Construction operation site real-time monitoring device and method based on three-dimensional laser

Abstract

The invention discloses a construction operation site real-time monitoring method based on three-dimensional laser. The method is characterized by comprising the following steps: S1, positioning and calibrating a safe operation boundary area; s2, positioning and calibrating the initial height of the crane; s3, generating a digital virtual light wall; s2, fitting the boundary point coordinate set in the S1 into a closed operation area boundary, and generating a digital virtual light wall by combining an operation initial height reference and an operation height upper limit; s4, coordinate cooperative calculation of the first positioning module and the second positioning module; s5, calculating the distances between the positioning module I and the virtual light wall and between the positioning module II and the virtual light wall; and S6, carrying out distance over-limit alarm. The real-time coordinates of the top end of the lifting arm and the lifting hook are collected through the double Beidou positioning modules respectively, the movement speed of the lifting arm and the lifting hook is calculated in combination with space coordinate points, the elevation angle of the lifting arm is calculated through an arc tangent algorithm, dynamic tracking of the three-dimensional posture of the construction machine is achieved, and the blank that in the prior art, only the distance can be measured through energy, and the posture cannot be tracked is filled.

Description

Technical Field

[0001] This invention relates to the field of power production operations, specifically a real-time monitoring device and method for construction operation sites based on three-dimensional laser technology. Background Technology

[0002] Almost all power production operations involve the threat of high-voltage electricity. When personnel enter the work area, they must strictly follow the operating procedures, go to the designated area according to the work requirements, and perform standardized operations on the designated targets. For correct operation and personnel safety, the commonly used intelligent supervision methods are: using positioning methods such as UWB and GPS to locate personnel entering the area, and using intelligent algorithms such as visual AI to intelligently analyze personnel's clothing and personnel identity.

[0003] With the innovation and development of science and technology, laser scanning technology, multi-sensor fusion technology, and computer vision technology have developed rapidly, providing a technical foundation for real-time monitoring of power construction operations.

[0004] Existing technologies include methods for measuring the spatial distance between hazard sources and targets using monocular vision. This involves controlling a laser projection device to scan the entire target scene, transmitting laser point cloud data, controlling a vision camera to acquire the entire field of view of the scene, and transmitting image data. A scene spatial model is then constructed using the laser point cloud data and the image data. The target spatial distance is determined by matching the point cloud data with the image data using the scene spatial model. While this technology can achieve distance monitoring within a certain range, it suffers from the following problems: Real-time tracking of mechanical three-dimensional posture is not achieved: It only realizes the function of measuring the distance between targets, but does not have the ability to track the three-dimensional posture of the crane in real time. It is impossible to dynamically grasp the changes in the spatial position of the equipment, and it is difficult to judge the collision trend of the crane's movement trajectory and obstacles. The lack of a digital virtual light wall: The lack of a digital virtual light wall early warning mechanism to achieve comprehensive three-dimensional protection, and the monitoring based solely on distance calculation, makes it impossible to set safety boundaries in advance to trigger proactive early warnings; The monocular ranging method has low reliability in distance measurement: The monocular + laser scheme is easily affected by occlusion due to the "single image feature + laser point cloud", which results in insufficient effective information to support the "point cloud-image matching", ultimately leading to low reliability in distance measurement.

[0005] In summary, existing technologies still have shortcomings in terms of scenario adaptation, functional completeness, and data fusion, and targeted breakthroughs are urgently needed. Summary of the Invention

[0006] This invention addresses the core risk posed by "exceeding the operating range" during power construction operations, which threatens the personal safety of workers and the stable operation of power equipment. This risk can easily lead to accidents such as crane collisions with power lines. Furthermore, existing monitoring technologies suffer from significant shortcomings in monitoring accuracy, scenario adaptability, and risk response efficiency, making it difficult to effectively control such risks. This invention aims to solve the problems of insufficient accurate monitoring of safe distances, effective identification of dynamic risks, and multi-scenario adaptability in power construction operations by optimizing the monitoring technology architecture, achieving efficient fusion of multi-source monitoring data, and precise tracking of operational targets. This will provide reliable technical protection for the safety of workers and power equipment.

[0007] Technical solution: This invention discloses a method for real-time monitoring of construction sites based on three-dimensional laser technology, the method comprising the following steps: S1. Safety operation boundary area positioning and calibration: The positioning module is placed at each corner point of the safety operation area to collect the pre-processed three-dimensional coordinates of each corner point in real time and store them as a set of boundary point coordinates. S2. Crane starting height positioning and calibration; then fix positioning module one to the top of the boom, use the elevation value of positioning module one in the initial state of the crane as the reference for the starting height of the operation, and confirm the upper limit parameter of the operating height. S3. Generation of digital virtual light wall: Fit the set of boundary point coordinates of S1 to the boundary of the closed working area, and generate a digital virtual light wall by combining the working start height benchmark and the working height upper limit; S4. Coordinate calculation of positioning module 1 and positioning module 2; Fix positioning module 2 at the hook, read the coordinates (Lon1, Lati1, Alt1) and (Lon2, Lati2, Alt2) of positioning module 1 and positioning module 2 in real time, and calculate the boom length and boom elevation angle. S5. Calculation of the distance from positioning module 1 and positioning module 2 to the virtual light wall; S6, Distance Exceeded Alarm.

[0008] Preferably, in S1, the positioning module one is activated and placed sequentially at each corner point of the safe working area, and the preprocessed three-dimensional coordinates of each corner point are collected in real time as the boundary point coordinates.

[0009] Preferably, in S3, the "Delaunay triangulation algorithm" is used to fit the set of boundary point coordinates into the boundary of the closed working area; and a light wall trigger threshold is preset. When the distance between the positioning module one or the positioning module two and the light wall is less than the threshold, an alarm is triggered.

[0010] Preferably, after the positioning module one or positioning module two collects data, the data is preprocessed; the preprocessing is an outlier removal + smoothing filtering algorithm, firstly removing outlier data with positioning accuracy greater than 5cm by the 3σ principle, and then smoothing coordinate fluctuations by moving average filtering to ensure the stability of positioning data.

[0011] Preferably, a timestamp synchronization mechanism is adopted between positioning module one and positioning module two. The software system reference clock is used as the anchor point to add timestamps to the positioning data of the two positioning modules. If the timestamp difference between the two data exceeds 50ms, the lagging data is automatically discarded and the system waits for the next round of acquisition to ensure that the time synchronization error of the data between the two modules does not exceed 30ms.

[0012] Specifically, in S4: The formula for calculating the boom length is: L = √[(Lon1 - Lon2)] 2 +(Lati1-Lati2) 2 +(Alt1-Alt2) 2 ] The formula for calculating the boom elevation angle is: θ=arctan [(Alt1-Alt_base) / √[(Lon1-Lon_base) 2 +(Lati1-Lati_base) 2 ]] In the formula, Alt_base represents the elevation coordinates of the crane base, Lon_base represents the longitude coordinates of the crane base, and Lati_base represents the latitude coordinates of the crane base.

[0013] Specifically, in S5, the shortest distance D_boom from the spatial point to the virtual light wall and the shortest distance T_boom from the upper boundary to the first positioning module to the virtual light wall are calculated respectively, as well as the shortest distance D_hook from the second positioning module to the virtual light wall and the shortest distance T_hook from the upper boundary to the second positioning module. An alarm is triggered when any one of D_boom, T_boom, D_hook, or T_hook is less than the threshold.

[0014] Specifically, S6 also includes the use of 3D LiDAR sensors for obstacle recognition and warning, specifically including: S6-1. The scene is scanned in real time by a 3D LiDAR sensor deployed at the top of the boom, and the laser point cloud data is returned. The scene video is captured by a video sensor and the image data is returned. The "timestamp interpolation method" is used to align the laser point cloud with the binocular vision image returned by the video sensor. Each frame of laser point cloud is matched with 2-3 frames of binocular vision image. The spatial deviation caused by the small displacement of the sensor is compensated by IMU data. S6-2. By preprocessing the laser point cloud and the binocular vision image, based on the pre-calibrated extrinsic parameters, the laser point cloud is transformed from the radar coordinate system to the binocular camera coordinate system. It is then projected onto the image plane through the camera intrinsic parameters to achieve the initial mapping between the point cloud and the image pixels. For the overlapping area of ​​the laser point cloud and the binocular vision image, the extrinsic parameters are fine-tuned using the ICP algorithm to minimize the distance error between the point cloud and the depth map and ensure spatial consistency. The S6-3 features a built-in lightweight YOLOv8 obstacle recognition model that performs real-time inference on the fused video frames, outputting the 2D bounding box of the obstacle and the corresponding 3D point cloud cluster. For the 3D point cloud cluster, a KD tree is used to search for the nearest point between the top of the boom and the point cloud cluster. The actual distance D_obstacle is calculated using the spatial distance formula and compared with a preset safety threshold D2. If D_obstacle is less than D2, an alarm is triggered.

[0015] Specifically, the construction process of the lightweight YOLOv8 obstacle recognition model mainly revolves around "reducing the number of model parameters, computational load and inference latency while ensuring recognition accuracy". The core is to carry out lightweight transformation and optimization of the original YOLOv8 architecture.

[0016] First, data preparation is performed. A dataset suitable for obstacle scenarios is selected, and after annotation, the samples are augmented through random cropping, flipping, and brightness adjustment. Normalization and resizing are then preprocessed. Next, the YOLOv8 architecture is lightweighted. The backbone uses Depthwise Separable Convolution instead of standard convolutions, or lightweight modules such as MobileNet and ShuffleNet are introduced to simplify the residual structure and reduce the number of parameters and computations. The Neck's PAN-FPN feature fusion layer compresses the channel dimension using 1x1 convolutions, simplifying cross-layer fusion and avoiding redundant computations. The decoupled classification and regression branches in the Head use a simplified convolution structure to reduce the number of channels and layers. Finally, model compression techniques are employed: INT8 quantization reduces weight precision, structured pruning removes redundant channels, or knowledge distillation is introduced to further compress the model. During the training phase, a lightweight backbone network is initialized using transfer learning, and a mini-batch optimizer, learning rate decay strategy, and regularization are used to prevent overfitting. Finally, the validation set is used to evaluate metrics such as mAP and FPS. The pruning ratio and distillation parameters are repeatedly fine-tuned and iteratively optimized until a balance between lightweight and performance is achieved for obstacle recognition requirements.

[0017] Specifically: D_obstacle= Where (X_boom, Y_boom, Z_boom) represent the three-dimensional coordinates of the top of the boom (obtained by the Ravage fusion detector), and (X_obs, Y_obs, Z_obs) represent the three-dimensional coordinates of the point closest to the top of the boom found from the obstacle 3D point cloud cluster through KD-tree search. This formula calculates the square root of the sum of the squares of the differences between two points in the X, Y, and Z dimensions in 3D space, yielding the straight-line distance between the two points, i.e., the actual physical distance D_obstacle.

[0018] This invention also discloses a real-time monitoring device for construction sites based on three-dimensional laser technology. It includes: a point cloud acquisition module, a video acquisition module, a data processing module, a recognition module, a ranging module, a positioning module one, a positioning module two, a virtual light wall construction module, a collision analysis module, and an alarm module. The data processing module receives data output from the point cloud acquisition module and the video acquisition module. The output of the data processing module is connected to the recognition module, and the output of the recognition module is connected to the ranging module. The outputs of positioning modules one, two, and the virtual light wall construction module are all connected to the collision analysis module. The output of positioning module one is also connected to the virtual light wall construction module. The outputs of the collision analysis module and the ranging module are connected to the alarm module. Specifically: Point cloud acquisition module: used to acquire point cloud data in real time; Video capture module: Used to capture video data in real time; Data processing module: used to align, register and fuse video and point cloud data, and then stitch together data collected by different cameras and LiDAR devices; Recognition module: Cleans and annotates the acquired images and point cloud data, and uses the model to detect obstacles in the video footage and laser point cloud; Distance measuring module: calculates the distance between a fixed obstacle and the boom in real time; Positioning Module 1: Used for collecting positioning data at various corner points in the safe working area, as well as for collecting positioning data at key nodes at the top of the boom; Positioning Module 2: Used to collect positioning data of key nodes of the hook; Collision analysis module: used to calculate the real-time distances between positioning module 1, positioning module 2, and the sides and top of the digital virtual light wall; Alarm module: Triggers visual and auditory alarms based on the distance output by the ranging module and the collision analysis module.

[0019] Beneficial effects of the present invention (1) In response to the problem that the existing technology “does not achieve real-time tracking of mechanical three-dimensional posture”, the present invention collects the real-time coordinates of the top of the boom and the hook respectively through dual Beidou positioning modules, calculates the movement speed of the boom and hook by combining spatial coordinate points, and calculates the boom elevation angle by arctangent algorithm, etc., to realize dynamic tracking of the three-dimensional posture of construction machinery, filling the gap of the existing technology that can only measure distance and cannot track posture. (2) In view of the limitation of “not constructing a digital virtual light wall”, this invention supports the calibration of irregular areas with 3-8 corner points. By fitting the closed operation boundary and combining the operation start height and the height limit, a digital virtual light wall of “plane + height” is generated, which replaces the passive monitoring of the existing technology that relies solely on distance calculation, and realizes “early prevention and control and proactive risk avoidance”. (3) In response to the problem of "low reliability of monocular ranging", this invention adopts the laser-visual fusion technology of laser radar and binocular camera to make up for the defect of single laser radar in recognizing low reflectivity targets, solve the depth error problem of monocular vision, and significantly improve the reliability and accuracy of distance measurement in complex environments. Attached Figure Description

[0020] Figure 1 This is a block diagram of the monitoring device of the present invention.

[0021] Figure 2 This is a flowchart of the monitoring method of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to embodiments, but the scope of protection of the present invention is not limited thereto: Combination Figure 1 This invention discloses a real-time monitoring device for construction sites based on three-dimensional laser technology. It includes: a point cloud acquisition module, a video acquisition module, a data processing module, a recognition module, a ranging module, a positioning module one, a positioning module two, a virtual light wall construction module, a collision analysis module, and an alarm module. The data processing module receives data output from the point cloud acquisition module and the video acquisition module. The output of the data processing module is connected to the recognition module, and the output of the recognition module is connected to the ranging module. The outputs of positioning modules one, two, and the virtual light wall construction module are all connected to the collision analysis module. The output of positioning module one is also connected to the virtual light wall construction module. The outputs of the collision analysis module and the ranging module are connected to the alarm module. Specifically: Point cloud acquisition module: used to acquire point cloud data in real time; Video capture module: Used to capture video data in real time; Data processing module: used to align, register and fuse video and point cloud data, and then stitch together data collected by different cameras and LiDAR devices; Recognition module: Cleans and annotates the acquired images and point cloud data, and uses the model to detect obstacles in the video footage and laser point cloud; Distance measuring module: calculates the distance between a fixed obstacle and the boom in real time; Positioning Module 1: Used for collecting positioning data at various corner points in the safe working area, as well as for collecting positioning data at key nodes at the top of the boom; Positioning Module 2: Used to collect positioning data of key nodes of the hook; Collision analysis module: used to calculate the real-time distances between positioning module 1, positioning module 2, and the sides and top of the digital virtual light wall; Alarm module: Triggers visual and auditory alarms based on the distance output by the ranging module and the collision analysis module.

[0023] Combination Figure 2 A monitoring method based on a three-dimensional laser-based real-time monitoring device for construction sites is described below: S1. Once the BeiDou positioning module is activated, it is placed sequentially at each corner point of the safe working area (supporting 3-8 corner points to meet the calibration requirements of irregular areas). It collects the preprocessed 3D coordinates of each corner point in real time and automatically stores them as a "boundary point coordinate set {(Lon1, Lati1, Alt1), (Lon2, Lati2, Alt2), ... (Lon...(Lon...)}". n Lati n Alt n )}”; Beidou positioning module: Relying on RTK differential positioning technology, it has centimeter-level real-time accurate positioning capability; The equipment adopts an adsorption installation method and is equipped with dual Beidou positioning modules. Beidou positioning module one is ultimately deployed on the top of the crane, and Beidou positioning module two is deployed at the hook position. Through the collaboration of the two modules, multi-point accurate positioning is achieved.

[0024] S2. When the crane is in the initial state, record the elevation value Alt0 of positioning module one as the "starting height reference for operation", and at the same time manually input the upper limit parameter H_max of the operation height (unit: m). S3. The "Delaunay triangulation algorithm" is used to fit the set of boundary point coordinates into the boundary of the closed working area. Combined with the starting height reference Alt0 and the upper limit of the working height H_max, a three-dimensional digital virtual light wall is generated (spatial range: within the boundary, elevation [Alt0, H_max]). The light wall trigger threshold D1 is preset (default 1m, supports user customization). That is, when the distance between the boom / hook and the light wall is less than D1, an alarm is triggered. S4. Real-time reading of the coordinates (Lon1, Lati1, Alt1) of positioning module one (top of boom) and the coordinates (Lon2, Lati2, Alt2) of positioning module two (hook), and calculation of boom length L=√[(Lon1-Lon2)] using the distance formula between two points in space. 2 +(Lati1-Lati2) 2 +(Alt1-Alt2) 2 The boom elevation angle θ is calculated by combining the preset coordinates (Lon_base, Lati_base, Alt_base) of the crane base with the coordinates: θ = arctan[(Alt1-Alt_base) / √[(Lon1-Lon_base)]. 2 +(Lati1-Lati_base) 2 This enables dynamic monitoring of the crane's posture. S5. Using the spatial point to virtual light wall distance algorithm, calculate the shortest distance D_boom from positioning module 1 (top of boom) to the side boundary of the virtual light wall and the shortest distance T_boom from the top boundary, and the shortest distance D_hook from positioning module 2 (hook) to the boundary of the virtual light wall and the shortest distance T_hook from the top boundary. The distance calculation cycle is synchronized with the positioning data acquisition cycle (100ms / time). S6. Compare D_boom, T_boom, D_hook, and T_hook with the preset threshold D1 respectively. If any value is less than D1, immediately generate a "distance over-limit alarm message" (including alarm type: boom over-limit / hook over-limit, current distance, trigger time), and transmit it synchronously to the work monitoring software module and the audible and visual alarm control module via TCP / IP protocol. The alarm response time shall not exceed 200ms.

[0025] In this embodiment, the audible and visual alarm uses LED beads and a buzzer as the core alarm components and triggers visual alarms by flashing light. The device adopts a magnetic installation method and can be flexibly deployed at the work site or in the crane cab. When an alarm signal is received, it can achieve synchronous audible and visual alarm by emitting sound through the buzzer and flashing LED indicator.

[0026] Specifically, the Rayvision-integrated obstacle recognition and alarm system includes: S6-1. The scene is scanned in real time by a 3D LiDAR sensor deployed at the top of the boom, and the laser point cloud data is returned. The scene video is captured by a video sensor and the image data is returned. The "timestamp interpolation method" is used to align the laser point cloud with the binocular vision image returned by the video sensor. Each frame of laser point cloud is matched with 2-3 frames of binocular vision image. The spatial deviation caused by the small displacement of the sensor is compensated by IMU data. In a preferred embodiment, a laser-visual fusion detector is used: it has a built-in 3D LiDAR sensor and a video sensor, which can achieve scanning coverage of 60° conical angle and 180° range, and collect three-dimensional point cloud data and two-dimensional video data of the environment around the crane boom in real time; the device adopts an adsorption installation design and is deployed on the top of the crane boom, and can complete environmental detection synchronously with the movement of the boom.

[0027] S6-2. By preprocessing the laser point cloud and the binocular vision image, based on the pre-calibrated extrinsic parameters, the laser point cloud is transformed from the radar coordinate system to the binocular camera coordinate system. It is then projected onto the image plane through the camera intrinsic parameters to achieve the initial mapping between the point cloud and the image pixels. For the overlapping area of ​​the laser point cloud and the binocular vision image, the extrinsic parameters are fine-tuned using the ICP algorithm to minimize the distance error between the point cloud and the depth map and ensure spatial consistency. The S6-3 features a built-in lightweight YOLOv8 obstacle recognition model that performs real-time inference on the fused video frames, outputting the 2D bounding box of the obstacle and the corresponding 3D point cloud cluster. For the 3D point cloud cluster, a KD tree is used to search for the nearest point between the top of the boom and the point cloud cluster. The actual distance D_obstacle is calculated using the spatial distance formula and compared with a preset safety threshold D2. If D_obstacle is less than D2, an alarm is triggered.

[0028] The construction process of the lightweight YOLOv8 obstacle recognition model mainly revolves around "reducing the number of model parameters, computational load and inference latency while ensuring recognition accuracy". The core is to carry out lightweight transformation and optimization of the original YOLOv8 architecture.

[0029] First, data preparation is performed by selecting datasets suitable for obstacle scenarios (such as vehicles, wires, etc.). After annotation, the samples are augmented through random cropping, flipping, and brightness adjustment, and preprocessing such as normalization and resizing is carried out. Next, the YOLOv8 architecture is lightweighted. In the backbone, Depthwise Separable Convolution replaces standard convolutions, or lightweight modules such as MobileNet and ShuffleNet are introduced to simplify the residual structure and reduce the number of parameters and computations. The PAN-FPN feature fusion layer in the Neck uses 1x1 convolutions to compress the channel dimension, simplifying cross-layer fusion and avoiding redundant computations. The decoupled classification and regression branches in the Head adopt a simplified convolutional structure, reducing the number of channels and layers. Then, model compression techniques are combined, such as INT8 quantization to reduce weight precision, structured pruning to remove redundant channels, or knowledge distillation (using the original YOLOv8 as a teacher model to transfer feature knowledge) to further compress the model. During the training phase, a lightweight backbone network is initialized using transfer learning, and a mini-batch optimizer, learning rate decay strategy, and regularization are used to prevent overfitting. Finally, the validation set is used to evaluate metrics such as mAP and FPS. The pruning ratio and distillation parameters are repeatedly fine-tuned and iteratively optimized until a balance between lightweight and performance is achieved for obstacle recognition requirements.

[0030] The monitoring process of this invention is as follows: Before the construction work begins, the on-site workers use the Beidou positioning device to place the Beidou positioning module one at each corner of the work area to locate and calibrate the work boundary position. After the positioning and calibration are completed, the Beidou module one is attached to the top of the crane (at this time the crane is stationary and the boom is not extended and is placed horizontally) to perform the positioning and calibration operation of the starting height of the work and set the working height parameters. The system automatically generates a digital virtual light wall based on the above positioning data. The first Beidou positioning module is attached to the top of the boom, and the second Beidou positioning module is attached to the hook. The first Beidou positioning module collects the node coordinates of the top of the boom in real time, and the second Beidou positioning module collects the node coordinates of the hook in real time. Together, they realize the dynamic monitoring of the crane's attitude. At the same time, the distance between the boom and hook and the digital virtual light wall is dynamically calculated. When the distance is less than the preset threshold, the alarm mechanism is automatically triggered, and the alarm information is synchronously transmitted to the operation monitor and the on-site audible and visual alarm. The radar-vision fusion detector is attached to the top of the boom, with its binocular cameras pointing upwards to identify obstacles above the boom in real time. It simultaneously calculates the actual distance between the boom and the obstacle and compares the calculated distance with a preset safety threshold. When the monitored distance is less than the preset safety threshold, an alarm mechanism is automatically triggered, and the alarm information is simultaneously transmitted to the operation monitor and the on-site audible and visual alarm.

[0031] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for real-time monitoring of construction sites based on three-dimensional laser technology, characterized in that... The method includes the following steps: S1. Safety operation boundary area positioning and calibration: The positioning module is placed at each corner point of the safety operation area to collect the pre-processed three-dimensional coordinates of each corner point in real time and store them as a set of boundary point coordinates. S2. Crane starting height positioning and calibration; then fix positioning module one to the top of the boom, use the elevation value of positioning module one in the initial state of the crane as the reference for the starting height of the operation, and confirm the upper limit parameter of the operating height. S3. Generation of digital virtual light wall: Fit the set of boundary point coordinates of S1 to the boundary of the closed working area, and generate a digital virtual light wall by combining the working start height benchmark and the working height upper limit; S4. Coordinate calculation of positioning module 1 and positioning module 2; Fix positioning module 2 at the hook, read the coordinates (Lon1, Lati1, Alt1) and (Lon2, Lati2, Alt2) of positioning module 1 and positioning module 2 in real time, and calculate the boom length and boom elevation angle. S5. Calculation of the distance from positioning module 1 and positioning module 2 to the virtual light wall; S6, Distance Exceeded Alarm.

2. The method for real-time monitoring of construction sites based on three-dimensional laser technology according to claim 1, characterized in that: In S1, the positioning module one is activated and placed sequentially at each corner point of the safe working area. The preprocessed three-dimensional coordinates of each corner point are collected in real time as the boundary point coordinates.

3. The method according to claim 1, characterized in that... In S3, the "Delaunay triangulation algorithm" is used to fit the set of boundary point coordinates into the boundary of the closed working area; and a preset light wall trigger threshold is set. When the distance between the positioning module 1 or the positioning module 2 and the light wall is less than the threshold, an alarm is triggered.

4. The method according to claim 1, characterized in that... After the positioning module one or positioning module two collects data, the data is preprocessed. The preprocessing is an outlier removal and smoothing filtering algorithm. First, outlier data with a positioning accuracy greater than 5cm is removed by the 3σ principle. Then, the coordinate fluctuation is smoothed by the moving average filter to ensure the stability of the positioning data.

5. The method according to claim 1, characterized in that... The positioning module 1 and positioning module 2 adopt a timestamp synchronization mechanism. The software system reference clock is used as the anchor point to add timestamps to the positioning data of the two positioning modules. If the timestamp difference between the two data exceeds 50ms, the lagging data is automatically discarded and the system waits for the next round of acquisition to ensure that the time synchronization error of the data between the two modules does not exceed 30ms.

6. The method according to claim 1, characterized in that... In S4: The formula for calculating the boom length is: L=√[(Lon1-Lon2) 2 +(Sides1-Sides2) 2 +(Alt1-Alt2) 2 ] The formula for calculating the boom elevation angle is: θ=arctan [(Alt1-Alt_base) / √[(Lon1-Lon_base) 2 +(Lati1-Lati_base) 2 ]] In the formula, Alt_base represents the elevation coordinates of the crane base, Lon_base represents the longitude coordinates of the crane base, and Lati_base represents the latitude coordinates of the crane base.

7. The method according to claim 1, characterized in that... In S5, the shortest distance D_boom from the spatial point to the virtual light wall and the shortest distance T_boom from the upper boundary to the location module 1 are calculated respectively, and the shortest distance D_hook from the location module 2 to the boundary of the virtual light wall and the shortest distance T_hook from the upper boundary to the location module 2 are calculated respectively. An alarm is triggered when any one of D_boom, T_boom, D_hook, or T_hook is less than the threshold.

8. The method according to claim 1, characterized in that... The S6 also includes the use of 3D LiDAR sensors for obstacle recognition and warning, specifically including: S6-1. The scene is scanned in real time by a 3D LiDAR sensor deployed at the top of the boom, and the laser point cloud data is returned. The scene video is captured by a video sensor and the image data is returned. The "timestamp interpolation method" is used to align the laser point cloud with the binocular vision image returned by the video sensor. Each frame of laser point cloud is matched with 2-3 frames of binocular vision image. The spatial deviation caused by the small displacement of the sensor is compensated by IMU data. S6-2. By preprocessing the laser point cloud and the binocular vision image, based on the pre-calibrated extrinsic parameters, the laser point cloud is transformed from the radar coordinate system to the binocular camera coordinate system. It is then projected onto the image plane through the camera intrinsic parameters to achieve the initial mapping between the point cloud and the image pixels. For the overlapping area of ​​the laser point cloud and the binocular vision image, the extrinsic parameters are fine-tuned using the ICP algorithm to minimize the distance error between the point cloud and the depth map and ensure spatial consistency. The S6-3 features a built-in lightweight YOLOv8 obstacle recognition model that performs real-time inference on the fused video frames, outputting the 2D bounding box of the obstacle and the corresponding 3D point cloud cluster. For the 3D point cloud cluster, a KD tree is used to search for the nearest point between the top of the boom and the point cloud cluster. The actual distance D_obstacle is calculated using the spatial distance formula and compared with a preset safety threshold D2. If D_obstacle is less than D2, an alarm is triggered.

9. The method according to claim 8, characterized in that... In S6-3, the construction of the lightweight YOLOv8 obstacle recognition model includes: Data preparation involves selecting a dataset suitable for obstacle scenarios, completing annotations, and then augmenting the samples through random cropping, flipping, brightness adjustment, and other data enhancement techniques. Normalization and size resizing are also performed as preprocessing steps. The YOLOv8 architecture is lightened by replacing standard convolutions with Depthwise SeparableConvolution in the Backbone, or by introducing lightweight modules such as MobileNet and ShuffleNet to simplify the residual structure and reduce the number of parameters and computations. The PAN-FPN feature fusion layer in the Neck uses 1x1 convolutions to compress the channel dimension, simplifying the cross-layer fusion hierarchy and avoiding redundant computations. The decoupled classification and regression branches in the Head adopt a simplified convolution structure to reduce the number of channels and layers. Then, model compression techniques are combined, such as reducing weight precision through INT8 quantization, removing redundant channels through structured pruning, or introducing knowledge distillation to further compress the model. During the training phase, a lightweight backbone network is initialized using transfer learning, and a mini-batch optimizer, learning rate decay strategy, and regularization are used to prevent overfitting. Finally, the validation set is used to evaluate the mAP and FPS metrics, and the pruning ratio and distillation parameters are repeatedly fine-tuned and iteratively optimized until the lightweight and performance balance required for obstacle recognition is achieved.

10. A real-time monitoring device for construction sites based on three-dimensional laser technology, characterized in that... include: The system comprises a point cloud acquisition module, a video acquisition module, a data processing module, a recognition module, a ranging module, a positioning module 1, a positioning module 2, a virtual light wall construction module, a collision analysis module, and an alarm module. The data processing module receives data from the point cloud acquisition module and the video acquisition module; its output is connected to the recognition module, and the recognition module's output is connected to the ranging module. The outputs of positioning modules 1 and 2, and the virtual light wall construction module are all connected to the collision analysis module; the output of positioning module 1 is also connected to the virtual light wall construction module. The outputs of the collision analysis module and the ranging module are connected to the alarm module. Specifically: Point cloud acquisition module: used to acquire point cloud data in real time; Video capture module: Used to capture video data in real time; Data processing module: used to align, register and fuse video and point cloud data, and then stitch together data collected by different cameras and LiDAR devices; Recognition module: Cleans and annotates the acquired images and point cloud data, and uses the model to detect obstacles in the video footage and laser point cloud; Distance measuring module: calculates the distance between a fixed obstacle and the boom in real time; Positioning Module 1: Used for collecting positioning data at various corner points in the safe working area, as well as for collecting positioning data at key nodes at the top of the boom; Positioning Module 2: Used to collect positioning data of key nodes of the hook; Collision analysis module: used to calculate the real-time distances between positioning module 1, positioning module 2, and the sides and top of the digital virtual light wall; Alarm module: Triggers visual and auditory alarms based on the distance output by the ranging module and the collision analysis module.

Images