hidden dangers in the target construction site image, the inspection record of the current inspection location is written into the inspection log; the inspection log includes the inspection time, inspection location coordinates, flight altitude, image number, and hidden danger identification result. At this time, the hidden danger identification result is marked as "no hidden dangers". This corresponds to step 130.

[0119] Step 212: Determine whether all inspection locations have been inspected.

[0120] In one optional implementation, when all inspection locations have been inspected, step 213 is executed; if there are uninspected inspection locations, the process returns to step 207, and the drone is controlled to fly to the next inspection location to continue the inspection task. This corresponds to step 120.

[0121] Step 213: The drone autonomously returns and lands.

[0122] In one optional implementation, when it is determined that all inspection locations have been inspected, the UAV performs a return-to-home operation, autonomously flying back to the takeoff point and performing a landing operation. This corresponds to step 120.

[0123] Step 214: Automatically generate inspection report.

[0124] In one optional implementation, all inspection logs and hazard identification results for this inspection task are obtained, and an inspection report is automatically generated based on the inspection logs and hazard identification results according to a preset report template.

[0125] Step 215: Issue tiered alarms and generate a hazard log.

[0126] In one optional implementation, the hazard category and identification confidence level are obtained from the hazard identification results. When the hazard category is an open flame hazard or illegal electrical use, the alarm level is determined to be a Level 1 alarm; when the hazard category is not wearing a safety helmet or lack of edge protection, the alarm level is determined to be a Level 2 alarm; when the hazard category is illegal scaffolding or uncovered bare soil, the alarm level is determined to be a Level 3 alarm. When the alarm level is Level 1, alarm information is simultaneously pushed to the project manager and safety officer via voice broadcast, SMS, and system notification; when the alarm level is Level 2, a rectification work order is generated and pushed to the safety officer via SMS and system notification; when the alarm level is Level 3, a rectification work order is generated and pushed to the safety officer via system notification. All hazard identification results are recorded in the hazard log, which includes the hazard number, discovery time, hazard location, hazard category, hazard image, and alarm level. This corresponds to step 131.

[0127] Step 216: Dispatch and archive of hazard rectification orders.

[0128] In one optional implementation, the rectification work order includes the location of the hazard, the hazard category, the hazard image, the discovery time, and the rectification requirements. The rectification work order is dispatched to the responsible party, and a rectification timer is started. In response to the responsible party's submission of rectification completion information, the drone is controlled to perform a follow-up flight mission according to the inspection location corresponding to the hazard location, collecting follow-up images. These images are then input into the target hazard identification model to obtain the follow-up identification results. Based on the follow-up identification results, it is determined whether the hazard has been eliminated. If so, the status of the rectification work order is updated to "rectified," and the rectification record is archived in the hazard ledger. If not, a new rectification work order is dispatched with updated rectification requirements. This corresponds to step 132.

[0129] Figure 4 This is a timing diagram illustrating a site hazard identification method according to an embodiment of the present invention. (Refer to...) Figure 4The interactive entities for the construction site hazard identification method include management users, business platforms, DJI SDK services, drone terminals, and AI recognition services.

[0130] The execution steps for initiating an inspection task and interacting with the communication link and equipment pre-inspection include: The management user logs into the business platform and sends a fixed-point inspection task containing preset inspection points, flight path configurations, and image acquisition parameters. The business platform initiates a communication connection establishment request to the DJI SDK service and simultaneously sends a drone equipment status verification command. The DJI SDK service performs a communication heartbeat handshake with the drone terminal and sends hardware status acquisition commands such as GPS positioning, battery level, gimbal camera, and obstacle avoidance module. The drone terminal collects all its hardware operating parameters and sends a normal equipment status message back to the DJI SDK service. The DJI SDK service feeds back the normal status of the drone terminal equipment to the business platform, completes the pre-device self-check, and confirms that it meets the conditions for automated inspection execution. The business platform sends a complete fixed-point flight path data package to the DJI SDK service. The data package contains the coordinates of all inspection locations, specific shooting and inspection parameters for each point, and cruise obstacle avoidance rules. The DJI SDK service parses the flight path data package and sends the standardized flight path to the drone terminal, which then loads and stores the flight path data locally.

[0131] After loading the flight path mission, the drone terminal takes off autonomously and cruises along the preset inspection route in a fully automatic manner. During the cruise, the following steps are continuously executed: the drone terminal collects real-time positioning coordinates, flight attitude, and remaining battery power data, and continuously pushes flight status data streams to the DJI SDK service; after receiving the drone terminal status data, the DJI SDK service forwards it to the business platform in real time, and the business platform displays the drone's flight status in real time, realizing remote full-process flight monitoring.

[0132] The execution steps for single-inspection point image acquisition, real-time transmission, and AI hazard inference interaction include: After the drone terminal arrives at the target inspection point, it automatically acquires high-definition images of the construction site according to preset parameters and binds image metadata such as shooting time and GPS latitude and longitude, continuously transmitting the images and supporting metadata back to the DJI SDK service; the DJI SDK service forwards the aerial images and image metadata to the business platform for local caching and storage; the business platform pushes the pre-processed aerial images to the AI ​​recognition service, triggering the construction site safety hazard inference and analysis process; the AI ​​recognition service calls the trained construction site-specific target hazard recognition model to detect and infer the images, and transmits the structured hazard recognition results back to the business platform, including hazard category, hazard boundary box pixel coordinates, recognition confidence, and image source metadata; after the business platform determines that the recognition result is a valid hazard, it pushes a visual pop-up alarm to the management user and marks the hazard location and associates it with the original inspection image on the front-end interface; the business platform simultaneously writes the inspection information, image data, and recognition results of this point into the system inspection log, completing the persistent retention of single-point inspection records and ensuring full-link traceability of the inspection process.

[0133] The execution steps for the completion of the full-area point inspection and the autonomous return interaction of the drone include: after the drone terminal traverses all the preset inspection locations and completes image acquisition, it autonomously performs the return landing action; after landing, it reports the completion status message of this inspection task to the DJI SDK service; the DJI SDK service forwards the task completion status to the business platform in a synchronized manner, marking the completion of the entire automated flight data acquisition process.

[0134] The steps for generating inspection reports and archiving logs include: the business platform summarizes all data from this inspection, including drone flight trajectories, all aerial images, AI hazard identification results for each location, and full-process operation logs, and automatically generates a standardized electronic inspection report; the business platform pushes the complete inspection report to management users, supporting online viewing, downloading, and tracing of historical inspection records; at the same time, the system automatically generates a standardized hazard log based on the identified hazard data, and performs tiered alarms, rectification work order assignment, rectification review, and archiving operations.

[0135] Please refer to Figure 5 The present invention also provides a construction site hazard identification terminal 500, including a memory 501, a processor 502, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the various steps of the construction site hazard identification method described above.

[0136] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for identifying potential hazards at construction sites, characterized in that, include: The inspection route is generated based on the preset inspection locations; Images of the target construction site are collected by drones at each of the inspection locations along the inspection route. The target construction site image is input into the target hazard identification model to obtain the hazard identification result corresponding to the target construction site image.

2. The method according to claim 1, characterized in that, The step of generating an inspection route based on preset inspection locations includes: Set the inspection locations and inspection parameters for the inspection locations in the preset construction site area map; An initial inspection route is generated based on all the inspection locations and the inspection parameters corresponding to the inspection locations. The inspection route is generated by identifying obstacles and no-fly zones in the initial inspection route and avoiding these obstacles and no-fly zones.

3. The method according to claim 1, characterized in that, The process of collecting target construction site images at various inspection locations along the inspection route using drones includes: Establish a communication link with the drone; The inspection route is sent to the UAV via the communication link, instructing the UAV to collect images of the target construction site at each inspection point along the inspection route according to the inspection parameters corresponding to the inspection point.

4. The method according to claim 1, characterized in that, Also includes: Construct an initial hazard identification model; Construct a construction site image training set, which includes construction site images and the corresponding real labels of the construction site images; Each of the construction site images is input into the initial hazard identification model to obtain a predicted label. The model is then trained based on the difference between the predicted label and the true label to obtain the trained target hazard identification model.

5. The method according to claim 4, characterized in that, The construction of the initial hazard identification model includes: Obtain the initial backbone network, initial neck network, and initial head network from the preset target detection model; An attention mechanism is added to the initial backbone network to obtain the target backbone network; A feature pyramid network and a path aggregation network are set in the initial neck network to obtain the target neck network; The output layer of the initial head network is set to a classification branch structure and a regression branch structure to obtain the target head network; The output of the target backbone network is used as the input of the target neck network, and the output of the target neck network is used as the input of the target head network to obtain the initial hazard identification model.

6. The method according to claim 4, characterized in that, The construction site image training set includes: Obtain original construction site images and perform standardization processing on the original construction site images. Construct an initial construction site image dataset based on the standardized original construction site images. Each construction site image in the initial construction site image dataset is labeled with a real hazard category and real bounding box coordinates, and the real hazard category and the real bounding box coordinates are used as the real label of the construction site image; Image augmentation is performed on each construction site image in the initial construction site image dataset to generate the final construction site image training set.

7. The method according to claim 6, characterized in that, The step of performing image augmentation operations on each construction site image in the initial construction site image dataset to generate the final construction site image training set includes: For each construction site image in the initial construction site image dataset, perform at least one of the following image enhancement operations: random cropping, horizontal flipping, brightness adjustment, Gaussian blurring, and Gaussian noise addition, to obtain an enhanced construction site image. Perform a coordinate transformation on the real bounding box coordinates in the real label of the construction site image that corresponds to the image enhancement operation to obtain the real label corresponding to the enhanced construction site image; An augmented construction site image training set is constructed based on the augmented construction site images and the corresponding ground truth labels. The initial construction site image dataset and the enhanced construction site image training set are merged to obtain the final construction site image training set.

8. The method according to claim 4, characterized in that, The process involves inputting each construction site image into the initial hazard identification model to obtain a predicted label, and then training the model based on the difference between the predicted label and the true label to obtain a trained target hazard identification model. Each of the construction site images is input into the initial hazard identification model for model training, and the initial hazard identification model outputs a predicted label for the construction site image; the predicted label includes the predicted hazard category and the predicted bounding box coordinates; The classification loss difference between the predicted hazard category in the predicted label and the actual hazard category in the actual label is calculated using a preset first loss function. The regression loss difference between the predicted bounding box coordinates in the predicted labels and the true bounding box coordinates in the true labels is calculated using a preset second loss function. The total loss difference between the predicted label and the true label is obtained by weighting the difference between the classification loss and the difference between the regression loss. Based on the total loss difference, backpropagation is performed through a preset optimizer to update the model parameters of the initial hazard identification model; The process is repeated until the preset training termination condition is met, resulting in the trained target hazard identification model.

9. The method according to claim 1, characterized in that, After obtaining the hazard identification result corresponding to the target construction site image, the method further includes: The alarm level is determined based on the hazard category in the hazard identification results. Perform the corresponding alarm operation according to the alarm level.

10. A hazard identification terminal for construction sites, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for identifying potential hazards at a construction site as described in any one of claims 1 to 9.

Images