Tower crane hoisting whole process dynamic mapping and sensing coverage method and system

Abstract

The invention relates to the technical field of intelligent control of engineering machinery, and provides a tower crane hoisting whole process dynamic mapping and sensing coverage method and system. The method comprises the following steps: respectively mounting two rotary radars on a rotatable bracket of the tower crane; the current hoisting operation stage of the tower crane is obtained in real time; switching a shooting view angle of the double-radar module according to the hoisting operation stage, and controlling the double-radar module to obtain continuous point cloud data of the corresponding operation stage; tower crane operation parameters are collected, and continuous lifting hook tracks are generated in combination with continuous point cloud data obtained by the double-radar module; identifying a dynamic obstacle and a static background contained in the target monitoring space from the continuous point cloud data, and constructing a construction environment point cloud model; performing space-time registration on the construction environment point cloud model and the physical coordinate system, and establishing a global dynamic map; and a dynamic sensing area coverage map is constructed by fusing the continuous lifting hook track and the global dynamic map, so that the real-time sensing of the space environment of the tower crane hoisting operation is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control of engineering machinery, and more particularly, to a tower crane hoisting whole-process dynamic mapping and perception coverage method and system. BACKGROUND

[0002] As a core heavy lifting equipment in the field of construction engineering, the tower crane is widely used in heavy vertical and horizontal transfer operations in scenarios such as high-rise buildings and bridge engineering. Due to the hoisting process involving high-altitude operation, dynamic transfer of heavy objects, and multi-device cooperation, there are multiple dynamic and static targets such as tower crane jibs, hooks, surrounding buildings, temporary construction machinery, and personnel in the operation space, and the operation safety and efficiency directly determine the project progress and personnel and property safety.

[0003] In the traditional technology, the visual observation of the on-site command personnel (such as the signalman) and the tower crane driver is mainly relied on, the information such as the hook position and the obstacle distance is transmitted through the intercom, the safety of the operation area is artificially judged, and the operation is guided. The field of vision of the operation personnel is easily blocked by the tower crane jib and the building to form a blind area, and is affected by physiological fatigue and subjective judgment errors, and the spatial environment of the hoisting operation process (such as lifting, transfer, and landing) cannot be continuously perceived. In addition, in some existing systems, cameras are arranged on the tower crane jib, the cab, and the operation site to collect two-dimensional image information of the hoisting area, which is transmitted to the display screen of the driving end to assist the driver in observing the hook state and the near-distance environment. However, the driver and the dispatching end can only rely on their own experience to judge the local relationship with the aid of the two-dimensional image information, and it is difficult to form a cognition of the overall spatial relationship, which may lead to decision-making errors in path planning, area blocking, and the like, and cause safety accidents.

[0004] In summary, the traditional tower crane operation mainly relies on artificial observation and experience judgment, and there are visual blind areas and lack of spatial judgment ability, which leads to safety hazards and low efficiency. SUMMARY

[0005] In this context, the embodiments of the present application aim to provide a tower crane hoisting whole-process dynamic mapping and perception coverage method and system, which can realize real-time perception of the spatial environment of the tower crane hoisting operation, and solve the technical problems of visual blind areas and lack of spatial judgment ability in the traditional tower crane.

[0006] In a first aspect of the embodiments of the present application, a tower crane hoisting whole-process dynamic mapping and perception coverage method is provided, comprising:

[0007] Two rotating radars are installed on the rotatable supports of the tower crane respectively to form a double-radar module, and the double-radar module is in communication connection with the control module;

[0008] The control module is used for acquiring a current hoisting operation stage of the tower crane in real time, the hoisting operation stage includes a lifting stage, a rotating stage, an amplitude changing stage and a descending stage; a shooting angle of the dual radar module is switched according to the hoisting operation stage, a key perception angle is used in the lifting stage and the descending stage, a layered perception angle is used in the rotating stage and the amplitude changing stage, and the dual radar module is synchronously controlled to acquire continuous point cloud data corresponding to the operation stage;

[0009] The tower crane operation parameters are collected, and the continuous point cloud data acquired by the dual radar module is combined to calculate a three-dimensional coordinate, a motion speed and a swing angle of a hook in real time, and generate a continuous hook trajectory;

[0010] A target perception algorithm is used to identify dynamic obstacles and static backgrounds contained in a target monitoring space from the continuous point cloud data, and construct a construction environment point cloud model;

[0011] The construction environment point cloud model is time-space registered with a physical coordinate system with the tower crane base as an origin, a global dynamic map containing the target monitoring space is established by using a time-space graph convolution network of a Gauss surface element mapping;

[0012] The continuous hook trajectory and the global dynamic map are fused, a dynamic perception area coverage map is constructed by using a trajectory differential algorithm and a dynamic ontology evolution algorithm, and the dynamic perception area coverage map is used to display the target monitoring space and the construction environment features in the target monitoring space which continuously change with each operation stage in the hoisting operation process.

[0013] In a second aspect of the embodiments of the present application, a tower crane hoisting whole-process dynamic mapping and perception coverage system is provided, comprising:

[0014] A construction unit is configured to install a plurality of rotating radars on rotatable supports of a tower crane to form a dual radar module, and the dual radar module is in communication connection with a control module;

[0015] A switching unit is configured to acquire a current hoisting operation stage of the tower crane in real time by using the control module, the hoisting operation stage includes a lifting stage, a rotating stage, an amplitude changing stage and a descending stage; a shooting angle of the dual radar module is switched according to the hoisting operation stage, a key perception angle is used in the lifting stage and the descending stage, a layered perception angle is used in the rotating stage and the amplitude changing stage, and the dual radar module is synchronously controlled to acquire continuous point cloud data corresponding to the operation stage;

[0016] A trajectory tracking unit is configured to collect tower crane operation parameters, combine the continuous point cloud data acquired by the dual radar module, calculate a three-dimensional coordinate, a motion speed and a swing angle of a hook in real time, and generate a continuous hook trajectory;

[0017] The point cloud modeling unit is configured to identify dynamic obstacles and static backgrounds contained in the target monitoring space from the continuous point cloud data by using a target perception algorithm, and construct a construction environment point cloud model.

[0018] The space-time registration unit is configured to take the tower base as the origin of a physical coordinate system, perform space-time registration on the construction environment point cloud model and the physical coordinate system, and establish a global dynamic map containing the target monitoring space by using a space-time graph convolution network based on a Gaussian surface element mapping method.

[0019] The coverage map generation unit is configured to fuse the continuous hook trajectory and the global dynamic map, and construct a dynamic perception area coverage map by using a trajectory differential algorithm and a dynamic ontology evolution algorithm. The dynamic perception area coverage map is used to display the target monitoring space and the construction environment features in the target monitoring space, which continuously change with each operation stage in the hoisting operation process.

[0020] In a third aspect of the embodiments of the present application, a terminal device is provided, which comprises at least one processor, a memory and an input-output unit. The memory is configured to store a computer program, and the processor is configured to invoke the computer program stored in the memory to execute the tower crane hoisting whole-process dynamic mapping and perception coverage method according to any one of the first aspect.

[0021] In a fourth aspect of the embodiments of the present application, a computer readable storage medium is provided, which comprises instructions that, when executed on a computer, cause the computer to execute the tower crane hoisting whole-process dynamic mapping and perception coverage method according to any one of the first aspect.

[0022] In a fifth aspect of the embodiments of the present application, a computer program product is provided, which comprises a computer program that, when executed by a processor, implements the tower crane hoisting whole-process dynamic mapping and perception coverage method according to any one of the first aspect.

[0023] According to an embodiment of the present application, a tower crane hoisting whole-process dynamic mapping and perception coverage method and system. First, two rotating radars are installed on the rotatable support of the tower crane to form a dual-radar module, and the dual-radar module is in communication connection with the control module. Then, the control module is used to obtain the current hoisting operation stage of the tower crane in real time, and the hoisting operation stage includes the lifting stage, the rotating stage, the amplitude changing stage and the descending stage. Next, the shooting angle of the dual-radar module is switched according to the hoisting operation stage, the key perception angle is used in the lifting stage and the descending stage, and the layered perception angle is used in the rotating stage and the amplitude changing stage, and the dual-radar module is controlled synchronously to obtain continuous point cloud data corresponding to the operation stage. The tower crane operation parameters are collected, the continuous point cloud data obtained by the dual-radar module is combined, the three-dimensional coordinates, the motion speed and the swing angle of the hook are calculated in real time, and the continuous hook trajectory is generated. A target perception algorithm is used to identify dynamic obstacles and static backgrounds contained in the target monitoring space from the continuous point cloud data, and a construction environment point cloud model is constructed. The construction environment point cloud model is time and space registered with the physical coordinate system with the tower crane base as the origin, a time and space convolution network of Gaussian surface element mapping is used to establish a global dynamic map containing the target monitoring space. Finally, the continuous hook trajectory and the global dynamic map are fused, a dynamic perception area coverage map is constructed through a trajectory differential algorithm and a dynamic ontology evolution algorithm, and the dynamic perception area coverage map is used to display the target monitoring space and the construction environment features in the target monitoring space which continuously change with each operation stage in the hoisting operation process. Through the cooperative operation of the dual-radar module and the dynamic mapping technology, the dynamic perception area coverage map is dynamically constructed, the whole-process coverage of the hoisting operation process is realized, the perception ability of the tower crane in the hoisting operation process is greatly improved, and the hoisting intelligent level and the operation safety redundancy capability of the construction site are further improved. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 A flowchart of a tower crane hoisting whole-process dynamic mapping and perception coverage method shown in the present application;

[0025] Figure 2 A schematic diagram of a tower crane hoisting whole-process dynamic mapping and perception coverage method shown in the present application;

[0026] Figure 3 A structural schematic diagram of a tower crane hoisting whole-process dynamic mapping and perception coverage system shown in the present application;

[0027] Figure 4 A structural schematic diagram of a medium of an embodiment of the present application is shown in the present application. DETAILED DESCRIPTION

[0028] Reference will now be made to Figure 1 ,Figure 1 A flowchart of a tower crane hoisting whole-process dynamic mapping and perception coverage method provided by an embodiment of the present application is shown. It should be noted that the embodiments of the present application can be applied to a tower crane operation scene.

[0029] To solve the above at least one technical problem, the embodiments of the present application provide a tower crane hoisting whole-process dynamic mapping and perception coverage method and system. Specifically, for the combination of a double-radar module and a rotatable support, by monitoring the hoisting operation stage in real time and switching the double-radar view angle of the corresponding stage, the double-radar module can collect continuous point cloud data that continuously and without dead angle covers the operation range in the whole process of the tower crane operation. This dynamic adjustment strategy effectively avoids the technical problem of visual blind area in the traditional single-radar or fixed-view-angle multi-radar scheme. Secondly, combined with the fusion calculation of the tower crane operation parameters and the continuous point cloud data, the real-time calculation accuracy of the hook three-dimensional coordinates is further improved. The trajectory generation algorithm effectively suppresses noise interference through trajectory differential processing, ensuring the accuracy of the motion state prediction. The target perception algorithm intelligently classifies dynamic obstacles and static backgrounds, and cooperates with the Gaussian face element mapping technology of the spatio-temporal graph convolution network to realize the dynamic update of the construction environment point cloud model. Through the point cloud fusion strategy of error weight accumulation, the calculation complexity is reduced while the detail features are preserved, improving the calculation efficiency and helping to realize real-time perception. The spatio-temporal registration mechanism with the tower crane base as the origin effectively solves the spatial reference unification problem of multi-source data, so that the constructed global dynamic map can accurately reflect the environmental evolution of the hoisting operation area. The finally generated dynamic perception area coverage map realizes dynamic early warning and safety boundary adaptive adjustment of the dangerous area by fusing the hook trajectory and the three-dimensional environment model, which can greatly shorten the collision warning response time and improve the protection efficiency of the anti-collision system.

[0030] Figure 1 The flow of a tower crane hoisting whole-process dynamic mapping and perception coverage method provided by an embodiment of the present application is shown, which includes:

[0031] In step S101, two rotating radars are respectively installed on the rotatable supports of the tower crane to form a double-radar module, and the double-radar module is in communication connection with the control module.

[0032] In the embodiments of the present application, two groups of rotatable radar sensors are installed at key positions of the tower crane, and the radar angle is dynamically adjusted through the rotatable support (such as a motorized gimbal) to construct a double-radar cooperative perception hardware system (such as shown in the figure) with angle adjustment capability. Figure 2 The two groups of radars are respectively defined as a first rotating radar and a second rotating radar (non-fixed primary and secondary, dynamically switchable), and are in bidirectional communication with the control module (such as an industrial-grade PLC, an edge computing controller, integrated with or connected to multiple types of sensors) to realize data interaction and control instruction transmission.

[0033] Alternatively, in the dual-radar module, the two rotating radars can scan synchronously at a frequency of 10Hz. For example, based on the technical parameters of TF-ALS-LIDAR-01 (detection distance ≥150 meters, accuracy ≤3 cm, angular resolution ≤0.05°), a rotating bracket meeting industrial-grade protection standards (IP67) is selected. The bracket material must have corrosion resistance, and the rotating mechanism uses a high-precision harmonic reducer to ensure a turning accuracy of 0.1 degrees. The dual radars are distributed and installed on both sides of the bracket, rigidly connected to the tower crane boom through a customized structure. Alternatively, the installation angle needs to be optimized using 3D modeling based on the tower crane's operating radius to ensure overlapping coverage areas and eliminate blind spots.

[0034] For example, the communication connection adopts a dual-redundant architecture of CAN bus and Ethernet. Each radar is configured with an independent CAN channel to transmit point cloud data, while gigabit-level data interaction between the control module and the radar is achieved through an industrial switch. The control module uses a 32-bit ARM processor, runs a real-time operating system (RTOS), and configures radar parameters (such as scanning frequency and sampling rate) through the MODBUS-TCP protocol. It also establishes a data buffer queue to process point cloud data streams exceeding 200,000 points per second. Dynamic calibration is required during installation. In a static state, a standard reflector (99% reflectivity) is used to calibrate the angle deviation. During dynamic testing, data from the tower crane tilt sensor is synchronously acquired to establish a real-time transformation model between the rotating support angle and the radar coordinate system, compensating for measurement drift caused by mechanical vibration. Ultimately, a 3D scanning system with a spatial resolution of 0.1 degrees is formed, which can generate more than 100,000 effective data points in a single scan, meeting the high-precision positioning requirements of the tower crane within a long-distance operating radius.

[0035] Step S102: The control module obtains the current hoisting operation stage of the tower crane in real time. In this embodiment, the hoisting operation stage includes the lifting stage, slewing stage, luffing stage, and descent stage.

[0036] Specifically, step S102 achieves accurate identification and dynamic switching of tower crane hoisting operation stages through multi-source sensor data fusion and kinematic model analysis. The control module uses tower crane operating parameters as the core basis, combined with lidar point cloud features and environmental perception data, to establish a stage feature discrimination model. During the hoisting stage, the weight sensor's sudden change signal at the moment the hook leaves the ground is monitored in real time. When the load weight reaches 90% of the rated value and the hoisting encoder speed exceeds 0.3 m / s, the hoisting stage is determined to have begun. In this case, the dual radar modules switch to a focus perception mode, performing a 180-degree fan-shaped scan centered on the hook, generating over 5000 three-dimensional point cloud data points per second. Combined with the tilt angle data from the IMU module, the Kalman filter algorithm is used to calculate the hook's spatial attitude, with the error controlled within ±2 cm. For example, in a building material hoisting scenario, when the hook carries 5 tons of steel bars 0.5 meters off the ground, the sling stress analysis algorithm is automatically triggered, using point cloud density changes to determine whether the load is swaying.

[0037] Upon entering the slewing or luffing phase, the control module establishes a dynamic model of rotational angular velocity and angular acceleration based on the slewing mechanism motor current fluctuation characteristics and tower top displacement data. When the slewing bearing torque exceeds a set threshold and the absolute value of the tower displacement gradient slope is less than 0.05 degrees per second, the system determines that it has entered the slewing or luffing phase. At this time, the dual radar modules employ a layered scanning strategy: the upper radar extracts environmental features at a resolution of 0.02 degrees, and the lower radar detects obstacles at a resolution of 0.1 degrees, with the scanning frequency simultaneously increased to 30Hz. For example, if temporary material is piled up within a 15-meter slewing radius of the tower crane, a point cloud registration algorithm identifies obstacles with a diameter exceeding 50cm, triggering a path correction command to reduce the rotational speed to 65% of the rated value. Simultaneously, a vibration compensation algorithm eliminates point cloud distortion caused by the robotic arm's rotation.

[0038] The descent phase identification is based on a dual criterion of hook height change rate and ground feature matching degree. When the lifting encoder detects a descent speed below 0.15m / s and the ground feature matching similarity in the LiDAR point cloud exceeds 85%, the descent phase is determined to begin. At this time, a high-precision positioning mode is activated, fusing RTK differential positioning data and visual SLAM algorithms to improve the hook's 3D coordinate positioning accuracy to ±3cm. For example, in precast component installation operations, when the hook approaches the target position by 20cm, the edge contour of the precast slab is identified through point cloud segmentation algorithms, and spatial conflict detection is performed in conjunction with the BIM model, automatically adjusting the landing point to a safe area. The entire phase switching process achieves seamless transition through a state machine model, with an average state recognition latency of less than 50ms, ensuring the continuity of perception coverage and timely control response in all aspects of the lifting operation.

[0039] Step S103: Switch the shooting angle of the dual radar modules according to the hoisting operation stage, and synchronously control the dual radar modules to acquire continuous point cloud data of the corresponding operation stage. In this embodiment, a key perception angle is used in the hoisting and descent stages, and a layered perception angle is used in the slewing and luffing stages.

[0040] Understandably, step S103 enhances environmental perception throughout the entire hoisting operation process through the coordinated optimization of a dynamic perspective switching mechanism and a point cloud data acquisition strategy. During the lifting phase, the dual radar modules construct a fan-shaped key perception area centered on the hook. The upper radar adopts a narrow beam scanning mode, compressing the scanning angle to within ±15 degrees, while simultaneously increasing the point cloud density to 200 measuring points per square meter, accurately capturing changes in the spatial relative position of the hook and the hoisted object. For example, in a precast floor slab hoisting scenario, when the hook carries an 8-ton component 0.3 meters off the ground, the forward radar performs continuous scanning with a resolution of 0.03 degrees, generating millimeter-level point cloud data of the object's edge contour. Combined with a tilt sensor, the coordinate system is corrected in real time to ensure that an audible and visual alarm is triggered when the object's swing amplitude exceeds a safety threshold. At this time, the rear radar simultaneously performs a low-frequency environmental scan, capturing static obstacle features within a 5-meter radius with a resolution of 0.1 degrees, forming complementary dual-channel data.

[0041] Upon entering the slewing or luffing phase, the dual radar modules switch to a layered scanning mode. The upper radar performs 360-degree environmental modeling with a 0.01-degree super-resolution, generating a high-precision point cloud of over 150,000 points per second. The lower radar employs a step-by-step scanning strategy, focusing on covering the area within a 10-meter slewing radius of the tower crane at 0.5-degree intervals. For example, in a scenario of multiple tower cranes operating simultaneously, when the tower crane rotates horizontally at 1.2 r / min, the upper radar uses multi-echo fusion technology to identify the metallic reflection characteristics of adjacent tower crane booms, constructing a 3D point cloud model including material properties. Simultaneously, the lower radar detects scaffolding components stacked on the ground and uses a point cloud clustering algorithm to distinguish hard obstacles with a diameter greater than 30 cm, triggering the path planning system to shift the rotation trajectory by 23 degrees to avoid collisions. At this time, the control module establishes a spatiotemporal synchronization compensation model based on the radar scanning frequency difference to eliminate point cloud distortion caused by changes in rotational angular velocity.

[0042] During the descent phase, the dual radar modules revert to a wide field-of-view configuration. The forward radar performs 3D reconstruction of the hook landing area with a coverage range of ±30 degrees, while the rearward radar simultaneously acquires ground elevation change data. For example, in steel structure installation, when the hook descends to 50cm from the target support, the forward radar uses phase modulation technology to analyze the spatial attitude of the steel structure connecting plate, generating a point cloud map containing 2000 feature points. This point cloud map is then matched with the BIM model at the millimeter level. Simultaneously, the rearward radar detects areas with ground settlement exceeding 5mm and generates a local terrain compensation model using interpolation algorithms, ensuring that the final hook positioning error is controlled within ±3cm. Throughout the process, the viewpoint switching delay of the dual radars is controlled within 80ms, and the point cloud data overlap rate remains at a preset ratio, forming a continuous and complete environmental perception sequence.

[0043] Optionally, in the key perception perspective, the first rotating radar is used to scan the three-dimensional space during the vertical lifting, lowering, and horizontal transfer of the hook. The second rotating radar is used to scan the micro-environment space of the hook's vertical path and landing area. In the layered perception perspective, the first rotating radar is used to scan the mid-to-long-range space at the front end of the boom, the side of the boom, and the mid-range space where the hook is located. The second rotating radar is used to scan the near-range space where the core lifting area is located. Based on this assumption, in step S103, switching the shooting perspective of the dual radar modules according to the lifting operation stage includes: when the lifting operation stage is in the lifting or lowering stage, or when the hook lifting height is detected to be lower than a preset lifting height threshold and the horizontal movement range is less than a preset horizontal range threshold, switching to the key perception perspective. When the lifting operation stage is in the slewing or luffing stage, or when the hook is detected to be performing long-distance horizontal transfer or large-range slewing operations, switching to the layered perception perspective.

[0044] Specifically, step S103 achieves refined environmental perception throughout the entire hoisting operation process by dynamically adjusting the spatial scanning strategy of the dual radar modules. Under the key perception perspective, the first rotating radar constructs a three-dimensional scanning field with the hook as the core. Its scanning range covers the three-dimensional space of the hook's vertical movement trajectory and horizontal transfer path, capturing real-time changes in the hoisted object's attitude through high-frequency point cloud acquisition. For example, in the hoisting operation of precast floor slabs, when the hook is vertically lifting an 8-ton component, the first radar performs a dense scan of a 0.5-meter radius area around the hoisted object with a resolution of 0.03 degrees, generating a three-dimensional point cloud map containing 2000 feature points. Combined with tilt sensor data, the coordinate system is corrected in real time to ensure that an audible and visual alarm is triggered when the hoisted object's swing amplitude exceeds a safety threshold. The second rotating radar simultaneously focuses on the micro-environment space below the hook's vertical path, using a narrow-beam scanning mode to perform sub-centimeter-level modeling of the ground landing area. When the hook approaches the target support, this radar can identify the protruding features of the embedded parts in the ground with an accuracy of ±2cm, providing data support for precise hook placement.

[0045] After entering the layered perception perspective, the first rotating radar switches to a mid-to-long-range environmental scanning mode, extending its scanning range to a 20-meter area at the front of the crane boom and a 15-meter area to the side. It constructs a three-dimensional topological model of obstacles surrounding the crane boom using multi-echo fusion technology. For example, in a scenario of multiple tower cranes operating simultaneously, when the tower crane rotates horizontally at 1.2 r / min, the first radar detects the metallic reflection characteristics of adjacent tower crane booms using the Doppler effect, identifying the outline of a standard tower crane section 50 meters away. It then uses a spatiotemporal graph convolutional network to establish a dynamic obstacle trajectory prediction model. The second rotating radar focuses on the mid-range space where the hook is located, employing a step-by-step scanning strategy to provide layered coverage of the core lifting area. When the hook makes a long-distance horizontal transfer, this radar performs layered scanning of an area within a 10-meter radius at 0.1-degree intervals, with each layer spaced 0.5 meters apart, forming a spatial feature dataset containing 5000 measuring points. This effectively identifies potential obstacles such as temporarily stacked scaffolding components.

[0046] The phase switching mechanism achieves intelligent decision-making through multi-source information fusion. When the hook lifting height is detected to be less than 3 meters and the horizontal movement range is less than 5 meters, it automatically switches to the key perception perspective. At this time, the dual radar scanning frequency is synchronously increased to 40Hz to ensure the dynamic tracking accuracy of the suspended object during low-altitude operations. For example, in steel structure installation operations, when the hook descends to 50cm from the target support, the second radar under the key perception perspective uses phase modulation technology to analyze the spatial attitude of the steel structure connection plate, generating a point cloud map containing 2000 feature points, and performing millimeter-level matching with the BIM model. When the tower crane performs large-range slewing operations, it determines to enter the layered perception mode by fusing data from the slewing bearing torque sensor and encoder. At this time, the scanning frequency of the first radar is increased to 60Hz, dynamically modeling obstacles within 50 meters of the front end of the boom, and predicting their movement trajectory using the Kalman filter algorithm, triggering the path correction command in advance to offset the rotation trajectory by a preset degree to avoid collision.

[0047] As an optional embodiment, in step S103, a hybrid target perception network is used to identify dynamic obstacles and static backgrounds contained in the target monitoring space based on the adjusted continuous point cloud map. Geometric features of the dynamic obstacles, geometric features of the static background, and real-time motion parameters of the dynamic obstacles are extracted from the adjusted continuous point cloud map. This includes: processing the adjusted continuous point cloud map through a semantic perception feature enhancement module; using an improved BEBLID descriptor and a FasterNet backbone network; fusing multi-scale point cloud density information to construct an adaptive feature pyramid; enhancing static structural boundary features through an edge-aware loss function; and outputting an enhanced point cloud feature map. Based on the aforementioned enhanced point cloud feature map, the ISS algorithm and normal vector constraints are used to calculate the normal vector and curvature of the point cloud. ISS feature scores are generated through neighborhood covariance analysis, and static feature points are screened. The DBSCAN clustering algorithm is used to remove discrete noise points. Multi-view geometric constraints are introduced to eliminate dynamic interference points. The curvature threshold is dynamically adjusted according to the weather information. Geometric features of the static background are extracted, and static background geometric features are simultaneously eliminated to obtain dynamic candidate point clouds. Based on the aforementioned dynamic candidate point cloud, a dynamic object detection network is constructed using a hybrid architecture of YOLOv5s and PointNet. A joint 2D and 3D detection head outputs the category, bounding box, and point cloud clustering mask of the dynamic obstacle to extract its geometric features. Motion optical flow is calculated using the Farneback optical flow method and combined with the point cloud data to obtain the initial velocity. The position, velocity, and acceleration parameters of the dynamic obstacle are then optimized using a Kalman filter algorithm, serving as the real-time motion parameters of the obstacle.

[0048] Step S103 achieves accurate separation of dynamic obstacles from static backgrounds and extraction of motion parameters through multi-level feature enhancement and hybrid architecture design. The semantic-aware feature enhancement module first preprocesses the adjusted continuous point cloud map, constructs a local feature matching framework using an improved BEBLID descriptor, and enhances the local geometric description capability of the point cloud through cosine similarity calculation between feature vectors. This module introduces the FasterNet backbone network to replace the traditional PointNet++ structure, utilizing its partial convolution (PConv) operation to reduce redundant computation, improving computational efficiency by more than 40% while maintaining feature expressiveness. The adaptive feature pyramid dynamically adjusts the contribution weights of features at different levels through a multi-scale point cloud density-weighted fusion strategy, preserving key structural features even in complex occlusion scenarios. For example, in tower crane cluster operation scenarios, when adjacent tower crane booms form a dense metal reflection area, this module effectively suppresses background noise interference and accurately extracts the edge contour features of obstacles within a 0.5-meter radius around the hook through a density-aware feature fusion strategy.

[0049] In the static background feature extraction stage, the ISS algorithm combines normal vector constraints to filter feature points. It constructs a 3D spatial feature scoring model by calculating the curvature variance and neighborhood covariance matrix eigenvalues ​​of each point. When the standard deviation of curvature is below a set threshold and the consistency of the principal curvature directions exceeds 80%, it is determined to be a static structural feature point. The DBSCAN clustering algorithm performs noise removal based on density peaks, and its neighborhood radius parameter is dynamically adjusted according to weather information. In rainy or foggy weather, the radius is expanded to 0.3 meters to compensate for the sparsity of the point cloud, while in sunny weather it is reduced to 0.1 meters to improve segmentation accuracy. Multi-view geometric constraints project the point cloud to multiple virtual camera views, utilizing perspective invariance to remove dynamic interference points, such as temporary objects like protective nets fluttering in strong winds; these are automatically filtered out when the difference in multi-view projection features exceeds a threshold.

[0050] The dynamic object detection network adopts a hybrid architecture of YOLOv5s and PointNet. The front-end YOLOv5s is responsible for quickly generating 2D candidate boxes, while the back-end PointNet extracts the 3D point cloud features of the candidate regions. The 2D detection head generates the initial target position through the anchor box mechanism, and the 3D detection head optimizes the bounding box size by combining normal vectors and curvature information. The final output includes the detection results containing class probability, spatial coordinates, and point cloud mask. In the precast floor slab hoisting scenario, when the hoisted object passes through the scaffolding area, the network detects a 50cm diameter steel pipe protrusion through a multi-scale feature pyramid. Combined with the point cloud mask, the network accurately segments the obstacle geometry, improving the recognition accuracy. The motion parameter estimation module uses the Farneback optical flow method to calculate the motion vector field between adjacent frame point clouds. The point cloud displacement relationship is established through dense optical flow matching. The initial velocity field is optimized by Kalman filtering, which can effectively suppress the influence of sensor noise. For example, in long-distance horizontal transfer operations of tower cranes, when the hook carries an 8-ton component and moves at a speed of 0.5 m / s, the three-dimensional motion trajectory of the component is calculated in real time, the position prediction error is controlled within ±8 cm, and the speed estimation fluctuation is less than 0.03 m / s, providing reliable data support for dynamic obstacle avoidance.

[0051] As an optional embodiment, in step S103, after calculating the motion optical flow based on the Farneback optical flow method and obtaining the initial velocity by combining it with point cloud data, and optimizing the position, velocity, and acceleration parameters of the dynamic obstacle using the Kalman filter algorithm as the real-time motion parameters of the dynamic obstacle, the real-time motion parameters and geometric features of the dynamic obstacle can also be used to extract the radial velocity of the dynamic obstacle by combining Doppler feature encoding, constructing a spatiotemporal correlation graph containing timestamps, coordinates, and velocity vectors, with the edge weights being the motion consistency probability. The detection sensitivity is dynamically adjusted according to the obstacle velocity. When the dynamic obstacle approaches the extracted static background, a joint feature optimization process is triggered to achieve interactive optimization of the static background and dynamic obstacle features.

[0052] Step S103 utilizes multimodal feature fusion and spatiotemporal modeling techniques to achieve high-precision analysis and adaptive perception optimization of the motion state of dynamic obstacles. After obtaining the position, velocity, and acceleration parameters of the dynamic obstacles, their Doppler feature codes are further extracted. By analyzing the Doppler frequency shift characteristics of the millimeter-wave radar echo signal, the radial velocity component of the obstacle relative to the radar is accurately calculated. For example, in a scenario of tower crane cluster cross-operation, when adjacent tower crane booms move towards each other at a speed of 2.5 m / s, the Doppler feature coding module analyzes the phase change of the echo signal and calculates that the relative radial velocity between the two reaches 4.8 m / s. This data is then weighted and fused with the point cloud motion parameters to form a spatiotemporal correlation graph node containing timestamps, three-dimensional coordinates, and velocity vectors. The edge weights between nodes are predicted using Kalman filtering and the difference between the prediction and the actual observation are used to calculate the motion consistency probability. When the difference exceeds 15%, an edge weight decay mechanism is triggered to ensure the timeliness of the dynamic topology.

[0053] For scenarios where obstacles are close to a static background, an adaptive sensitivity adjustment algorithm is employed. This algorithm dynamically adjusts detection parameters by analyzing the angle between the obstacle's velocity vector and the static background's normal vector. When the angle is less than 30° and the relative velocity exceeds 1 m / s, a joint feature optimization process is triggered. Specifically, firstly, the point cloud features of the dynamic obstacle are cross-modal matched with the ISS features of the static background, and a graph attention network is used to calculate feature similarity weights. Secondly, a spatiotemporal graph convolutional network is used to aggregate the motion features of neighboring nodes, establishing a spatiotemporal correlation model that includes velocity trend prediction. Finally, an adversarial training strategy is used to optimize the feature fusion weights. In the case of steel structure factory hoisting, this mechanism improves the accuracy of collision warnings between the hook and the steel beam. For example, when a precast wall panel is hoisted to a distance of 0.8 meters from a concrete column, a sudden change in the angle between the wall panel edge point cloud and the column surface normal vector is detected in real time, triggering a multi-scale feature enhancement module. This module extracts contact surface geometric features with a precision of 0.1 meters through an expanded convolutional network, verifies the collision risk using a BIM model, and finally generates a dynamic safety plan that includes avoidance paths.

[0054] Step S104: Collect tower crane operating parameters, combine them with continuous point cloud data obtained from dual radar modules, calculate the three-dimensional coordinates, movement speed and swing angle of the hook in real time, and generate a continuous hook trajectory.

[0055] Understandably, step S104 utilizes multi-source data fusion and kinematic modeling techniques to achieve real-time tracking and trajectory generation of the hook's motion state. First, tower crane operating parameters are collected, including basic motion parameters such as hoisting mechanism speed, slewing bearing angular velocity, and trolley travel encoder pulse signals. Simultaneously, continuous point cloud data generated by dual radar modules is acquired. The dual radars employ a heterogeneous layout: the forward radar covers the hook's vertical motion trajectory with a narrow beam, while the rearward radar monitors the horizontal swing range with a wide field of view. During the data fusion phase, a timestamp alignment algorithm synchronizes the tower crane operating parameters and radar point cloud data to millisecond-level accuracy, eliminating errors caused by mechanical transmission delays. For example, in a precast floor slab hoisting scenario, when the hook carries an 8-ton component for vertical lifting, the forward radar uses phase modulation technology to analyze the relative position offset between the hook's center point and the hoisting rope. Combined with the linear velocity data output by the tower crane's hoisting encoder, a motion compensation algorithm is used to eliminate point cloud distortion caused by rope swing, and the hook's three-dimensional coordinates are calculated in real time. At this time, the rearward radar synchronously scans the 0.5-meter radius area around the hook, analyzes the hook swing amplitude by changing the point cloud density, and triggers a dynamic filtering mechanism when the swing angular velocity exceeds a preset threshold, using a particle filtering algorithm to optimize the swing angle estimate.

[0056] The trajectory generation module constructs a motion state equation based on a Kalman filter framework, recursively calculating the hook's three-dimensional coordinates, velocity vector, and angular displacement parameters. Under complex working conditions, tower crane structural parameters are introduced to establish a motion constraint model. For example, during overlapping operations of multiple tower cranes, geometric constraints are constructed using the tower crane boom length and slewing center coordinates to correct the hook's trajectory in real time. For instance, when the tower crane performs boom slewing operations, if the rear radar detects that the hook's horizontal offset due to inertia exceeds a safety threshold, historical trajectory data and current angular velocity parameters are immediately retrieved. The inverse kinematics algorithm is used to calculate the actual swing center coordinates of the hook, and combined with the tower crane's structural features in the BIM model, a dynamic obstacle avoidance trajectory including a safety buffer is generated.

[0057] Step S105: Using a target perception algorithm, identify dynamic obstacles and static backgrounds contained in the target monitoring space from the continuous point cloud data, and construct a point cloud model of the construction environment.

[0058] Understandably, step S105 utilizes multimodal perception fusion and 3D modeling technology to achieve dynamic obstacle recognition and spatial structure modeling in the construction environment. In this scenario, the target perception algorithm specifically refers to a multi-layered perception architecture that integrates dual radar point clouds and visual features. Its core lies in achieving semantic segmentation of dynamic obstacles and extraction of static background features through an improved deep learning model. The construction environment point cloud model is a 3D spatial topology constructed based on temporal point cloud data, containing digital representations of dynamic obstacle trajectories, static facility geometry, and environmental physical properties.

[0059] For example, the point cloud data collected by the dual radar modules is first optimized for view alignment and fusion. In scenarios involving overlapping operations of tower cranes, the forward radar scans a 5-meter radius around the hook with a 0.1° angular resolution, while the backward radar covers a 20-meter radius area in wide-angle mode. Multipath interference is eliminated through voxel grid downsampling and feature-level fusion algorithms. When rain or fog is detected, real-time precipitation intensity data is obtained via a meteorological API. The fusion weights are dynamically adjusted based on a point cloud density attenuation model; for example, the forward radar weight is increased to 0.7 in rainy weather to compensate for reduced visibility, while it is decreased to 0.3 in sunny weather to highlight visual features.

[0060] The hybrid target perception network adopts a heterogeneous architecture of YOLOv7 and PointNet++. The front-end YOLOv7 detects candidate obstacle regions through a multi-scale feature pyramid, while the back-end PointNet++ extracts the normal vector and curvature features of the candidate regions. In the prefabricated component hoisting scenario, when the hook passes through the scaffolding area, the network focuses on the edge features of the metal pipes through an attention mechanism. Combined with the point cloud density information after weather compensation, it accurately identifies a 30cm diameter steel pipe protrusion, with its bounding box positioning error controlled within ±5cm. The dynamic obstacle motion parameters are jointly optimized using optical flow and Kalman filtering. For example, in a strong wind environment, when the swaying of the protective net is detected, the optical flow displacement vector and radar velocity measurement results are weighted and fused to reduce the velocity estimation fluctuation amplitude to 0.05m / s.

[0061] In the geometric feature extraction stage, an improved GICP algorithm is used for feature registration, and the surface structural characteristics of the point cloud are described by constructing a local covariance matrix. In the steel structure factory, when the hook approaches the steel beam, point cloud clusters within a 0.2-meter range of the contact surface are extracted, and their principal curvature and normal vector distribution characteristics are calculated to generate a geometric descriptor containing 128-dimensional feature vectors. Static background modeling utilizes the TSDF algorithm to construct an implicit surface of the environment. By dynamically adjusting the voxel resolution, for example, using a 0.1-meter voxel size in key areas (such as within a 5-meter slewing radius of the tower crane) and expanding to 0.5 meters in the outer areas, the BIM model matching accuracy reaches ±3cm. Finally, the construction environment point cloud model is dynamically updated through a spatiotemporal correlation graph. Nodes contain obstacle geometric features, motion states, and environmental semantic labels, and edge weights reflect the spatiotemporal consistency probability, providing a high-precision environmental semantic map for intelligent obstacle avoidance and path planning.

[0062] Optionally, in step S105, the continuous point cloud data collected by the dual radar modules are fused and optimized according to the shooting angle and their corresponding acquisition angles to obtain a continuous point cloud map of the target monitoring space; the weather information of the geographical area where the tower crane is located is acquired in real time, and the point cloud weights under different weather conditions are adjusted by the point cloud density attenuation model; a hybrid target perception network is used to identify the dynamic obstacles and static background contained in the target monitoring space based on the adjusted continuous point cloud map, and the geometric features of the dynamic obstacles, the geometric features of the static background, and the real-time motion parameters of the dynamic obstacles are extracted from the adjusted continuous point cloud map; the improved generalized iterative nearest point (GICP) algorithm and the dynamic fusion truncated symbolic distance function (TSDF) algorithm are used to construct the construction environment point cloud model with the extracted geometric features and real-time motion parameters.

[0063] Specifically, in step S105, during the dual-radar point cloud fusion stage, a heterogeneous sensor spatiotemporal alignment algorithm is employed to map the point cloud data of the forward narrow-beam radar and the backward wide-angle radar to a unified coordinate system. The forward radar captures high-precision motion trajectories in the vertical direction through its narrow-beam characteristics, while the backward radar covers the horizontal space with a wide angle. Multipath interference is eliminated through voxel grid downsampling and feature-level fusion strategies. A weather adaptive adjustment mechanism is introduced to dynamically adjust sensor weights based on meteorological data. An exponential decay model is used to compensate for point cloud quality loss under severe weather conditions, ensuring robustness of perception in complex environments.

[0064] The hybrid target perception network employs a heterogeneous architecture of YOLOv7 and PointNet++ to achieve feature complementarity. The front-end YOLOv7 detects candidate obstacle regions through a multi-scale feature pyramid, while the back-end PointNet++ extracts the normal vectors and curvature features of the candidate regions. The network dynamically focuses on key geometric features through a channel attention mechanism, combined with weather-compensated point cloud density information, to achieve accurate identification and classification of dynamic obstacles. The motion parameters of dynamic obstacles are jointly optimized using optical flow and Kalman filtering, and environmental interference is eliminated using a motion compensation algorithm to establish a dynamic correlation model between the obstacle's motion state and spatial position.

[0065] In the geometric feature extraction stage, an improved generalized iterative nearest-point algorithm is used for feature registration. A local covariance matrix is ​​constructed to describe the surface structure characteristics of the point cloud, and principal component analysis is combined to extract geometric feature descriptors. For static background modeling, a dynamic truncation signed distance function algorithm is employed. An implicit surface for the environment is constructed using adaptive voxel resolution. High-resolution voxels are used in key areas to ensure detailed features, while low-resolution voxels are used in peripheral areas to balance computational efficiency. The environment model is dynamically updated through a spatiotemporal correlation graph. Nodes contain obstacle geometric features, motion states, and environmental semantic labels, and edge weights reflect the spatiotemporal consistency probability, achieving 3D semantic modeling and dynamic evolution of the construction environment.

[0066] The final constructed point cloud model of the construction environment integrates multi-source sensing data, and achieves digital representation of obstacle geometry, motion state, and environmental structure through feature fusion and semantic annotation. The model adopts a hierarchical storage structure, with dynamic obstacle information and static background data stored in separate layers, supporting real-time querying and dynamic updates. This model provides a high-precision environmental semantic map for intelligent obstacle avoidance systems, supporting real-time decision-making for path planning and collision warning.

[0067] Step S106: Using the tower crane base as the origin of the physical coordinate system, the construction environment point cloud model is spatiotemporally registered with the physical coordinate system, and a global dynamic map containing the target monitoring space is established using a spatiotemporal graph convolutional network with Gaussian surface mapping.

[0068] As an optional embodiment, in step S106, a three-dimensional physical coordinate system is established with the geometric center of the tower crane base as the origin. The installation position coordinates and attitude parameters of the dual radar modules are obtained, and the original coordinates of each point in the construction environment point cloud model are mapped to the physical coordinate system through a coordinate transformation matrix. Based on the timestamps of the point cloud data collected by each rotating radar, a time interpolation algorithm is used to synchronize the point cloud data at different times, eliminating the spatiotemporal deviation caused by radar data acquisition delay, and aligning the construction environment point cloud model with the physical coordinate system in time and space. The aligned point cloud model is discretized into a Gaussian grid using a Gaussian element mapping method. Each Gaussian grid stores the mean of static background geometric features, the probability distribution of dynamic obstacle motion parameters, and the point cloud density confidence level for the corresponding spatial region. A spatiotemporal graph convolutional network is constructed, using a Gaussian grid of facets as graph nodes. A Gaussian distribution model is used to uniformly represent the stored geometric feature mean, motion parameter probability distribution, and point cloud density confidence level, serving as the node features for each graph node. Node features include the facet's geometric parameters, timestamp, and dynamic attribute labels. Edge weights are set based on the spatial distance and temporal correlation between facets. The spatiotemporal graph convolutional network fuses the Gaussian facet node features over continuous time, updating the motion parameter distribution of facets containing dynamic obstacles in real time while preserving the geometric features of static background facets. This generates a global dynamic map containing temporal evolution information, which dynamically updates the environmental features of the target monitoring space as the hoisting operation progresses.

[0069] Specifically, step S106 achieves global perception and dynamic evolution modeling of the construction environment through spatiotemporal registration and dynamic modeling techniques. A three-dimensional physical coordinate system is constructed with the geometric center of the tower crane base as the origin. Coordinate transformation relationships are established through the installation parameters of the dual radar modules, mapping each data point in the point cloud model to a unified coordinate system. This process uses quaternion rotation and translation compensation algorithms to eliminate sensor installation deviations. For example, when the dual radars are installed at a 90-degree angle, the rotation matrix is ​​calculated using the Rodrigues formula, and spatial alignment is achieved by combining it with the Euler angle parameters of the base coordinate system. To address the issue of asynchronous radar data, a cubic spline interpolation algorithm is used to resample the point cloud at different times. When the hook moves at a speed of 2.5 m / s, the time interpolation accuracy can reach the millisecond level, effectively eliminating spatiotemporal misalignment caused by differences in sensor refresh rates.

[0070] The Gaussian surface mapping module discretizes the registered point cloud space into a regular grid, with each grid representing environmental features through a Gaussian distribution function. In areas with dense dynamic obstacles, the grid resolution is increased to 0.1 meters to capture motion details, while in static areas it is expanded to 0.5 meters to balance computational efficiency. Each surface stores the velocity vector probability distribution of dynamic obstacles, the mean geometric features of the static background, and the point cloud density confidence score. For example, within a 5-meter slewing radius of a tower crane, the mean geometric features of the surface include normal vector and curvature information, and the dynamic parameter distribution is updated using Kalman filtering. The spatiotemporal graph convolutional network treats the surface as graph nodes, fusing Gaussian distribution parameters and timestamp information into node features. Edge weights are calculated using spatial Euclidean distance and time difference, forming a dynamic spatiotemporal relational structure.

[0071] The network employs a hierarchical convolutional architecture. The spatial layer aggregates features from neighboring nodes through a graph attention mechanism, while the temporal layer captures temporal dependencies using dilated causal convolution. In scenarios involving multiple tower cranes operating simultaneously, when adjacent tower crane booms enter the monitoring area, the network enhances the correlation between nodes in the interaction area through a dynamic edge weight adjustment mechanism, updating the distribution of obstacle motion parameters in real time. Static background elements utilize a feature freezing strategy, allowing only the mean of geometric features to adjust slowly with environmental changes, while the motion parameter distribution of dynamic elements is updated every 0.5 seconds to ensure real-time performance and stability. The resulting global dynamic map, calculated recursively through a spatiotemporal graph convolutional network, achieves deep fusion of obstacle trajectory prediction and static environmental semantics, providing an environmental model containing spatiotemporal evolution features for intelligent obstacle avoidance of tower cranes.

[0072] Step S107: The continuous hook trajectory is fused with the global dynamic map, and a dynamic sensing area coverage map is constructed using a trajectory differentiation algorithm and a dynamic ontology evolution algorithm.

[0073] In this embodiment of the application, the dynamic sensing area coverage map is used to display the target monitoring space that changes continuously with each stage of the hoisting operation, as well as the construction environment characteristics within the target monitoring space.

[0074] Understandably, step S107 constructs a perception coverage map reflecting the spatiotemporal evolution characteristics of the construction process by fusing the dynamic trajectory of the hoisting operation with the 3D environmental model. The continuous hook trajectory is generated by Kalman filtering and spline interpolation algorithms using hook position data collected by dual radar modules, and includes a 3D coordinate sequence with timestamps, reflecting the continuous movement path of the hook during the hoisting operation. The global dynamic map integrates point cloud density, obstacle motion parameters, and environmental semantic information through a spatiotemporal graph convolutional network to form a spatiotemporal correlation model that includes the geometric features of static facilities and the probability distribution of dynamic obstacles.

[0075] In this embodiment, the trajectory differentiation algorithm uses a linear tracking differentiator to perform kinematic analysis on the hook trajectory, and transforms discrete position data into a continuous and differentiable sequence of motion parameters by constructing a state observer. The trajectory differentiation algorithm suppresses high-frequency noise interference through a pole placement strategy, extracting derivative features such as velocity and acceleration while maintaining trajectory smoothness. For example, it can still accurately output an acceleration estimate on the order of 0.1 m / s² even under sudden hook stop conditions. The dynamic ontology evolution algorithm, based on a change propagation mechanism, identifies entity attributes in the environment model that need updating through ontology difference calculation, and uses a heuristic search strategy to generate change paths with the smallest impact range, ensuring synchronous adjustment of associated semantic labels when dynamic obstacle positions are updated.

[0076] The dynamic sensing area coverage map achieves multi-dimensional information fusion through a rasterized spatiotemporal coordinate system. Each grid cell stores the dynamic obstacle contact probability, static facility damage level, and environmental parameter distribution characteristics. In tower crane group operation scenarios, the hook trajectory is spatiotemporally aligned with the global dynamic map. Convolution operations are used to extract the intersection features between the trajectory and obstacle areas, and a coverage heatmap is generated through dynamic threshold segmentation. The dynamic sensing area coverage map adopts a four-dimensional spatiotemporal coding structure, including spatial coordinates, timestamps, and environmental state vectors, supporting real-time querying of the environmental parameter change trajectory of a specific area over historical periods. This reduces environmental perception latency, improves obstacle recognition accuracy, and provides a spatiotemporally consistent intelligent decision-making basis for tower crane group collaborative operations.

[0077] As an optional embodiment, in step S107, the continuous hook trajectory is spatiotemporally aligned with the global dynamic map, ensuring that the hook trajectory coordinates are consistent with the physical coordinate system of the global dynamic map, and synchronously associating the hook movement sequence with the real-time motion parameters of dynamic obstacles. Based on a trajectory differentiation algorithm, the continuous hook trajectory is differentiated to extract the velocity change rate, acceleration change rate, and swing angle change trend of the hook at different operation stages, generating hook motion characteristic parameters. A dynamic ontology evolution algorithm is used to construct an ontology model containing the lifting operation stage, hook motion characteristic parameters, static background geometric features, and real-time motion parameters of dynamic obstacles. The association weights of each element in the ontology model are dynamically updated according to the operation stage. Centered on the current hook position, and combining the hook motion characteristic parameters and environmental features in the global dynamic map, a dynamic perception boundary is generated through the ontology model. This dynamic perception boundary adjusts in real-time with the hook movement, covering the target monitoring space within a preset safety range of the hook. The dynamic perception boundary is superimposed on the static background and dynamic obstacles in the global dynamic map to generate the dynamic perception area coverage map. The dynamic sensing area coverage map displays in real time the spatial relationship between the hook trajectory and environmental features, early warning information of dynamic obstacles entering the dynamic sensing boundary, and changes in the sensing area at different operation stages.

[0078] Specifically, step S107 achieves global perception and intelligent early warning for hoisting operations through spatiotemporal alignment and dynamic modeling techniques. First, the continuous hook trajectory is spatiotemporally registered with the global dynamic map. A coordinate transformation matrix is ​​used to eliminate sensor installation deviations and data asynchrony issues, ensuring a strict correspondence between the hook's three-dimensional coordinates and the physical coordinate system. During this process, a time interpolation algorithm is used to resample the hook trajectory at sub-second levels, synchronously associating the real-time position and motion parameters of dynamic obstacles to establish a spatiotemporally consistent correlation model.

[0079] The trajectory differentiation algorithm performs kinematic analysis of the hook trajectory based on an adaptive tracking differentiator, transforming the discrete position sequence into a continuously differentiable sequence of motion parameters by constructing a state observer. This algorithm employs a pole placement strategy to suppress high-frequency noise, accurately extracting the rate of change of velocity and acceleration even under conditions of sudden hook stops or changes in direction. For example, when the hook is lifted with an acceleration of 3 m / s², the algorithm can output an acceleration estimate on the order of 0.05 m / s³ in real time. The dynamic ontology evolution algorithm identifies entity attributes in the environment model that need updating through ontology difference calculation. It uses a heuristic search strategy to generate change paths with the smallest impact range, ensuring synchronous adjustment of associated semantic labels when dynamic obstacle positions are updated. For example, when a newly arrived tower crane is detected, the algorithm automatically enhances the association weight of its motion parameters in the environment model.

[0080] The dynamic sensing area coverage map achieves multi-dimensional information fusion through a rasterized spatiotemporal coordinate system. Each raster cell stores the dynamic obstacle contact probability, static facility damage level, and environmental parameter distribution characteristics. Centered on the current position of the hook, the dynamic sensing boundary is generated through an ontology model by combining motion feature parameters extracted by the trajectory differential algorithm with environmental features in the global dynamic map. This boundary uses a probability density function to describe the safe area around the hook. When a dynamic obstacle is detected entering a preset threshold area, a multi-level early warning mechanism is triggered, and the obstacle entry probability distribution is displayed on the coverage map in the form of a heat map. In scenarios of multiple cranes operating simultaneously, the coverage map identifies changes in the sensing area at different stages of the operation through a dynamic threshold segmentation algorithm. For example, it automatically expands the sensing radius to 5 meters during the hoisting and positioning stage, while shrinking it to 2 meters during the rotation process to improve computational efficiency.

[0081] This reduces environmental perception latency to less than 150ms, improves obstacle collision warning accuracy, and shortens dynamic perception boundary adjustment response time to 50ms. Through spatiotemporal alignment and ontological evolution technology, the semantic features and dynamic evolution patterns of the hoisting operation environment can be reflected in real time, providing intelligent decision support with spatiotemporal consistency for tower crane group collaborative operations, and significantly reducing collision risks and operational errors under complex working conditions.

[0082] As an optional embodiment, after step S107, the real-time motion parameters of the hook and the real-time motion parameters of the dynamic obstacles can be called to calculate the Euclidean distance and relative motion trend between the hook and each obstacle. A preset collision probability model is used to quantify the risk value, and the safety risk level is marked on the dynamic perception area coverage map. High-risk areas are marked with dense markers, medium-risk areas with sparse markers, and low-risk areas with safe markers. This collision probability model is constructed based on a distance threshold and relative velocity coupling algorithm. Obstacle classification features are extracted from the construction environment point cloud model. These features include the geometric dimensions and fixed coordinates of static obstacles, the real-time position of dynamic obstacles, and their motion trajectory vectors. Differentiated contour symbols are superimposed on the dynamic perception area coverage map. Static obstacles are marked with solid graphics, and dynamic obstacles are marked with dynamic graphics. The position coordinates of dynamic obstacles are updated synchronously, and a motion trajectory curve of the dynamic obstacle obtained by connecting multiple consecutive frame position coordinate points is generated.

[0083] Specifically, following step S107, a deeper enhancement of construction safety early warning is achieved through dynamic risk quantification and visualization technologies. The motion parameters of the hook and dynamic obstacles are acquired in real time. Based on a spatiotemporally aligned coordinate system, the Euclidean distance and relative motion vector between them are calculated. A distance-velocity coupling algorithm is used to construct the collision probability model, mapping risk factors to spatial geometric relationships through nonlinear functions. For example, when an obstacle enters the hook's swing radius, the model automatically increases the collision probability weight. A multi-level threshold strategy is adopted for risk level classification. Different color coding rules are set based on historical accident data and engineering specifications. The difference in grid density and fill pattern intuitively reflects the risk intensity, enabling operators to quickly identify high-risk areas.

[0084] The obstacle classification feature extraction module separates static and dynamic entity attributes from the point cloud model. Static obstacles are extracted using a region growing algorithm to obtain their geometric contours and coordinate information, while dynamic obstacles are generated using a trajectory prediction algorithm to produce motion vectors. In the visualization stage, symbolic encoding rules are used to differentiate the rendering of different obstacle categories. Static obstacles are represented by fixed color blocks to indicate their geometric shape, while dynamic obstacles are displayed using continuous trajectory lines to show their motion state. The dynamic perception effect is enhanced by combining motion direction arrows and velocity vector lines. Risk level labeling employs dynamic permeation technology. When obstacle risk parameters change, a smooth transition is achieved through color gradients and transparency adjustments to avoid misjudgments caused by abrupt visual changes.

[0085] This shortens risk warning response time and reduces false collision rates. Through multi-dimensional information fusion display, operators' perception efficiency of complex working conditions is improved, especially in scenarios involving multiple tower cranes operating simultaneously, increasing the accuracy of dynamic obstacle trajectory prediction. Differentiated labeling strategies enhance the visualization accuracy of spatial distribution of safety risks, effectively supporting real-time decision-making and path optimization in hoisting operations, and providing a self-adaptive intelligent safety protection system for large-scale construction projects.

[0086] For example, dynamic risk visualization technology is used to intuitively present the construction safety situation. Based on real-time calculation of the spatial relationship and motion state between the hook and obstacles, a multi-level risk quantification model is used to dynamically partition and label the environment. In the dense red grid area, the system updates the obstacle contact probability 200 times per second through high-frequency point cloud sampling and motion trajectory prediction algorithms, combined with Kalman filtering to optimize the relative velocity estimation accuracy, ensuring real-time early warning capability for high-risk areas. In the sparse yellow grid area, an adaptive sampling strategy is adopted. When an obstacle enters the area, the point cloud density is automatically increased to 0.2-meter resolution, and a motion prediction line for the next 3 seconds is generated through trajectory extrapolation algorithm, providing operators with buffer decision-making time. In the green grid area, the environmental baseline characteristics are maintained through spatial interpolation algorithms, and historical data caching technology is used to reduce the frequency of sensor data acquisition and reduce system load.

[0087] This reduces the response time for hazardous area identification and lowers the false alarm rate. The differentiated labeling strategy improves the accuracy of risk area identification for operators, especially in scenarios involving multiple tower cranes operating simultaneously. The dynamic grid adjustment mechanism can automatically expand the coverage area of ​​high-risk areas to an 8-meter radius. Through the coordinated optimization of grid density and color coding, GPU resource consumption is reduced while maintaining a rendering speed of 50 frames per second. The deep integration of the 3D dynamic perception overlay map with the driver's AR display module enables crane operators to quickly locate risk sources through visual focus, shortening the response delay of operational commands and thus improving operational safety and efficiency in complex conditions.

[0088] As an optional embodiment, after step S107, the tower crane's rated slewing radius, lifting height range, and obstacle distribution in the dynamic sensing area coverage map are obtained. A fast randomized expanding tree algorithm is used to calculate the passable path from the hook's current position to the target landing point. In the dynamic sensing area coverage map, the optimal passable path is marked with a solid line, potential passable paths with dashed lines, and inaccessible areas obstructed by obstacles or exceeding the tower crane's performance are marked with another type of solid line. The position coordinates of key nodes in each passable path are marked, and speed limit thresholds for these key nodes are marked based on the hook's motion parameters. The dynamic sensing area coverage map is synchronously updated at a preset frequency, automatically refreshing the safety risk level, obstacle distribution, and each passable path. The dynamic sensing area coverage map is transmitted to the dispatching and operating terminals. The dispatching terminal interface supports the overlay of dynamic sensing area coverage maps from multiple tower cranes. The overlaid dynamic sensing area coverage map is projected into a three-dimensional display space to assist users in global path planning and high-risk area closure decisions. The driver interface displays the safety risk level of key areas around the hook to the user based on the dynamic perception area coverage map, and enlarges the risk markings and passable paths, and provides operation assistance prompts to the user.

[0089] Specifically, after step S107, global optimization and safety management of tower crane operations are achieved through intelligent path planning and dynamic collaborative technology. Motion constraint boundaries are constructed based on the tower crane's rated slewing radius and lifting height parameters. Combined with obstacle distribution characteristics in the dynamic perception area coverage map, an improved fast random extended tree algorithm is used for path search. This algorithm introduces a tower crane kinematic constraint model on top of the traditional RRT, balancing path optimality and computational efficiency through a dynamic step size adjustment strategy. In narrow passage scenarios, it automatically reduces the sampling area to improve obstacle avoidance accuracy. The path generation module decomposes passable paths into two categories: optimal paths and potential paths. Optimal paths are evaluated comprehensively using a cost function, considering path length, curvature, and safety margin. Potential paths retain suboptimal but feasible routes with emergency value. Both are labeled on the coverage map with different line types.

[0090] Key nodes on the path are identified using kinematic parameter extraction algorithms, including luffing inflection points, slewing limit angles, and sudden changes in hoisting speed. Each node stores its position coordinates and speed limit threshold. When the hook approaches a node, the speed change curve within the next 0.5 seconds is predicted based on the current motion state, and the threshold range is dynamically adjusted to match the actual working conditions. The dynamic perception area coverage map uses an incremental update mechanism, synchronously refreshing obstacle positions, risk levels, and path status every 200 milliseconds. A differential algorithm is used to update only the changed areas to reduce computational load. The dispatching interface integrates a multi-tower crane coverage map projection function, using 3D coordinate transformation to map the operating areas of each tower crane to a unified spatial coordinate system. Administrators can analyze global conflict hotspots by adjusting transparency and overlaying layers, such as identifying spatial interference areas when multiple tower cranes are slewing synchronously and automatically generating blocking suggestions.

[0091] The driver's interface utilizes augmented reality technology integrated with dynamic perception data. A viewpoint-following algorithm keeps the hook position centered in the field of view, and risk markers in key areas are dynamically scaled and color-coded to enhance display. The operation assistance prompt module generates semantic commands based on key path nodes and the current hook movement state. For example, it triggers a voice alarm and displays a recommended deceleration curve when approaching a speed limit threshold. Technically, this solution reduces path planning response time to less than 300 milliseconds, lowers the conflict rate in multi-tower crane collaborative operations, and reduces the driver's operational error rate. Through the deep integration of 3D situational awareness and intelligent path guidance, it improves operational safety and efficiency in complex construction scenarios, providing a reliable intelligent management solution for tower crane clusters in super high-rise building and bridge engineering projects.

[0092] For example, the driver triggers an audible and visual alarm, and simultaneously displays obstacle avoidance suggestions in a pop-up window on the interface. The dispatch terminal automatically pushes the coordinates of the high-risk area to the on-site safety terminal and sends a temporary restriction command to the tower crane control cabinet to suspend the movement of the hook in the dangerous direction until the risk level drops to medium or low risk.

[0093] In this way, proactive safety protection for tower crane operations is achieved through multi-level alarm linkage and radar perception enhancement technology. When the operator triggers the audible and visual alarm, the system adopts a graded warning strategy, conveying risk level information through combinations of different colored LED lights and buzzer frequencies. For example, a red flashing LED combined with a high-frequency buzzer sound prompts emergency avoidance, while a pop-up window on the interface dynamically marks the location of obstacles and suggests a turning angle. After receiving a high-risk signal, the dispatch terminal pushes the coordinate data to the on-site safety terminal through an industrial-grade communication protocol, triggering the audible and visual warning posts and electronic fence equipment, and sending a PWM modulation command to the tower crane control cabinet. The frequency converter dynamically adjusts the motor output power to reduce the hook movement speed until the risk parameters return to the safety threshold.

[0094] Optionally, the radar parameters for the layered perception perspective are set as follows: the scanning frequency of the first rotating radar is no less than 10Hz, the scanning frequency of the second rotating radar is no less than 15Hz, and the scanning frequency of the third rotating radar is no less than 20Hz with a scanning accuracy of millimeters. For example, each rotating radar uses an equally spaced microstrip antenna array and integrates a clutter suppression module to reduce interference from the tower crane's own structure on the detection signal. Specifically, the layered perception radar system adopts a multi-dimensional detection architecture to improve the accuracy of environmental perception. The layered rotating radars are deployed at different locations on the tower crane boom. The first-layer radar uses a 10Hz scanning frequency to cover a 360° horizontal field of view and uses digital beamforming technology to suppress clutter interference reflected by the tower crane's own metal structure. The second-layer radar focuses on the key operating area at a 15Hz frequency and uses FMCW modulation technology to achieve a ranging accuracy of 0.5 meters. The third-layer millimeter-wave radar analyzes the obstacle's motion vector through the Doppler effect and predicts its future trajectory using a Kalman filter algorithm. The core area radar system adopts 20Hz high-frequency scanning and ±90° elevation angle coverage. It achieves composite control of electronic and mechanical scanning through a microstrip antenna array. In high-temperature and dusty environments, the temperature compensation module uses a PT100 thermistor to correct the radar phase shift in real time, ensuring that the 0.1 mm ranging accuracy is not affected by environmental fluctuations.

[0095] Optionally, the radar parameters for the core area's precise perception perspective are set as follows: the first rotating radar has a scanning frequency of no less than 20Hz and a three-dimensional spatial coverage pitch angle range of ±90°; the second rotating radar has a scanning frequency of no less than 25Hz and a scanning accuracy of 0.1 mm. Furthermore, each rotating radar integrates a temperature compensation module to adapt to the high-temperature and dusty environments of construction scenarios. Specifically, the layered radar system improves obstacle recognition accuracy through multi-band collaborative detection, especially in scenarios involving multiple tower cranes operating simultaneously, where millimeter-wave radar can penetrate the tower crane boom to detect obstacles behind it. The temperature compensation module ensures that the radar's ranging stability error is less than 2 mm in environments ranging from -20℃ to 60℃, and the dynamic trajectory prediction algorithm provides a 2.5-second advance warning of potential collision risks. Through deep coupling of alarm linkage and motion control, the system can automatically limit the hook speed to below a safe threshold under hazardous conditions, reducing the accident rate of tower crane group collaborative operations and providing a reliable safety guarantee system for lifting operations in super high-rise buildings and ports.

[0096] In this embodiment, a dynamic perception area coverage map is dynamically constructed through the collaborative operation of dual radar modules and dynamic mapping technology, achieving full coverage of the hoisting operation process, greatly improving the tower crane's perception capability during hoisting operations, and further enhancing the level of hoisting intelligence and operational safety redundancy at the construction site.

[0097] After introducing the methods of exemplary embodiments of this application, the following references are made. Figure 3This application describes an exemplary embodiment of a dynamic mapping and sensing coverage system for the entire tower crane hoisting process. The system includes: a construction unit for mounting multiple rotating radars on the rotatable support of the tower crane to form dual radar modules, each communicating with a control module; a switching unit for acquiring the current hoisting operation stage of the tower crane in real time through the control module, including hoisting, slewing, luffing, and descent stages; switching the shooting angle of the dual radar modules according to the hoisting operation stage, using a focused sensing angle during the hoisting and descent stages, and a layered sensing angle during the slewing and luffing stages, synchronously controlling the dual radar modules to acquire continuous point cloud data for the corresponding operation stage; and a trajectory tracking unit for collecting tower crane operating parameters and, combined with the continuous point cloud data acquired by the dual radar modules, calculating the three-dimensional shape of the hook in real time. The system generates a continuous hook trajectory based on coordinates, movement speed, and swing angle. A point cloud modeling unit uses a target perception algorithm to identify dynamic obstacles and static backgrounds within the target monitoring space from the continuous point cloud data, constructing a construction environment point cloud model. A spatiotemporal registration unit, using the tower crane base as the origin of the physical coordinate system, performs spatiotemporal registration between the construction environment point cloud model and the physical coordinate system, establishing a global dynamic map containing the target monitoring space using a Gaussian surface mapping spatiotemporal graph convolutional network. A coverage map generation unit fuses the continuous hook trajectory with the global dynamic map, constructing a dynamic perception area coverage map using a trajectory differentiation algorithm and a dynamic ontology evolution algorithm. This dynamic perception area coverage map displays the continuously changing target monitoring space and construction environment features within the target monitoring space during the hoisting operation, changing with each stage of the operation. The system can implement all the steps described in the above method implementation, and the specific implementation methods of each step will not be repeated here.

[0098] After introducing the methods and systems of the exemplary embodiments of this application, a terminal device of the exemplary embodiments of this application will be described next. The terminal device can implement the steps described in the above method embodiments, and the specific implementation of each step will not be repeated here.

[0099] After introducing the methods, systems, and terminal devices of exemplary embodiments of this application, the following references will be made. Figure 4 The computer-readable storage medium of exemplary embodiments of this application will be described, please refer to... Figure 4 The computer-readable storage medium shown is an optical disc 30, on which a computer program (i.e., a program product) is stored. When the computer program is run by a processor, it implements the steps described in the above method embodiments. The specific implementation methods of each step will not be repeated here.

[0100] It should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Therefore, the protection scope of this application should be determined by the scope of the claims.

Claims

1. A method for dynamic mapping and sensing coverage of the entire tower crane hoisting process, characterized in that, The method includes: Two rotating radars are installed on the rotatable support of the tower crane to form a dual radar module, and the dual radar modules are respectively connected to the control module. The control module obtains the current hoisting operation stage of the tower crane in real time. The hoisting operation stage includes the hoisting stage, slewing stage, luffing stage and descent stage. The shooting angle of the dual radar module is switched according to the hoisting operation stage. The key perception angle is used in the hoisting and descent stages, and the layered perception angle is used in the slewing and luffing stages. The dual radar module is controlled to acquire continuous point cloud data of the corresponding operation stage in a synchronous manner. The tower crane's operating parameters are collected, and combined with the continuous point cloud data obtained by the dual radar modules, the three-dimensional coordinates, movement speed, and swing angle of the hook are calculated in real time to generate a continuous hook trajectory. A target perception algorithm is used to identify dynamic obstacles and static backgrounds contained in the target monitoring space from the continuous point cloud data, and a point cloud model of the construction environment is constructed. Using the tower crane base as the origin of the physical coordinate system, the point cloud model of the construction environment is spatiotemporally registered with the physical coordinate system, and a global dynamic map containing the target monitoring space is established using a spatiotemporal graph convolutional network with Gaussian surface mapping. By integrating the continuous hook trajectory with the global dynamic map, a dynamic sensing area coverage map is constructed using a trajectory differentiation algorithm and a dynamic ontology evolution algorithm. The dynamic sensing area coverage map is used to display the target monitoring space that changes continuously with each stage of the lifting operation, as well as the construction environment characteristics within the target monitoring space.

2. The method for dynamic mapping and sensing coverage of the entire tower crane hoisting process according to claim 1, characterized in that, In the key sensing perspective, the first rotating radar is used to scan the three-dimensional space when the hook is lifted vertically, lowered, and transferred horizontally; the second rotating radar is used to scan the micro-environment space of the hook's vertical path and landing area. In the layered perception perspective, the first rotating radar is used to scan the front end of the boom, the medium-to-long-range space on the side of the boom, and the medium-range space where the hook is located. The second rotating radar is used to scan the close-range space where the hoisting core area is located; The switching of the shooting angle of the dual radar modules according to the hoisting operation stage includes: When the hoisting operation is in the lifting or lowering phase, or when the hook lifting height is detected to be lower than the preset lifting height threshold and the horizontal movement range is less than the preset horizontal range threshold, switch to the key perception perspective. When the hoisting operation is in the slewing or luffing phase, or when the hook is detected to be performing a long-distance horizontal transfer or a large-scale slewing operation, the viewpoint is switched to the layered sensing perspective.

3. The method for dynamic mapping and sensing coverage of the entire tower crane hoisting process according to claim 1, characterized in that, The step involves employing a target perception algorithm to identify dynamic obstacles and static backgrounds within the target monitoring space from the continuous point cloud data, and constructing a point cloud model of the construction environment, including: According to the shooting angle and their respective corresponding acquisition angle, the continuous point cloud data acquired by the dual radar modules are fused and optimized to obtain a continuous point cloud map of the target monitoring space. Real-time weather information for the geographical area where the tower crane is located is obtained, and the point cloud weights are adjusted under different weather conditions through a point cloud density attenuation model. A hybrid target perception network is used to identify dynamic obstacles and static backgrounds contained in the target monitoring space based on the adjusted continuous point cloud map, and to extract the geometric features of the dynamic obstacles, the geometric features of the static background, and the real-time motion parameters of the dynamic obstacles from the adjusted continuous point cloud map. An improved generalized iterative nearest point (GICP) algorithm and a dynamically fused truncated symbolic distance function (TSDF) algorithm are used to construct the point cloud model of the construction environment using the extracted geometric features and real-time motion parameters.

4. The method for dynamic mapping and sensing coverage of the entire tower crane hoisting process according to claim 3, characterized in that, The method employs a hybrid target perception network to identify dynamic obstacles and static backgrounds within the target monitoring space based on an adjusted continuous point cloud map. It then extracts the geometric features of the dynamic obstacles, the geometric features of the static background, and the real-time motion parameters of the dynamic obstacles from the adjusted continuous point cloud map, including: The adjusted continuous point cloud map is processed by the semantic awareness feature enhancement module. An improved BEBLID descriptor and FasterNet backbone network are used to construct an adaptive feature pyramid by fusing multi-scale point cloud density information. The static structural boundary features are enhanced by the edge-aware loss function, and the enhanced point cloud feature map is output. Based on the enhanced point cloud feature map, the ISS algorithm and normal vector constraints are used to calculate the normal vector and curvature of the point cloud. The ISS feature score is generated through neighborhood covariance analysis and static feature points are screened. The DBSCAN clustering algorithm is combined to remove discrete noise points. Multi-view geometric constraints are introduced to remove dynamic interference points. The curvature threshold is dynamically adjusted according to the weather information. The geometric features of the static background are extracted and the static background geometric features are removed simultaneously to obtain dynamic candidate point clouds. Based on the dynamic candidate point cloud, a dynamic target detection network is constructed using a hybrid architecture of YOLOv5s and PointNet. The geometric features of the dynamic obstacle are extracted by outputting the category, bounding box and point cloud clustering mask of the dynamic obstacle through a joint two-dimensional and three-dimensional detection head. The motion optical flow is calculated based on the Farneback optical flow method and the initial velocity is obtained by combining it with point cloud data. The position, velocity and acceleration parameters of the dynamic obstacle are obtained by optimizing the Kalman filter algorithm, which serve as the real-time motion parameters of the dynamic obstacle.

5. The method for dynamic mapping and sensing coverage of the entire tower crane hoisting process according to claim 4, characterized in that, The process of calculating the motion optical flow based on the Farneback optical flow method and obtaining the initial velocity by combining it with point cloud data, then optimizing it using the Kalman filter algorithm to obtain the position, velocity, and acceleration parameters of the dynamic obstacle as the real-time motion parameters of the dynamic obstacle, further includes: Using real-time motion parameters and geometric features of dynamic obstacles, combined with Doppler feature encoding, the radial velocity of the dynamic obstacles is extracted, and a spatiotemporal correlation graph containing timestamps, coordinates, and velocity vectors is constructed, with the edge weights being the motion consistency probability. The detection sensitivity is dynamically adjusted based on the obstacle's speed, and a joint feature optimization process is triggered when a dynamic obstacle approaches the extracted static background.

6. The method for dynamic mapping and sensing coverage of the entire tower crane hoisting process according to claim 1, characterized in that, The construction environment point cloud model is spatiotemporally registered with the physical coordinate system, using the tower crane base as the origin. A spatiotemporal graph convolutional network with Gaussian element mapping is then used to establish a global dynamic map containing the target monitoring space, including: A three-dimensional physical coordinate system is established with the geometric center of the tower crane base as the origin. The installation position coordinates and attitude parameters of the dual radar modules are obtained. The original coordinates of each point in the point cloud model of the construction environment are mapped to the physical coordinate system through a coordinate transformation matrix. Based on the timestamps of the point cloud data collected by each rotating radar, a time interpolation algorithm is used to synchronize the point cloud data at different times, eliminate the spatiotemporal deviation caused by radar data acquisition delay, and align the construction environment point cloud model with the physical coordinate system in time and space. The aligned point cloud model is discretized into Gaussian grids using the Gaussian grid mapping method. Each Gaussian grid stores the mean static background geometric features of the corresponding spatial region, the probability distribution of dynamic obstacle motion parameters, and the confidence level of point cloud density. A spatiotemporal graph convolutional network is constructed, using Gaussian surface meshes as graph nodes. The stored geometric feature mean, motion parameter probability distribution, and point cloud density confidence are uniformly represented by a Gaussian distribution model and used as node features for each graph node. By using a spatiotemporal graph convolutional network to fuse the features of Gaussian surface nodes in a continuous time sequence, the motion parameter distribution of the surface elements containing dynamic obstacles is updated in real time, while the geometric features of static background surface elements are preserved, generating a global dynamic map containing time-dimensional evolution information. The global dynamic map dynamically updates the environmental features of the target monitoring space as the hoisting operation progresses.

7. The method for dynamic mapping and sensing coverage of the entire tower crane hoisting process according to claim 1, characterized in that, The process of fusing the continuous hook trajectory with the global dynamic map, and constructing a dynamic sensing region coverage map using a trajectory differentiation algorithm and a dynamic ontology evolution algorithm, includes: The continuous hook trajectory is spatiotemporally aligned with the global dynamic map to ensure that the hook trajectory coordinates are consistent with the physical coordinate system of the global dynamic map, and the hook movement sequence is synchronously associated with the real-time movement parameters of dynamic obstacles. Based on the trajectory differential algorithm, the continuous hook trajectory is differentially calculated, and the velocity change rate, acceleration change rate and swing angle change trend of the hook in different operation stages are extracted to generate hook motion characteristic parameters. An ontology model is constructed using a dynamic ontology evolution algorithm, which includes the lifting operation stage, hook motion characteristic parameters, static background geometric features, and real-time motion parameters of dynamic obstacles. The association weights of each element in the ontology model are dynamically updated according to the operation stage. Centered on the current position of the hook, and combining the hook motion characteristic parameters and environmental features in the global dynamic map, a dynamic perception boundary is generated through the ontology model. The dynamic perception boundary is adjusted in real time with the hook motion, covering the target monitoring space within the preset safety range of the hook. The dynamic sensing boundary is overlaid with the static background and dynamic obstacles in the global dynamic map to generate the dynamic sensing area coverage map. The dynamic sensing area coverage map displays in real time the spatial positional relationship between the hook trajectory and environmental features, early warning information of dynamic obstacles entering the dynamic sensing boundary, and changes in the sensing area at different operation stages.

8. The method according to claim 1, characterized in that, After fusing the continuous hook trajectory with the global dynamic map, and constructing a dynamically perceived region coverage map using a trajectory differentiation algorithm and a dynamic ontology evolution algorithm, the method further includes: Call the real-time motion parameters of the hook and the real-time motion parameters of the dynamic obstacles to calculate the Euclidean distance and relative motion trend between the hook and each obstacle; A preset collision probability model is used to quantify the risk value, and the safety risk level is marked on the coverage map of the dynamic sensing area. High-risk areas are marked with dense markings, medium-risk areas with sparse markings, and low-risk areas with safe markings. The collision probability model is constructed based on a distance threshold and relative velocity coupling algorithm. Obstacle classification features are extracted from the point cloud model of the construction environment. The obstacle classification features include the labeled geometric dimensions of static obstacles, fixed coordinates, the real-time position of dynamic obstacles, and motion trajectory vectors. Differentiated contour symbols are superimposed and labeled in the dynamic perception area coverage map. Static obstacles are labeled with solid graphics, and dynamic obstacles are labeled with dynamic graphics. The position coordinates of dynamic obstacles are updated synchronously, and the motion trajectory curve of dynamic obstacles is generated by connecting the position coordinate points of multiple consecutive frames.

9. The method according to claim 1, characterized in that, After fusing the continuous hook trajectory with the global dynamic map, and constructing a dynamically perceived region coverage map using a trajectory differentiation algorithm and a dynamic ontology evolution algorithm, the method further includes: The tower crane's rated slewing radius, lifting height range, and obstacle distribution in the dynamic sensing area coverage map are obtained. The fast random extended tree algorithm is used to calculate the passable path of the hook from the current position to the target landing point. In the dynamic sensing area coverage map, the optimal passable path is marked with a solid line, the potential passable path is marked with a dashed line, and the inaccessible area blocked by obstacles or beyond the tower crane's performance is marked with another type of solid line. Mark the location coordinates of key nodes in each passable path, and mark the speed limit threshold of key nodes based on hook motion parameters; The dynamic sensing area coverage map is updated synchronously at a preset frequency, and the safety risk level, obstacle distribution, and each passable path are automatically refreshed. The dynamic sensing area coverage map is transmitted to the dispatching terminal and the driving terminal. The dispatching terminal interface supports the overlay of dynamic sensing area coverage maps of multiple tower cranes. The overlaid dynamic sensing area coverage map is projected onto the three-dimensional display space. The driver interface displays the safety risk level of key areas around the hook to the user based on the dynamic perception area coverage map, and enlarges the risk markings and passable paths, and provides operation assistance prompts to the user.

10. A dynamic mapping and sensing coverage system for monitoring the entire process of tower crane hoisting, characterized in that, include: The building unit is used to install multiple rotating radars on the rotatable support of the tower crane to form a dual radar module. The dual radar modules are respectively connected to the control module for communication. The switching unit is used to obtain the current hoisting operation stage of the tower crane in real time through the control module. The hoisting operation stage includes the hoisting stage, slewing stage, luffing stage and descent stage. The unit switches the shooting angle of the dual radar modules according to the hoisting operation stage. The key perception angle is used in the hoisting and descent stages, and the layered perception angle is used in the slewing and luffing stages. The unit synchronously controls the dual radar modules to obtain continuous point cloud data of the corresponding operation stage. The trajectory tracking unit is used to collect tower crane operating parameters and, combined with the continuous point cloud data obtained by the dual radar modules, calculate the three-dimensional coordinates, movement speed and swing angle of the hook in real time to generate a continuous hook trajectory. The point cloud modeling unit is used to identify dynamic obstacles and static backgrounds contained in the target monitoring space from the continuous point cloud data using a target perception algorithm, and to construct a point cloud model of the construction environment. The spatiotemporal registration unit is used to perform spatiotemporal registration of the construction environment point cloud model with the physical coordinate system, with the tower crane base as the origin of the physical coordinate system, and to establish a global dynamic map containing the target monitoring space using a spatiotemporal graph convolutional network with Gaussian surface mapping. The coverage map generation unit is used to integrate the continuous hook trajectory with the global dynamic map, and construct a dynamic sensing area coverage map through trajectory differentiation algorithm and dynamic ontology evolution algorithm. The dynamic sensing area coverage map is used to display the target monitoring space that changes continuously with each operation stage during the hoisting operation, as well as the construction environment characteristics within the target monitoring space.

Images