Bridge construction dynamic three-dimensional risk area monitoring method and system based on monocular 3D detection algorithm

Abstract

The invention discloses a bridge construction dynamic three-dimensional risk area monitoring method and system based on a monocular 3D detection algorithm. The method comprises the steps that a video image of a construction area is collected through a monocular camera; utilizing the trained 3D detection model to identify the three-dimensional space positions and sizes of the construction machinery and the operating personnel; performing optical flow detection on the video image to determine the motion state of the construction machinery; generating and adjusting a dynamic three-dimensional risk area capable of moving along with the construction machinery around the construction machinery based on a detection result; judging whether the personnel are authorized operating personnel or not through comparison between face recognition and an identity library; the position of an operator is monitored, and when the construction machinery runs, a person not operating a driver enters the three-dimensional risk area to trigger an alarm; and when the construction machinery is static, unauthorized team operators enter the three-dimensional risk area to trigger an alarm. The method has the advantages that reliable technical guarantee can be provided for construction site safety management, and the real-time performance, precision and adaptability of construction safety monitoring are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of highway construction risk monitoring, and in particular to a method and system for monitoring dynamic three-dimensional risk areas in bridge construction based on a monocular 3D detection algorithm. Background Art

[0002] Bridge construction, especially large-scale cross-sea bridge projects, faces problems such as high construction difficulty, complex environment, and small site. In order to build the main bridge, it is usually necessary to build an offshore trestle as a passage and construction platform, which makes the construction platform space small. There are a large number of mechanical equipment and workers on the temporary offshore trestle and bridge deck construction site, and they move frequently, which is easy to cause spatial intersection, resulting in a significant increase in safety risks.

[0003] In the safety monitoring of bridge construction areas, traditional monitoring methods mainly rely on two-dimensional image detection technology or manual monitoring by personnel, including video monitoring, sensor detection and GPS positioning. These methods have the following shortcomings:

[0004] 1. Limited ability to judge spatial position:

[0005] Since two-dimensional image detection can only obtain planar information and cannot identify the precise positional relationship between machinery and equipment and operators in three-dimensional space, it is difficult to accurately judge the relative distance between machinery and personnel, and cannot effectively monitor potential spatial cross-risks.

[0006] 2. Inaccurate division of dynamic risk areas:

[0007] Traditional monitoring methods usually rely on manual safety distance setting and cannot dynamically adjust the risk area according to the real-time position changes of construction machinery and personnel and the changes in the movement status of construction machinery. This is prone to mismatch of risk area ranges when construction machinery moves frequently, increasing safety hazards.

[0008] 3. Insufficient real-time performance and adaptability:

[0009] In a complex environment such as a bridge construction site, machinery and personnel move frequently and their paths are not fixed. Traditional monitoring technology cannot track and identify the specific locations of all machinery and personnel in real time and flexibly, making it difficult to provide accurate risk warning information.

[0010] 4. High hardware cost:

[0011] Although some existing technologies (such as lidar) can provide three-dimensional spatial information, their equipment costs are high and their installation is complex, making them unsuitable for large-scale applications in high-density and dynamic construction scenarios. Summary of the invention

[0012] The purpose of the present invention is to provide a method and system for monitoring dynamic three-dimensional risk areas in bridge construction based on a monocular 3D detection algorithm, thereby solving the aforementioned problems existing in the prior art.

[0013] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0014] A method for monitoring dynamic three-dimensional risk areas of bridge construction based on a monocular 3D detection algorithm comprises the following steps:

[0015] S1. Image acquisition: real-time video images of the construction area are collected through a monocular camera;

[0016] S2. Image processing and 3D detection: Use the trained 3D detection model to identify the three-dimensional spatial position and size of construction machinery and workers at the construction site;

[0017] S3, optical flow detection: perform optical flow detection on the video image to analyze the motion state of the construction machinery;

[0018] S4. Risk area generation: Based on the detection results of the 3D detection model and the optical flow detection results, dynamically generate and adjust the three-dimensional risk area around the construction machinery as the construction machinery moves;

[0019] S5. Identity recognition: Through face recognition and identity database comparison, the identity of the person entering the risk area is identified in real time to confirm whether he is an authorized operator of the construction machinery operation team;

[0020] S6. Early warning: Real-time monitoring of the position of workers. When the construction machinery is running, if a person other than the operating driver enters the three-dimensional risk area, an alarm will be triggered; when the construction machinery is stationary, if an unauthorized team worker enters the three-dimensional risk area, an alarm will be triggered.

[0021] Preferably, step S1 specifically includes the following contents:

[0022] S11. Select a monocular camera with high resolution and wide viewing angle to ensure coverage of the main construction areas;

[0023] S12, install at a high position to avoid obstruction and ensure that the camera can monitor key construction locations to the maximum extent;

[0024] S13, the monocular camera collects video images of the construction area in real time.

[0025] Preferably, the 3D detection model is built using the MMDetection3D framework; the training process of the 3D model is:

[0026] S21. Collect image data of various construction machines and workers, annotate their 3D positions and sizes, and construct a training data set of construction scenes;

[0027] S22, using ResNet101 as the backbone, first perform multi-level feature extraction on the image data, and then pass these features to the 3D detection head to complete the specific task; the introduction of DCN enables the backbone to more flexibly adapt to the geometric deformation of the target in the image, thereby enhancing the effect of feature extraction;

[0028] S23, predicting the three-dimensional information and the object category of the object in combination with the 3D detection head; the three-dimensional information includes the value of the center offset in the x direction, the value of the center offset in the y direction, the depth information, the width of the object, the length of the object, the height of the object, and the rotation angle of the object;

[0029] S24. Use supervised learning technology to train the 3D detection model so that it can predict the three-dimensional spatial position and size of construction machinery and workers from the video images obtained by the monocular camera.

[0030] Preferably, step S3 specifically includes the following contents:

[0031] S31, input two consecutive frames of images into the LiteFlowNet2 network, which estimates the optical flow information layer by layer through a cascaded convolutional neural network, estimates the pixel displacement in the entire image from coarse to fine, and outputs the optical flow field;

[0032] S32, using the 2D target detection model Faster R-CNN to detect construction machinery targets, outputting a 2D target detection bounding box, and extracting the optical flow data of the target area from the optical flow field based on the 2D target detection bounding box;

[0033] S33. Determine whether the motion state of the construction machinery is "moving" or "stationary" by calculating the average amplitude of the optical flow vector in the target object area.

[0034] Preferably, step S4 specifically includes the following contents:

[0035] S41, setting a risk area expansion factor, which is used to determine the size of the risk area;

[0036] S42, setting the value of the risk zone expansion factor according to the movement state of the construction machinery;

[0037] S43, extracting basic tensor data of the construction machinery boundary box, including position coordinates, length and width;

[0038] S44. Expand the specified tensor data using the risk area expansion factor.

[0039] Preferably, step S42 is specifically as follows: when the construction machinery is in a "moving" state, the risk zone expansion factor is set to 1, so that the size of the generated risk area is equivalent to 1 times the distance of the construction machinery boundary box; when the construction machinery is in a "stationary" state, the risk zone expansion factor is set to 0.5, so that the size of the generated risk area is reduced to 0.5 times the distance of the construction machinery boundary box.

[0040] Preferably, step S44 specifically includes proportionally adjusting only the length and width parameters of the construction machinery boundary box, and using the adjusted tensor data to generate a new three-dimensional risk area boundary box;

[0041] The length and width of the tensor data are multiplied by (1 + expansion factor) to implement the expansion operation, while the height remains unchanged.

[0042] Preferably, step S5 specifically includes the following contents:

[0043] S51, collect and register the facial features of the operating team members, generate feature codes and store them in the identity database;

[0044] S52, capturing images of people entering the risk area in real time;

[0045] S53. Use the face detection model and the face recognition model to identify and extract the face feature vector of the person's image, compare the face feature vector of the person with all the feature vectors in the identity database, and calculate the similarity between the two to determine whether the corresponding person is an authorized person.

[0046] The present invention also aims to provide a bridge construction dynamic three-dimensional risk area monitoring system based on a monocular 3D detection algorithm, comprising:

[0047] Image acquisition module: collects video images of the construction area in real time through a monocular camera;

[0048] Image processing and 3D detection module: Use the trained 3D detection model to identify the 3D spatial position and size of construction machinery and workers at the construction site;

[0049] Optical flow detection module: performs optical flow detection on video images to analyze the motion state of construction machinery;

[0050] Risk area generation: Based on the detection results of the 3D detection model and the optical flow detection results, the three-dimensional risk area around the construction machinery is dynamically generated;

[0051] Identity recognition module: Through face recognition and identity database comparison, the identity of the person entering the risk area is identified in real time to confirm whether he is an authorized operator of the construction machinery operation team;

[0052] Early warning module: real-time monitoring of the location of workers. When the construction machinery is running, if anyone other than the operator enters the three-dimensional risk area, an alarm will be triggered; when the construction machinery is stationary, if an unauthorized team worker enters the three-dimensional risk area, an alarm will be triggered.

[0053] Preferably, the system further comprises:

[0054] Management platform: Receive and display risk warning information for real-time monitoring by construction managers; view the current status of risk areas, construction machinery movement information, and identity verification results of entrants, so that managers can quickly respond to potential risks.

[0055] The beneficial effects of the present invention are as follows: 1. Enhanced three-dimensional spatial recognition capability: The present invention can extract the three-dimensional spatial position of construction machinery and operators from the two-dimensional image of a single camera through a monocular 3D detection algorithm. This technology solves the deficiency that traditional two-dimensional image detection is difficult to accurately judge spatial relationships, and provides more accurate data support for risk monitoring. 2. Flexibility of dynamic risk zone division: The present invention automatically generates and dynamically adjusts the three-dimensional risk area around the construction machinery based on the real-time position and motion state of the construction machinery, so that risk monitoring can adapt to the movement of the construction machinery and changes in the construction environment. Compared with the statically divided risk area, the dynamic risk area is more flexible and can adapt to complex construction scenes more accurately. 3. Improved real-time warning responsiveness: The present invention can monitor in real time whether the operator enters the risk area by judging the relative position of the personnel and the construction machinery, and trigger the warning in time. This real-time warning mechanism is more efficient than manual monitoring, which helps to improve the safety of the construction site. 4. Reduce costs and simplify installation: The present invention uses a monocular camera for three-dimensional space detection, without the need to use a laser radar or a multi-camera system, thereby reducing equipment costs. At the same time, the installation process is simple and convenient, which is convenient for rapid deployment at the construction site. 5. Enhanced environmental adaptability: The present invention is suitable for complex construction environments, especially in small, high-density areas such as bridge construction, and can maintain a stable monitoring effect. The system can adapt to changes under different lighting conditions to ensure all-weather safety monitoring. 6. Improved construction safety management level: The present invention pushes real-time warning information to the management platform and relevant personnel through wireless transmission, which helps construction management personnel to understand the safety status of the site in real time, facilitates rapid response and decision-making, and further improves construction safety from the management level. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Detailed description of the invention Figure 1 is a flow chart of a monitoring method in an embodiment of the present invention.

[0057] Figure 2 Schematic diagram of the monitoring system in the embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation methods described herein are only used to explain the present invention and are not used to limit the present invention.

[0059] In order to solve the safety monitoring problem caused by complex space and dense construction equipment in bridge construction, the present invention provides a dynamic three-dimensional risk area monitoring method and system for bridge construction based on a monocular 3D detection algorithm. Traditional two-dimensional image monitoring methods cannot accurately judge the positional relationship between mechanical equipment and operators in three-dimensional space, and it is difficult to effectively monitor the safety risks in the construction area. Based on a monocular 3D detection algorithm, the present invention obtains the three-dimensional position of construction machinery and personnel from a two-dimensional image through a depth estimation method, generates and dynamically adjusts a three-dimensional risk area. When an operator enters a risk area, the system automatically triggers an early warning and notifies relevant personnel in a timely manner. The present invention has the advantages of low cost, easy deployment, and high real-time performance. It is suitable for high-density and complex construction scenes such as cross-sea bridges, and provides a reliable technical guarantee for construction site safety management.

[0060] Embodiment 1

[0061] In this embodiment, Figure 1 As shown in the figure, the dynamic three-dimensional risk area monitoring method for bridge construction based on monocular 3D detection algorithm includes the following six parts:

[0062] 1. Image acquisition

[0063] The video images of the construction area are collected in real time through a monocular camera. The specific contents include the following:

[0064] 1.1. Select a monocular camera with high resolution and wide viewing angle to ensure coverage of the main construction areas.

[0065] 1.2. Install it at a high position to avoid obstruction and ensure that the camera can monitor key construction locations to the maximum extent.

[0066] 1.3. The monocular camera collects video images of the construction area in real time.

[0067] 2. Image Processing and 3D Detection

[0068] The trained 3D detection model is used to identify the three-dimensional spatial position and size of construction machinery and workers at the construction site. The 3D detection model is built using the MMDetection3D framework; the training process of the 3D model is as follows:

[0069] 2.1. Collect image data of various construction machines and workers, annotate their 3D positions and sizes, and construct a training dataset for construction scenes.

[0070] 2.2. Using ResNet101 as the backbone, we first perform multi-level feature extraction on the image data, and then pass these features to the 3D detection head to complete specific tasks; the introduction of DCN enables the backbone to more flexibly adapt to the geometric deformation of the target in the image, thereby enhancing the effect of feature extraction.

[0071] 2.3. Combine the 3D detection head to predict the three-dimensional information and object category of the object; the three-dimensional information includes the value of the center offset in the x direction, the value of the center offset in the y direction, depth information, the width of the object, the length of the object, the height of the object, and the rotation angle of the object.

[0072] 2.4. Use supervised learning technology to train the 3D detection model so that it can predict the 3D spatial position and size of construction machinery and workers from the video images obtained by the monocular camera.

[0073] 3. Optical flow detection

[0074] Perform optical flow detection on video images to analyze the motion state of construction machinery. Specifically, it includes the following contents:

[0075] 3.1. Input two consecutive frames of images into the LiteFlowNet2 network. The LiteFlowNet2 network estimates the optical flow information layer by layer through a cascaded convolutional neural network, estimates the pixel displacement in the entire image from coarse to fine, and outputs the optical flow field.

[0076] 3.2. Use the 2D target detection model Faster R-CNN to detect construction machinery targets, output the 2D target detection bounding box, and extract the optical flow data of the target area from the optical flow field based on the 2D target detection bounding box;

[0077] 3.3. By calculating the average amplitude of the optical flow vector in the target object area, the motion state of the construction machinery is determined to be "moving" or "stationary".

[0078] 4. Risk Area Generation

[0079] Based on the detection results of the 3D detection model and the optical flow detection results, the three-dimensional risk area around the construction machinery is dynamically generated and adjusted as the construction machinery moves. Specifically, it includes the following contents:

[0080] 4.1. Set the risk zone expansion factor, which is used to determine the size of the risk area.

[0081] 4.2. Set the value of the risk zone expansion factor according to the movement state of the construction machinery; when the construction machinery is in the "moving" state, set the risk zone expansion factor to 1, so that the size of the generated risk area is equivalent to 1 times the distance of the construction machinery boundary box; when the construction machinery is in the "stationary" state, set the risk zone expansion factor to 0.5, so that the size of the generated risk area is reduced to 0.5 times the distance of the construction machinery boundary box.

[0082] 4.3. Extract the basic tensor data of the construction machinery bounding box, including position coordinates, length and width;

[0083] 4.4. Use the risk area expansion factor to expand the specified tensor data. Only the length and width parameters of the construction machinery bounding box are proportionally adjusted, and the adjusted tensor data is used to generate a new three-dimensional risk area bounding box; specifically, the length and width in the tensor data are multiplied by (1+expansion factor) to achieve the expansion operation, while the height remains unchanged.

[0084] In this embodiment, the risk area can move automatically with the movement of the construction machinery, and there is no need to manually re-define the risk area.

[0085] 5. Identity Recognition

[0086] Through face recognition and identity database comparison, the identity of the person entering the risk area can be identified in real time to confirm whether he is an authorized operator of the construction machinery operation team. The specific contents include the following:

[0087] 5.1. Collect and register the facial features of the operating team members, generate feature codes and store them in the identity database.

[0088] 5.2. Capture images of people entering risk areas in real time.

[0089] 5.3. Use the face detection model and the face recognition model to identify and extract the facial feature vector of the person's image, compare the facial feature vector of the person with all the feature vectors in the identity database, and calculate the similarity between the two to determine whether the corresponding person is an authorized person.

[0090] VI. Early Warning

[0091] Real-time monitoring of the location of workers. When the construction machinery is running, an alarm will be triggered if anyone other than the operator enters the three-dimensional risk area. When the construction machinery is stationary, an alarm will be triggered if an unauthorized team worker enters the three-dimensional risk area.

[0092] In this embodiment, Figure 2 As shown, a dynamic three-dimensional risk area monitoring system for bridge construction based on a monocular 3D detection algorithm is provided. The system can implement the above-mentioned method. The system includes the following seven modules:

[0093] 1. Image acquisition module

[0094] 1.1. Equipment selection: Choose a monocular camera with high resolution and wide viewing angle to ensure coverage of the main construction areas.

[0095] 1.2. Installation location: Install it at a high position to avoid obstruction and ensure that the camera can monitor key construction locations to the maximum extent.

[0096] 1.3. Real-time transmission: Use the network to transmit the collected images to the edge computing device in real time for 3D detection and risk analysis.

[0097] In this embodiment, a monocular camera is used in combination with a 3D detection model to achieve three-dimensional spatial positioning of construction machinery and personnel. Compared with the traditional multi-sensor solution, the monocular solution significantly reduces the hardware cost and deployment complexity, is suitable for large-scale applications, and meets the needs of precise on-site positioning.

[0098] 2. Image Processing and 3D Detection Module

[0099] Based on the MMDetection3D framework, a 3D detection model of construction machinery and workers is established to identify the three-dimensional spatial position and size of construction machinery and workers at the construction site.

[0100] 2.1 Model training:

[0101] (1) Data preparation: Collect image data of various construction machines and workers, annotate their 3D positions and sizes, and construct a training dataset for construction scenes.

[0102] (2) Depth estimation: Using ResNet101 as the backbone, we first extract multi-level features from the image, and then pass these features to other modules (such as the detection head) to complete specific tasks. The introduction of DCN (deformable convolution) enables the backbone to more flexibly adapt to the geometric deformation of the target in the image, thereby enhancing the effect of feature extraction.

[0103] (3) Combined with the 3D detection head (FCOS3D) to predict the 3D information and object category of the object. The 3D information includes Δx: the value of the center offset in the x direction. Δy: the value of the center offset in the y direction. d: depth information. w: the width of the object. l: the length of the object. h: the height of the object. θ: the rotation angle of the object.

[0104] (4) Training process: The supervised learning technology is used to train the model so that it can predict the 3D position and size of construction machinery and workers from monocular images.

[0105] 2.2 Real-time inference:

[0106] (1) Input data: Receive camera image data in real time and input it into the trained 3D detection model.

[0107] (2) Output results: The model outputs the categories, three-dimensional positions, and sizes of construction machinery and workers, providing basic data for the risk area generation module.

[0108] 3. Optical flow detection module

[0109] The optical flow module is introduced to analyze the motion state of construction machinery and adjust the size of the risk area according to the state.

[0110] 3.1. Optical flow calculation: Two consecutive frames of images are input into LiteFlowNet2 to calculate the optical flow field.

[0111] Select LiteFlowNet2 network to process the image pairs and obtain the optical flow field. LiteFlowNet2 network is an efficient optical flow estimation algorithm. The main steps are as follows.

[0112] An image pair is two consecutive frames of images (frame_t, frame_(t+1)), representing the scene at a certain moment t and the next moment t+1.

[0113] (1) Input network: Two frames of images are input into the LiteFlowNet2 network. The network estimates the optical flow information layer by layer through a cascaded convolutional neural network, and estimates the pixel displacement in the entire image from coarse to fine.

[0114] (2) Obtaining the optical flow field: The output is an optical flow field, which is a tensor of shape (H, W, 2), where H and W are the height and width of the image, and 2 represents the horizontal (dx) and vertical (dy) displacement of each pixel. The output optical flow field can be understood as the displacement vector of each pixel between two frames of images.

[0115] 3.2. Extract the optical flow information of the target area: Use the 2D detection model Faster R-CNN to detect construction machinery targets, output the 2D target detection bounding box, and extract the optical flow data of the target object area from the optical flow field.

[0116] (1) Use a 2D object detection model (Faster R-CNN) to detect construction machinery in the image.

[0117] (2) The object detection model provides a 2D bounding box for the construction machinery, that is, the location of the construction machinery in the image.

[0118] (3) According to the bounding box of 2D target detection, the optical flow information of the target area is extracted from the optical flow field.

[0119] 3.3. Determine the state of motion: Determine whether the object is moving by calculating the average amplitude of the optical flow vector in the target area. The optical flow module determines whether the state of the construction machinery is "moving" or "stationary".

[0120] In this embodiment, the optical flow module is used to monitor the motion state of the construction machinery in real time and intelligently adjust the expansion factor of the risk area. This dynamic adjustment method enables the system to adapt to the motion or static state of the construction machinery, improves environmental adaptability, and ensures the real-time and accuracy of the risk area.

[0121] 4. Risk Area Generation

[0122] Combined with the construction machinery status information output by the optical flow detection module, the generation and adjustment of dynamic risk areas are realized.

[0123] 4.1. Setting of risk area expansion factor: Set a risk area expansion factor, which is used to determine the size of the risk area.

[0124] 4.2. Dynamic setting of risk zone expansion factor: Set the value of the risk zone expansion factor based on the judgment results of the movement state of the construction machinery.

[0125] When the construction machinery is in the "moving" state, the risk zone expansion factor is set to 1, so that the size of the generated risk area is equivalent to 1 times the distance of the construction machinery boundary box.

[0126] When the construction machinery is in a “stationary” state, the risk zone expansion factor is set to 0.5, so that the size of the generated risk area is reduced to 0.5 times the distance of the construction machinery boundary box.

[0127] 4.3. Bounding box data acquisition and preprocessing: Extract the basic tensor data of the 3D bounding box (construction machinery bounding box) and generate its copy, including position coordinates (x, y, z), length, and width.

[0128] 4.4. Risk area expansion operation: On the generated 3D bounding box tensor, the specified parameters are expanded using the risk area expansion factor.

[0129] Only the length and width parameters of the bounding box are proportionally adjusted to form the expanded risk area bounding box. Specifically, the expansion operation is achieved by multiplying the length and width of the tensor by (1 + expansion factor) respectively, while the height parameter remains unchanged to ensure the rationality of the three-dimensional structure of the risk area.

[0130] The expanded tensor data is used to generate a new three-dimensional risk area bounding box object.

[0131] In this embodiment, based on the three-dimensional position and motion state of the construction machinery, the system realizes automatic follow-up adjustment of the dynamic risk area. Unlike the traditional 2D risk area, which usually needs to be manually delineated due to the lack of depth information in 2D images, the 3D risk area can automatically move with the movement of the vehicle or construction machinery, ensuring that the risk area always accurately covers the high-risk area, greatly improving the automation and accuracy of risk control.

[0132] 5. Identity Recognition Module

[0133] 5.1 Face Detection Module

[0134] (1) Function: Identify the face area in the video stream or image.

[0135] (2) Technology selection: Use a deep learning-based face detection model, MTCNN (Multi-task Cascaded Convolutional Networks).

[0136] (3) Process: Input real-time image frames and input frame data into the face detection model. The model outputs the bounding box coordinates of the face for subsequent extraction of facial features.

[0137] 5.2. Face feature extraction module

[0138] (1) Function: Encode the detected face area and generate a feature vector that can be used for identity comparison.

[0139] (2) Technology selection: Use a pre-trained face recognition model (ArcFace) to extract the feature vector of each face.

[0140] (3) Process: The detected face area is captured and input into the feature extraction model. The model outputs a fixed-length feature vector (usually 128 or 512 dimensions) that represents the unique feature information of the face. The feature vector is used for identity database comparison.

[0141] 5.3. Identity database construction and management module

[0142] (1) Function: Establish and manage the identity database of operating team members. The database contains the facial feature code of each authorized operator.

[0143] (2) Data collection: Before the system is started, multiple facial images of each operating team member are collected.

[0144] (3) Feature encoding: Each face image is passed through a feature extraction module to generate a feature vector and store it in the identity database.

[0145] (4) Database management: Provides the functions of adding, deleting, modifying and checking the identity database to ensure real-time updating of authorized personnel information.

[0146] 5.4. Identity matching and authorization verification module

[0147] (1) Function: To authenticate the identity of persons entering an area and determine whether they are authorized persons.

[0148] (2) Comparison method: Calculate the similarity (Euclidean distance) between the newly detected facial feature vector and each feature vector in the identity database. Set a similarity threshold (e.g., 0.8). When the similarity is higher than the threshold, the person is judged to be an authorized person, otherwise, he is an unauthorized person.

[0149] (3) Process: When a face entering the area is detected, its feature vector is extracted. The feature vector is compared with all feature vectors in the identity database and the similarity is calculated. If the highest similarity exceeds the preset threshold, the person is confirmed to be an authorized operator; if the condition is not met, the person is marked as an unauthorized person.

[0150] (4) Output: Based on the comparison results, the control system performs different operations:

[0151] (5) Authorized personnel: allowed to enter the risk area, and the "Allow entry" prompt is displayed. Unauthorized personnel: trigger the warning system and are prohibited from entering the risk area.

[0152] 5.5 System Workflow

[0153] (1) Initialization: Collect and register the facial features of the operating team members, generate feature codes and store them in the identity database.

[0154] (2) Real-time monitoring: During system operation, images of people entering risk areas are captured in real time.

[0155] (3) Identity verification: Use the face detection and feature extraction module to identify and extract the facial feature vector of the new person. Compare the new features with the identity database, calculate the similarity and determine whether the person is an authorized person.

[0156] 6. Early Warning Module

[0157] Real-time monitoring of the location of workers. When the construction machinery is running, an alarm will be triggered if anyone other than the operator enters the three-dimensional risk area. When the construction machinery is stationary, an alarm will be triggered if an unauthorized team worker enters the three-dimensional risk area.

[0158] 6.1. Relative position calculation: Utilize the data from the 3D detection module to calculate the relative distance between the operator and the construction machinery in real time.

[0159] Real-time monitoring of the location of workers. When the construction machinery is running, if a non-operating driver enters the risk area, an alarm will be triggered; when the construction machinery is stationary, if a non-authorized person from the work team enters the risk area, an alarm will be triggered to notify relevant personnel.

[0160] 6.2. Alarm mechanism: When the system detects the triggering alarm conditions, it will immediately trigger the sound and light alarm to alert the on-site workers.

[0161] The alarm information is sent to the management platform and the mobile devices of relevant managers through the wireless communication module to ensure that the early warning information is received in time.

[0162] In this embodiment, a multi-level intelligent early warning mechanism combining 3D detection, dynamic risk area division and optical flow analysis is provided. The system automatically determines the risk level and triggers different levels of early warning, including on-site sound and light alarm, remote notification of management personnel, real-time recording of risk events, etc. Through multi-level response and closed-loop control, the early warning signal is ensured to be synchronized with the actual risk situation, effectively improving the safety management efficiency and intelligence level of the construction site.

[0163] 7. Management Platform

[0164] Receive and display risk warning information for real-time monitoring by construction managers; view the current status of risk areas, construction machinery movement information, and identity verification results of entrants, so that managers can quickly respond to potential risks.

[0165] 7.1. Platform construction:

[0166] A risk monitoring interface is designed on the management platform to display the location of construction machinery and workers, risk area status and early warning information in real time.

[0167] The platform receives early warning information from front-end equipment and saves alarm records for later analysis.

[0168] 7.2、Function realization:

[0169] (1) Real-time monitoring: Dynamically display construction machinery, workers, and risk areas in a 3D scene.

[0170] (2) Warning records: Save the time, location and construction machinery information of the warning trigger for tracking and review.

[0171] (3) Remote control: Managers can adjust system parameters (such as safety distance) through the platform to meet the needs of different construction scenarios.

[0172] By adopting the above technical solution disclosed in the present invention, the following beneficial effects are obtained:

[0173] The present invention provides a method and system for monitoring dynamic three-dimensional risk areas of bridge construction based on a monocular 3D detection algorithm, and the three-dimensional spatial recognition capability is enhanced: the present invention can extract the three-dimensional spatial position of construction machinery and operators from the two-dimensional image of a single camera through a monocular 3D detection algorithm. This technology solves the deficiency that traditional two-dimensional image detection is difficult to accurately judge spatial relationships, and provides more accurate data support for risk monitoring. Flexibility of dynamic risk area division: the present invention automatically generates and dynamically adjusts the three-dimensional risk area around the construction machinery based on the real-time position and motion state of the construction machinery, so that risk monitoring can adapt to the movement of the equipment and changes in the construction environment. Compared with the statically divided risk area, the dynamic risk area is more flexible and can adapt to complex construction scenes more accurately. Improved real-time warning responsiveness: the present invention can monitor in real time whether the operator enters the risk area by judging the relative position of the personnel and the construction machinery, and trigger the warning in time. This real-time warning mechanism is more efficient than manual monitoring, which helps to improve the safety of the construction site. Reduce costs and simplify installation: the present invention uses a monocular camera for three-dimensional space detection, without the need to use a laser radar or a multi-camera system, thereby reducing equipment costs. At the same time, the installation process is simple and convenient, which is convenient for rapid deployment at the construction site. Enhanced environmental adaptability: The present invention is suitable for complex construction environments, especially in small, high-density areas such as bridge construction, and can maintain a stable monitoring effect. The system can adapt to changes under different lighting conditions to ensure all-weather safety monitoring. Improved construction safety management level: The present invention pushes real-time warning information to the management platform and relevant personnel through wireless transmission, which helps construction management personnel understand the safety status of the site in real time, facilitates rapid response and decision-making, and further improves construction safety from the management level.

[0174] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention.

Claims

1. A method for monitoring dynamic three-dimensional risk areas of bridge construction based on a monocular 3D detection algorithm, characterized in that: The following steps are included: S1. Image acquisition: real-time video images of the construction area are collected through a monocular camera; S2. Image processing and 3D detection: Use the trained 3D detection model to identify the three-dimensional spatial position and size of construction machinery and workers at the construction site; S3, optical flow detection: perform optical flow detection on the video image to analyze the motion state of the construction machinery; S4. Risk area generation: Based on the detection results of the 3D detection model and the optical flow detection results, dynamically generate and adjust the three-dimensional risk area around the construction machinery as the construction machinery moves; S5. Identity recognition: Through face recognition and identity database comparison, the identity of the person entering the risk area is identified in real time to confirm whether he is an authorized operator of the construction machinery operation team; S6. Early warning: Real-time monitoring of the position of workers. When the construction machinery is running, if a person other than the operating driver enters the three-dimensional risk area, an alarm will be triggered; when the construction machinery is stationary, if an unauthorized team worker enters the three-dimensional risk area, an alarm will be triggered.

2. The method for monitoring dynamic three-dimensional risk areas in bridge construction based on a monocular 3D detection algorithm according to claim 1 is characterized in that: Step S1 specifically includes the following contents: S11. Select a monocular camera with high resolution and wide viewing angle to ensure coverage of the main construction areas; S12, install at a high position to avoid obstruction and ensure that the camera can monitor key construction locations to the maximum extent; S13, the monocular camera collects video images of the construction area in real time.

3. The method for monitoring dynamic three-dimensional risk areas in bridge construction based on a monocular 3D detection algorithm according to claim 1 is characterized in that: The 3D detection model is built using the MMDetection3D framework; the training process of the 3D model is as follows: S21. Collect image data of various construction machines and workers, annotate their 3D positions and sizes, and construct a training data set of construction scenes; S22, using ResNet101 as the backbone, first perform multi-level feature extraction on the image data, and then pass these features to the 3D detection head to complete the specific task; the introduction of DCN enables the backbone to more flexibly adapt to the geometric deformation of the target in the image, thereby enhancing the effect of feature extraction; S23, predicting the three-dimensional information and the object category of the object in combination with the 3D detection head; the three-dimensional information includes the value of the center offset in the x direction, the value of the center offset in the y direction, the depth information, the width of the object, the length of the object, the height of the object, and the rotation angle of the object; S24. Use supervised learning technology to train the 3D detection model so that it can predict the three-dimensional spatial position and size of construction machinery and workers from the video images obtained by the monocular camera.

4. The method for monitoring dynamic three-dimensional risk areas in bridge construction based on a monocular 3D detection algorithm according to claim 1 is characterized in that: Step S3 specifically includes the following contents: S31, input two consecutive frames of images into the LiteFlowNet2 network, which estimates the optical flow information layer by layer through a cascaded convolutional neural network, estimates the pixel displacement in the entire image from coarse to fine, and outputs the optical flow field; S32, using the 2D target detection model Faster R-CNN to detect construction machinery targets, outputting a 2D target detection bounding box, and extracting the optical flow data of the target area from the optical flow field based on the 2D target detection bounding box; S33. By calculating the average amplitude of the optical flow vector in the target object area, it is determined whether the motion state of the construction machinery is "moving" or "stationary".

5. The method for monitoring dynamic three-dimensional risk areas in bridge construction based on a monocular 3D detection algorithm according to claim 1 is characterized in that: Step S4 specifically includes the following contents: S41, setting a risk area expansion factor, which is used to determine the size of the risk area; S42, setting the value of the risk zone expansion factor according to the movement state of the construction machinery; S43, extracting basic tensor data of the construction machinery boundary box, including position coordinates, length and width; S44. Expand the specified tensor data using the risk area expansion factor.

6. The method for monitoring dynamic three-dimensional risk areas in bridge construction based on a monocular 3D detection algorithm according to claim 5 is characterized in that: Specifically, step S42 is as follows: when the construction machinery is in a "moving" state, the risk zone expansion factor is set to 1, so that the size of the generated risk area is equivalent to 1 times the distance of the construction machinery boundary box; when the construction machinery is in a "stationary" state, the risk zone expansion factor is set to 0.5, so that the size of the generated risk area is reduced to 0.5 times the distance of the construction machinery boundary box.

7. The method for monitoring dynamic three-dimensional risk areas in bridge construction based on a monocular 3D detection algorithm according to claim 5, characterized in that: Step S44 specifically includes proportionally adjusting only the length and width parameters of the construction machinery boundary box, and using the adjusted tensor data to generate a new three-dimensional risk area boundary box; The length and width of the tensor data are multiplied by (1 + expansion factor) to implement the expansion operation, while the height remains unchanged.

8. The method for monitoring dynamic three-dimensional risk areas of bridge construction based on a monocular 3D detection algorithm according to claim 1, characterized in that: Step S5 specifically includes the following contents: S51, collect and register the facial features of the operating team members, generate feature codes and store them in the identity database; S52, capturing images of people entering the risk area in real time; S53. Use the face detection model and the face recognition model to identify and extract the face feature vector of the person's image, compare the face feature vector of the person with all the feature vectors in the identity database, and calculate the similarity between the two to determine whether the corresponding person is an authorized person.

9. A dynamic three-dimensional risk area monitoring system for bridge construction based on a monocular 3D detection algorithm, characterized in that: include, Image acquisition module: collects video images of the construction area in real time through a monocular camera; Image processing and 3D detection module: Use the trained 3D detection model to identify the 3D spatial position and size of construction machinery and workers at the construction site; Optical flow detection module: performs optical flow detection on video images to analyze the motion state of construction machinery; Risk area generation: Based on the detection results of the 3D detection model and the optical flow detection results, the three-dimensional risk area around the construction machinery is dynamically generated; Identity recognition module: Through face recognition and identity database comparison, the identity of the person entering the risk area is identified in real time to confirm whether he is an authorized operator of the construction machinery operation team; Early warning module: real-time monitoring of the location of workers. When the construction machinery is running, if anyone other than the operator enters the three-dimensional risk area, an alarm will be triggered; when the construction machinery is stationary, if an unauthorized team worker enters the three-dimensional risk area, an alarm will be triggered.

10. The bridge construction dynamic three-dimensional risk area monitoring system based on monocular 3D detection algorithm according to claim 9, characterized in that: The system also includes, Management platform: Receive and display risk warning information for real-time monitoring by construction managers; view the current status of risk areas, construction machinery movement information, and identity verification results of entrants, so that managers can quickly respond to potential risks.

Images