Multi-device cooperative path planning methods and systems considering geometric and kinematic constraints

Abstract

This invention relates to the field of smart construction site technology, specifically to a method and system for multi-device collaborative path planning that considers geometric and kinematic constraints. The method includes the following steps: performing dynamic oriented bounding volume tree processing on multi-source point cloud data to obtain spatiotemporal bounding volume features of equipment or obstacles; obtaining traversable manifold features by differentially flattening the motion data of construction machinery; fusing spatiotemporal bounding volume features, traversable manifold features, and a perceptual uncertainty feature map to obtain comprehensive safety potential field features; based on spatiotemporal collaborative constraints, using homotopy-class sampling optimization to process the semantic map of the construction site and traversable manifold features to obtain initial collision-free path cluster features; obtaining preferred collaborative path features by topological homotopy culling of the initial collision-free path cluster features and dynamic right-of-way features; and combining the dynamic right-of-way features, comprehensive safety potential field features, and preferred collaborative path features, using spatiotemporal corridor quadratic planning to complete multi-device collaborative path planning.

Description

Technical Field

[0001] This invention relates to the field of smart construction site technology, specifically to a multi-device collaborative path planning method and system that considers geometric and kinematic constraints. Background Technology

[0002] In the process of multi-machine collaborative operation in smart construction sites, there are common prominent technical pain points such as unreliable environmental perception, conflicts of right-of-way among multiple devices, complex motion constraints, and insufficient safety redundancy. On the one hand, construction site scenarios are complex and ever-changing, with frequent occurrences of harsh environments such as dust, dense fog, and equipment obstruction. Single-sensor perception is prone to problems such as data sparsity and feature distortion, resulting in insufficient accuracy in environmental and equipment status perception and an inability to accurately capture the location of obstacles and the movement trajectory of equipment. On the other hand, heterogeneous equipment such as excavators, concrete pouring trucks, and dump trucks operate simultaneously on construction sites, and the lack of a dynamic right-of-way allocation mechanism easily leads to conflicts such as multiple devices vying for the same passage and low-priority devices interfering with the operation of high-priority devices, which seriously affect construction efficiency. Meanwhile, construction machinery is mostly multi-jointed and heterogeneous chassis structure with complex motion constraints. Traditional path planning methods rely on static maps and fixed weight design, which cannot adapt to dynamic operation scenarios on construction sites. They are difficult to match the kinematic characteristics of the equipment itself, and cannot take into account the temporal constraints of construction procedures. As a result, defects such as uneven trajectory, excessive energy consumption of equipment, low efficiency of multi-machine collaboration, and prominent collision risks are likely to occur. Ultimately, traditional methods cannot meet the safety and real-time control requirements of unmanned and intelligent construction on smart construction sites, thus restricting the implementation and application of unmanned operations on smart construction sites. Summary of the Invention

[0003] To address the shortcomings of existing methods and the needs of practical applications, and in order to solve the aforementioned problems, this invention provides a multi-device cooperative path planning method considering geometric and kinematic constraints, comprising the following steps: Dynamic oriented bounding volume tree processing is performed on multi-source point cloud data to obtain spatiotemporal bounding volume features of equipment or obstacles. By using differential flat mapping to map the motion data of construction machinery, traversable manifold features are obtained. The spatiotemporal bounding volume features, traversable manifold features, and a perceptual uncertainty feature map are fused to obtain comprehensive safety potential field features. Based on spatiotemporal cooperative constraints, homotopy-class sampling optimization is used to process the semantic map of the construction site and the traversable manifold features to obtain initial collision-free path cluster features. Topological homotopy culling is used to process the initial collision-free path cluster features and dynamic right-of-way features to obtain preferred cooperative path features. Combining the dynamic right-of-way features, comprehensive safety potential field features, and preferred cooperative path features, spatiotemporal corridor quadratic planning is used to complete multi-equipment cooperative path planning.

[0004] Optionally, the step of performing dynamic directed bounding volume tree processing on multi-source point cloud data to obtain the spatiotemporal bounding volume features of devices or obstacles includes the following steps: Register multi-source point cloud data and construct directed bounding volumes according to the device structure by region; supplement spatiotemporal attributes by associating temporal point clouds to form spatiotemporal bounding volumes that can characterize the dynamic features of devices or obstacles; construct a hierarchical bounding volume tree and obtain bounding volume features with spatiotemporal attributes through a dynamic update mechanism.

[0005] Optionally, obtaining the traversable manifold features by mapping the motion data of the engineering machinery through differential flat mapping includes the following steps: Based on the specific characteristics of engineering machinery, a categorized and constrained flattened kinematic model is constructed; based on real-time motion data of the equipment, the traversable manifold in the spatiotemporal dimension is represented by the flattened kinematic model; the dimensionality reduction of the traversable manifold is performed to obtain traversable manifold features.

[0006] Optionally, the process of fusing the spatiotemporal bounding volume features, the traversable manifold features, and the perceived uncertainty feature map to obtain the comprehensive security potential field features includes the following steps: A multi-head attention layer is constructed, and attention weights are calculated for different feature dimensions. The spatiotemporal bounding volume feature, the traversable manifold feature, the perceptual uncertainty feature map, and the corresponding attention weights are combined to obtain the comprehensive security potential field feature.

[0007] Optionally, constructing the perceived uncertainty feature map includes the following steps: A fusion feature map with both geometric accuracy and semantic integrity is generated using a visual radar fusion network; the perceptual uncertainty of the fusion feature map is quantified based on the Monte Carlo dropout strategy; and a perceptual uncertainty feature map is generated using the perceptual uncertainty.

[0008] Optionally, the step of using homotopy-based sampling optimization to process the semantic map of the construction site and the accessible manifold features based on spatiotemporal cooperative constraints to obtain initial collision-free path cluster features includes the following steps: Based on spatiotemporal collaborative constraints, the semantic map of the construction site and the traversable manifold are integrated to obtain a feasible region in the spatiotemporal dimension. Through homotopy class partitioning and semantic priority sampling, the feasible region is divided into multiple topologically equivalent path families. The path families are filtered using kinematic feasibility verification and path optimization to obtain the initial collision-free path cluster features.

[0009] Optionally, the step of obtaining preferred cooperative path features by topological homotopy removal of the initial collision-free path cluster features and dynamic right-of-way features includes the following steps: Based on dynamic right-of-way characteristics, a subset of conflict-free paths is identified from the initial collision-free path cluster; the safety potential field value of the path subset is evaluated according to the comprehensive safety potential field characteristics to obtain low-risk paths; the low-risk paths are ranked using collaborative benefits to generate preferred collaborative path characteristics.

[0010] Optionally, analyzing the dynamic right-of-way characteristics includes the following steps: Each device is considered a player in the game, and a payoff function is constructed. The potential field function is used to characterize the repulsive and attractive forces between devices. The dynamic right-of-way characteristics are solved by combining the payoff function and the potential field function.

[0011] Optionally, the step of combining the dynamic right-of-way characteristics, the comprehensive security potential field characteristics, and the preferred cooperative path characteristics to complete multi-device cooperative path planning using spatiotemporal corridor secondary planning includes the following steps: With the optimization of collaborative paths as the core, and combining equipment geometric characteristics, perception uncertainty and dynamic right-of-way constraints, a spatiotemporal corridor is constructed; based on the spatiotemporal corridor, the collaborative paths of multiple devices are solved using a multi-objective quadratic programming objective function and constraints.

[0012] This invention addresses collaborative operation scenarios involving multiple construction machines in smart construction sites. It constructs an integrated intelligent path planning system based on multi-source perception, motion modeling, spatiotemporal constraints, collaborative planning, and safety optimization. Environmental and uncertainty modeling is achieved through laser / visual fusion perception and Monte Carlo dropout; the kinematic feasible region of equipment is represented based on differential flatness and manifold transformation; spatiotemporal collaborative constraints are formed by combining BIM semantics and construction progress; safe, collision-free path clusters are generated using homotopy sampling and topology culling; dynamic right-of-way allocation is achieved using game theory and artificial potential fields; a comprehensive safety potential field is constructed by fusing multiple features through an attention mechanism; and finally, trajectory smoothing and real-time optimization are achieved within the spatiotemporal corridor using quadratic planning.

[0013] This invention can effectively solve problems such as dust obstruction at construction sites, heterogeneous equipment, multi-machine conflicts, and complex process constraints, and achieves collision-free, low-energy-consumption, high-smoothness, and highly collaborative autonomous operation, significantly improving construction safety, equipment utilization and overall operation efficiency, and providing reliable technical support for the unmanned, intelligent and refined management of smart construction sites.

[0014] Secondly, to efficiently execute the multi-device cooperative path planning method considering geometric and kinematic constraints provided by this invention, this invention also provides a multi-device cooperative path planning system considering geometric and kinematic constraints, comprising: an input device, an output device, a processor, and a memory, wherein the input device, output device, processor, and memory are interconnected, and the memory stores program instructions for using the multi-device cooperative path planning method considering geometric and kinematic constraints. The multi-device cooperative path planning system considering geometric and kinematic constraints provided by this invention has a compact structure and stable performance, and can stably execute the multi-device cooperative path planning method considering geometric and kinematic constraints provided by this invention, further improving the overall applicability and practical application capability of this invention. Attached Figure Description

[0015] Figure 1 A flowchart of a multi-device collaborative path planning method considering geometric and kinematic constraints is provided for an embodiment of the present invention; Figure 2 This is a framework diagram of a multi-device collaborative path planning system that considers geometric and kinematic constraints, provided as an embodiment of the present invention. Detailed Implementation

[0016] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0017] Throughout this specification, references to an embodiment, example, or illustration mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, phrases appearing in various places throughout the specification, such as "in one embodiment," "in an embodiment," "an example," or "an illustration," do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in any suitable combination and / or sub-combination in one or more embodiments or examples. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0018] Please see Figure 1 To address the aforementioned problems, this invention provides a multi-device cooperative path planning method that considers geometric and kinematic constraints, such as... Figure 1 As shown, in one embodiment, the method includes the following steps: S1. Perform dynamic oriented bounding volume tree processing on multi-source point cloud data to obtain the spatiotemporal bounding volume features of equipment or obstacles. Obtain the passable manifold features by differentially flat mapping the motion data of engineering machinery.

[0019] The process of performing dynamic directed bounding volume tree processing on multi-source point cloud data to obtain the spatiotemporal bounding volume features of devices or obstacles includes the following steps: S111. Register multi-source point cloud data and construct directed bounding volumes according to the device structure and region.

[0020] Multi-source radar point clouds are easily affected by complex environmental interference in smart construction site scenarios (such as rough and uneven ground, dust obstruction, and temporary obstacle obstruction). Direct registration will result in low fusion accuracy and failure to accurately extract the geometric features of equipment and obstacles. Therefore, it is necessary to first modify the registration algorithm based on the characteristics of the construction site scenario. To address this, a ground point prior masking technique is introduced. Combined with statistical patterns of construction site scenarios, the ground point cloud is automatically identified and removed through a dual judgment of point cloud elevation threshold and ground normal vector. Only target point clouds such as equipment and temporary obstacles (e.g., steel bar piles and formwork) are retained. Then, ICP registration is performed on the target point cloud.

[0021] To address the mobile characteristics of construction site equipment, the iteration termination condition of ICP registration was optimized (the iteration error threshold was adjusted from the usual 0.1m to 0.05m), improving the fusion accuracy of multi-source point clouds (LiDAR + millimeter-wave radar) and ensuring that the positional deviation of the target point cloud after registration is controlled within the millimeter level, thus meeting the accuracy requirements for geometric feature extraction of construction machinery.

[0022] Furthermore, in the construction site environment, dust can cause millimeter-wave radar to generate a large number of false noise points. These noise points have no actual movement trajectory and exhibit abnormal speeds, significantly differing from the movement characteristics of real equipment and obstacles. This is further compounded by the actual operating speed of construction site equipment (the movement speed of various types of construction machinery). (Set the obstacle speed to 0), and set the speed threshold range ( (This process removes noisy clouds whose speed exceeds the specified range.)

[0023] Meanwhile, spatial clustering analysis was performed on the remaining point cloud, and point clouds with a spatial distance of less than 0.3m were divided into the same cluster. Clusters with fewer than 5 small noise points were removed to further improve the purity of the point cloud and avoid false noise from interfering with the subsequent construction of the bounding volume.

[0024] For a single-frame fused point cloud, a directed bounding volume (OBV) is constructed for each region based on the geometric features of the engineering machinery: First, semantic segmentation is performed on the fused point cloud after denoising and registration to distinguish between the whole machine, key components of the equipment (such as the cab, boom, stick, and bucket of an excavator, and the cab and bucket of a loader) and obstacles (temporarily piled steel bars, formwork, and slag heaps).

[0025] Then, for different geometric characteristics of the target, directed bounding volumes are constructed respectively: for key components of the equipment, the bounding volume direction strictly follows the joint movement direction (e.g., the boom OBV is constructed along its rotation axis, and the bucket OBV is constructed along the opening and closing direction), ensuring that the bounding volume can accurately wrap the movement trajectory of the component; for the whole equipment, an overall OBV is constructed to represent the overall space occupation range of the equipment; for irregular obstacles (e.g., slag heaps, steel bar heaps), the minimum bounding volume algorithm is used to construct an OBV that fits its shape, reducing the redundancy of the bounding volume and improving the efficiency of subsequent collision detection.

[0026] S112. Associate the temporal point cloud to supplement the spatiotemporal attributes and form a spatiotemporal bounding volume that can characterize the dynamic features of the equipment or obstacle.

[0027] By associating temporal point clouds and adding timestamps and motion velocity / acceleration attributes to each OBV, a spatiotemporal bounding volume is formed. By combining device positioning time-series data (GNSS / IMU) with point cloud timestamps, the OBVs of multiple consecutive frames are temporally correlated. The positional change of each OBV between adjacent frames is calculated, thereby obtaining its velocity and acceleration. A corresponding timestamp label is added to each OBV to clarify its spatial position and motion state at a specific moment, ultimately forming a spatiotemporal bounding volume. For example, at time t1, the spatial coordinates of the excavator bucket OBV are (X1, Y1, Z1), its velocity is 0.8 m / s, and its acceleration is... By using the spatiotemporal bounding volume, the movement trajectory of the bucket can be tracked in real time, and its spatial position at the next moment can be predicted, providing support for collision warning for subsequent path planning.

[0028] S113. Construct a hierarchical bounding volume tree and obtain bounding volume features with spatiotemporal attributes through a dynamic update mechanism.

[0029] The bounding volume tree is constructed in a hierarchical manner based on the equipment as a whole, key components, and obstacles: The bounding volume tree is constructed in a top-down hierarchical manner, and is divided into three levels: the root node is the OBV of the entire equipment, which is used to quickly determine the spatial positional relationship of the equipment as a whole with other equipment and obstacles; the child nodes are the OBV of key components of the equipment, which are used to accurately determine whether the movement trajectory of each component of the equipment (such as boom and bucket) collides with other targets; the leaf nodes are the OBV of various obstacles, including fixed obstacles (such as fences and building components) and temporary obstacles (such as steel bar piles and dump trucks).

[0030] This hierarchical structure enables layered screening for collision detection. First, targets with no collision potential are quickly eliminated through the root node, and then precise collision detection is performed through child nodes and leaf nodes, greatly improving the efficiency of collision detection.

[0031] Furthermore, the bounding volume tree is updated according to a preset period. In this embodiment, the real-time requirements of smart construction site path planning (control period) are taken into account. The update cycle of the bounding volume tree is set to 100ms. During the update process, a local update strategy is adopted instead of reconstructing the entire tree: by comparing the point cloud data of the new frame with that of the previous frame, it is determined whether the position and size of each OBV have changed. If the point cloud of the new frame shows a change in the position of the obstacle (such as the temporary pile of steel bars being moved or the dump truck leaving), only the leaf node OBV of the corresponding obstacle is updated. If the critical components of the equipment move (such as the excavator boom being raised), only the corresponding child node OBV is updated. If the entire equipment moves, only the root node OBV and the related child node OBV are updated.

[0032] In this embodiment, for each new input point cloud frame, the newly detected bounding volume is matched with the bounding volumes of existing leaf nodes in the tree using an association algorithm (such as nearest neighbor matching). For nodes that match successfully, only their bounding volume parameters (position, size) and timestamp are updated. For newly added bounding volumes that fail to match, they are inserted into the tree as new leaf nodes, and the bounding volumes of their parent nodes are updated from bottom to top (using a bounding volume merging algorithm, such as calculating the minimum directed bounding volume that can contain the bounding volumes of all child nodes). For leaf nodes that remain unmatched (indicating that the object has disappeared), they are marked as invalid and pruned from the tree after several cycles. This incremental update avoids the overhead of rebuilding the entire tree every frame.

[0033] The process of obtaining traversable manifold features by mapping the motion data of engineering machinery using differential flat mapping includes the following steps: S121. Based on the specific characteristics of engineering machinery, construct a categorized, constrained, flattened kinematic model.

[0034] To address the heterogeneity of construction machinery (excavators: tracked + multi-jointed; loaders: wheeled + single-jointed), a categorical differential flat model is constructed: Excavators use tracked chassis, which cannot turn on the spot, and are equipped with three core joints: boom, stick, and bucket, resulting in high freedom of movement and strong non-linearity. Loaders use wheeled chassis, which are flexible in steering, and are equipped with only one working joint: the bucket, resulting in a relatively simple movement mode.

[0035] Based on this heterogeneity, differential flat models adapted to both types of equipment are constructed. The equipment's center of mass position plus joint angles are used as the flat output quantities. The equipment's center of mass position is used to represent the overall spatial displacement of the equipment, and the joint angles are used to represent the motion state of the working parts. Subsequently, the real-time state quantities of the equipment, such as chassis speed, chassis steering angle, and joint rotational angular velocity, are transformed into first- and second-order differential functions of the flat output quantities through differential transformation. This achieves the flattening and simplification of the nonlinear kinematic model, reduces the complexity of subsequent motion space calculations, and ensures that the model can accurately reflect the motion characteristics of the equipment.

[0036] Furthermore, a site motion constraint correction model is introduced: combining the actual operating parameters of the construction machinery in the smart construction site, the physical motion constraints are embedded into the domain of the flat output to avoid the model generating invalid solutions that do not conform to the actual motion capabilities of the equipment.

[0037] Specifically, for tracked equipment (such as excavators), considering the limitations of the construction site (such as narrow working areas and winding passages), a minimum turning radius constraint is set (approximately 1.5m, based on the parameters of commonly used tracked equipment on construction sites), and the range of chassis turning angles is clearly defined to prevent the model from generating paths that exceed the equipment's turning capabilities. For multi-joint equipment (such as excavators), considering their operational requirements, a maximum joint rotation angle constraint is set (e.g., the maximum lifting angle of the excavator boom is approximately 75°, the maximum rotation angle of the stick is approximately 120°, and the maximum opening and closing angle of the bucket is approximately 90°), limiting the range of joint angle values. At the same time, considering the load-bearing capacity of the equipment's hydraulic system, a joint rotation angular velocity constraint is added (e.g., the maximum rotation angular velocity of the boom is approximately 30° / s) to ensure that the motion state output by the model conforms to the mechanical performance of the equipment and to avoid mechanical damage or operational safety hazards caused by the lack of motion constraints.

[0038] S122. Based on the real-time motion data of the device, the traversable manifold in the spatiotemporal dimension is characterized by the flattened kinematic model.

[0039] Substitute encoder / sensor time-series data into the flattening model to calculate the feasible motion space of the device at each time step: The system collects real-time time-series data from the chassis encoder (wheel speed, steering angle), joint encoder (rotation angle of each joint), and hydraulic sensor (joint driving force) of the equipment (the sampling frequency is set to 100Hz to keep consistent with the update cycle of the bounding volume tree to ensure data synchronization). These time-series data are then substituted frame by frame into the flattening model of the corresponding equipment.

[0040] By calculating the feasible motion parameters of the equipment at each time stamp through model calculation, a feasible motion space is constructed. This space is a high-dimensional set that includes all spatial coordinate ranges reachable by the equipment's center of mass, as well as the achievable angle combinations of each joint at the corresponding coordinates, while also taking into account the constraints of chassis steering and joint movement. For example, the feasible motion space of the excavator at time t2 includes not only the X, Y, and Z three-dimensional coordinate ranges that its center of mass can move under the current working conditions, but also all angle combinations that the boom, stick, and bucket can achieve at that position. This clarifies where the equipment can go and how it can move at that time, providing a single-time motion basis for subsequent manifold construction.

[0041] The temporal feasible motion space is stitched together into a passable manifold (a continuous surface in a high-dimensional space, representing all feasible motion states of the equipment in the spatiotemporal dimension): Since the operation of engineering machinery is a continuous dynamic process, the feasible motion space at a single moment cannot reflect the continuous motion capability of the equipment. Therefore, it is necessary to temporally stitch together all feasible motion spaces in a continuous time series (such as within 10 seconds).

[0042] During the splicing process, it is necessary to ensure that the feasible motion space of adjacent moments is smoothly connected, and to eliminate abnormal motion space caused by sensor data noise (such as the centroid coordinates calculated at a certain moment exceeding the construction site operation area). The final passable manifold is a high-dimensional continuous surface, whose dimensions are composed of three-dimensional spatial coordinates + one-dimensional time + multiple dimensions of joint angles. It includes feasible regions in the spatial dimension, motion sequences in the time dimension, and feasible combinations of joint angles, fully representing all feasible motion states of the equipment in the spatiotemporal dimension.

[0043] Taking a tracked excavator as an example, its flatness output variables are selected as the two-dimensional plane coordinates (x, y) of the center point of the rear axle of the equipment and the heading angle of the entire vehicle. Equipment state variables (left and right track linear speeds) , The differential relationship between the output variable and the flattened output variable is established using the following flattening transformation formula: ; Where L is the track spacing. For wheel loaders, the flatness output variable is selected as the coordinates of the front axle center point. and front wheel steering angle Vehicle kinematic constraints are represented by the following relationship: , , ,in The front axle center speed, For vehicle body heading, Let be the wheelbase. Based on the real-time motion data of the device, the state space is transformed into a flat output space using the corresponding flattening transformation formula mentioned above. By sampling the polynomial parameters (such as fifth-order spline coefficients) of the flat output variables in the time domain, the passable manifold in the spatiotemporal dimension can be represented in the flat output space.

[0044] S123. Perform dimensionality reduction on the walkable manifold to obtain walkable manifold features.

[0045] A high-dimensional traversable manifold contains multiple dimensions (e.g., the traversable manifold of an excavator contains seven dimensions: X, Y, Z spatial coordinates, time, boom angle, stick angle, and bucket angle). Principal component analysis (PCA) algorithm is used to reduce the dimensionality of the high-dimensional manifold, preserving its core features. Finally, the high-dimensional manifold is reduced to three spatial coordinates (X, Y, Z) and one time dimension (t), forming a four-dimensional manifold feature.

[0046] Furthermore, the dimensionality-reduced walkable manifold only considers the kinematic constraints of the equipment itself, without taking into account obstacles in the construction site scenario (such as temporary steel rebar piles, other operating equipment, and barriers). Therefore, it combines the spatiotemporal bounding volume features of equipment or obstacles for adaptation and screening. Specifically, the features of the dimensionality-reduced walkable manifold are projected onto the constructed spatiotemporal bounding volume tree, and each spatiotemporal node (spatial coordinates + time) on the manifold is checked point by point to determine whether the bounding volume of the equipment overlaps with the bounding volume of the obstacle. If there is an overlap, the node is marked as an infeasible region and removed; if there is no overlap, the node is retained as a feasible region.

[0047] Through this screening, the resulting traversable manifold features not only satisfy the kinematic constraints of the equipment itself, but also avoid all obstacles in the construction site scenario, ensuring that the subsequent path planned based on this manifold is both kinematically feasible and geometrically collision-free.

[0048] S2. By fusing the spatiotemporal bounding volume features, the traversable manifold features, and the perceived uncertainty feature map, a comprehensive security potential field feature is obtained.

[0049] In this embodiment, constructing the perceived uncertainty feature map includes the following steps: S211. Generate a fused feature map with both geometric accuracy and semantic integrity using a visual radar fusion network.

[0050] PointNet++ is used to extract radar point cloud features, and CNN is used to extract visual image features. These features are then fused into a 3D feature map through an attention mechanism.

[0051] Specifically, the PointNet++ network is first used to extract features from the fused point cloud of LiDAR / millimeter-wave radar. This network can effectively handle irregular point cloud data. Through multi-layer sampling and feature aggregation, it extracts local geometric features (such as device outlines and obstacle edges) and global spatial features (such as the relative positions of devices and obstacles) of the point cloud, and outputs a point cloud feature map with dimensions C×H×W (C is the number of feature channels, and H and W are the feature map dimensions). At the same time, a lightweight CNN network (such as MobileNet) is used to extract features from the RGB / infrared images acquired by the visual camera. The focus is on extracting semantic features (such as device type and obstacle category) and texture features (such as pixel distribution in dusty areas and surface texture of devices) in the image, and outputting an image feature map with the same size as the point cloud feature map.

[0052] Finally, a channel attention mechanism is introduced to perform weighted fusion of the two types of feature maps and automatically allocate feature weights. For example, in dust-occluded areas, the feature weight of radar point cloud is increased (accounting for 0.7) and the feature weight of image is decreased (accounting for 0.3). In unoccluded areas, the feature weights of the two types are balanced (each accounting for 0.5). Finally, the feature maps are fused to generate a three-dimensional feature map.

[0053] In this embodiment, a dust mask (pixel value features of dusty areas at construction sites) can be added to the fusion network to improve the feature extraction accuracy of occluded areas. During smart construction site operations, equipment such as excavators and dump trucks generate a large amount of dust. The image pixels in dusty areas exhibit blurred and uniform grayscale characteristics, which can easily lead to distortion of image features extracted by CNN, thereby affecting the accuracy of fusion features.

[0054] To address this, a dust masking module is added after the image feature extraction module of the fusion network. First, by statistically analyzing the pixel value distribution patterns of construction site dust images, the pixel threshold range (grayscale value 150-200, saturation below 30) for dusty areas is determined. Based on this threshold, dusty areas in the image are automatically identified, generating a binary dust mask (dusty areas are marked as 1, and non-dusty areas are marked as 0). Subsequently, this mask is used to correct the image feature map. For dusty areas marked as 1, their image feature signals are weakened, while the radar point cloud feature signals of the corresponding areas are strengthened to compensate for the loss of image features caused by dust occlusion. For non-dusty areas, the normal fusion of image features and point cloud features is maintained, so that the fused feature map can still accurately capture the core features of equipment and obstacles in dusty and occluded scenarios.

[0055] S212. Based on the Monte Carlo dropout strategy, quantify the perceptual uncertainty of the fused feature map.

[0056] While fusion feature maps combine the advantages of vision and radar, factors such as dust, occlusion, and sensor noise in construction site scenarios can still lead to uncertainties in perception results (such as deviations in equipment location identification in dusty areas and false features caused by noise). If this uncertainty is not quantified, subsequent path planning will be based on unreliable perception results, increasing the risk of collisions.

[0057] Therefore, a dropout layer is introduced into the output layer of the fusion network, with a dropout rate of 0.2 (optimized for the noise characteristics of construction site data). The core function of the dropout layer is to simulate the uncertainty of the network by randomly discarding some network neurons, thereby quantifying the reliability of the perception results. Embedding the dropout layer into the output layer of the fusion network ensures that some fused feature nodes are randomly discarded during each forward propagation, simulating the perception results under different noise scenarios and laying the foundation for subsequent multiple rounds of sampling.

[0058] Perform 50 forward propagations (sampling) on ​​the same frame of fused data to obtain 50 feature maps, and calculate the feature variance for each spatial location (the larger the variance, the higher the perceptual uncertainty). To ensure the accuracy of uncertainty quantification, the same frame of visual-radar fusion data needs to be repeatedly sampled, and the perceptual uncertainty is quantified by statistically analyzing the dispersion of different sampling results.

[0059] Specifically, with fixed parameters for the fusion network, the same frame of fused data (including radar point clouds and visual images) is propagated 50 times forward. During each propagation, the dropout layer randomly discards 20% of the output feature nodes, resulting in 50 fused feature maps of consistent size with subtle differences in features. Subsequently, for each spatial location in the feature maps (corresponding to a spatial point in the construction site scene), the feature values ​​for that location are extracted from the 50 feature maps, and their variance is calculated. The larger the variance, the higher the dispersion of different sampling results, meaning the perception result at that spatial location is more affected by noise and occlusion, resulting in higher perception uncertainty. The smaller the variance, the more stable the sampling result, and the more reliable the perception result. For example, in the spatial location of a dust-occluded area, the feature values ​​of the 50 feature maps have a large dispersion and high variance, corresponding to high uncertainty; while in the unoccluded channel area, the feature values ​​have a small dispersion and low variance, corresponding to low uncertainty.

[0060] S213. Generate a perceptual uncertainty feature map through the perceptual uncertainty.

[0061] After Monte Carlo dropout sampling and variance calculation, the perceived uncertainty quantification value (variance) for each spatial location is obtained. However, the variance value is a continuous numerical value and cannot be directly used for subsequent safe potential field construction and path planning decisions. Therefore, it is necessary to transform the variance value into an intuitive and operable uncertainty level and associate it with spatiotemporal bounding volume features to generate a perceived uncertainty feature map.

[0062] Specifically, variance values ​​are mapped to uncertainty levels (0-5, where 0 = no uncertainty and 5 = completely unreliable). Considering the actual needs of construction site scenarios, the variance values ​​are graded and mapped to ensure that the level classification closely reflects the actual distribution of perceived uncertainty on the construction site, facilitating subsequent risk weight allocation. First, we statistically analyzed the variance distribution range of all spatial locations. Based on the characteristics of the construction site perception scenario, we set a grading threshold: variance < 0.05 is grade 0 (no uncertainty, such as unobstructed and noiseless passage areas, where the perception results are completely reliable). A variance of <0.1 indicates Level 1 (extremely low uncertainty, slight noise influence, and generally reliable perception results). Variance < 0.2 is Level 2 (low uncertainty, minor dust obstruction, reliable perception results). A variance of <0.3 indicates a level 3 (moderate uncertainty, moderate dust or partial obstruction; perception results should be used with caution). A variance < 0.4 indicates a level 4 (high uncertainty, severe dust or large-area obstruction, low reliability of perception results); variance Level 5 (completely unreliable, such as heavy dust or complete equipment obstruction, where the perception results cannot be used as a basis for path planning).

[0063] This hierarchical mapping transforms continuous variance values ​​into discrete uncertainty levels, intuitively reflecting the perceived reliability of each spatial location and providing a clear risk grading standard for the subsequent construction of a safety potential field.

[0064] Furthermore, the uncertainty level of the device / obstacle position is marked in the feature map to form a perceived uncertainty feature map. The core value of perceived uncertainty is to provide a basis for judging the collision risk between the device and the obstacle. Therefore, it is necessary to bind the uncertainty level to the specific target object (device, obstacle) to avoid the uncertainty feature from being decoupled from the target.

[0065] Specifically, spatiotemporal bounding volume features are extracted to obtain the spatial coordinate range and bounding volume outline of each device and obstacle. These bounding volume outlines are projected onto the generated uncertainty level map to clarify the uncertainty level of the spatial region where each device and obstacle is located, and are visually marked in the feature map—different colors are used to distinguish the uncertainty level (e.g., level 0 green, level 1-2 blue, level 3 yellow, level 4-5 red). At the same time, the ID of the device / obstacle and the corresponding uncertainty level are labeled (e.g., excavator 1, uncertainty level 2 rebar pile, uncertainty level 3). Finally, a perceptual uncertainty feature map is formed, which contains the spatial geometric information of the construction site scene, the perceptual reliability of each spatial location, and is also bound to the target information of the device and obstacle.

[0066] The process of fusing the spatiotemporal bounding volume features, the traversable manifold features, and the perceived uncertainty feature map to obtain the comprehensive safe potential field features includes the following steps: S221. Construct a multi-head attention layer and calculate the attention weights for different feature dimensions.

[0067] The spatiotemporal bounding volume features, traversable manifold features, and perceived uncertainty feature maps have significant differences in dimensionality and representation form (the spatiotemporal bounding volume focuses on geometry and motion state, the traversable manifold focuses on kinematic feasibility range, and the perceived uncertainty focuses on risk level). It is necessary to standardize and unify the coding of the three types of features to ensure the accuracy and rationality of subsequent attention weight calculation and feature fusion.

[0068] Specifically, the spatiotemporal bounding volume features are encoded as spatial location + geometric dimensions + motion velocity + collision risk value (collision risk value = bounding volume overlap probability): First, the core information of the spatiotemporal bounding volume is extracted. The spatial location adopts the unified UTM coordinate system of the BIM model, accurately marking the centroid coordinates (X, Y, Z) and boundary coordinates of each bounding volume to ensure that the spatial positioning is consistent with other features. The geometric dimensions are clearly defined as the length, width, and height of the bounding volume. The motion velocity directly adopts the motion attributes of the spatiotemporal bounding volume, including linear velocity and angular velocity, to clarify the dynamic motion state of the equipment. The collision risk value is quantitatively calculated through the bounding volume overlap probability, specifically the percentage of the overlap volume between the current equipment bounding volume and the bounding volumes of other equipment and obstacles. The overlap probability ranges from 0 to 1. When the overlap probability = 1, it indicates a complete collision, and the collision risk value = 1. When the overlap probability = 0, there is no collision risk, and the collision risk value = 0. By quantifying the collision risk, core indicators are provided for the subsequent risk assessment of the safety potential field.

[0069] The walkable manifold features are encoded as spatial coordinates + kinematic feasibility (feasibility 0-1, 1 = fully feasible): Based on the walkable manifold features, the dimensionality-reduced four-dimensional manifold (X, Y, Z, t) is first split by timestamps, with each timestamp corresponding to a set of spatial coordinates. During encoding, it is necessary to ensure that the spatial coordinates are consistent with the coordinate system of the spatiotemporal bounding volume and the perceptual uncertainty feature map. The kinematic feasibility is quantified according to the effectiveness of the manifold. Combining the equipment kinematic constraints and obstacle screening results, for nodes on the manifold that fully meet the equipment kinematic requirements and do not overlap with obstacles, the feasibility is set to 1 (fully feasible); for nodes that are close to the kinematic constraint boundary (such as joint angles close to the maximum limit) but do not exceed the constraint, the feasibility is set to 0.6-0.9; for nodes that slightly exceed the kinematic constraints or are close to obstacles, the feasibility is set to 0.1-0.5; for nodes that completely exceed the kinematic constraints or overlap with obstacles, the feasibility is set to 0 (completely infeasible).

[0070] The perceived uncertainty feature map is encoded as spatial coordinates + uncertainty level + risk weight (risk weight = 1 + uncertainty level × 0.2): The core information of the perceived uncertainty feature map is extracted. The spatial coordinates are consistent with the first two types of features, accurately corresponding to each spatial point in the construction site scenario. The uncertainty level directly adopts the 0-5 level classification results, clearly defining the perceived reliability of each spatial location. The risk weight is quantified based on the uncertainty level, with the calculation formula set as risk weight = 1 + uncertainty level × 0.2, ensuring that high uncertainty areas correspond to higher risk weights—level 0 (no uncertainty) risk weight = 1.0, level 1 = 1.2, level 2 = 1.4, level 3 = 1.6, level 4 = 1.8, level 5 = 2.0. Through risk weight assignment, perceived uncertainty is transformed into a quantifiable risk coefficient.

[0071] The importance of various features varies significantly in different scenarios (such as dust-covered areas and unobstructed passage areas). For example, in dust-covered areas, the importance of features related to perceiving uncertainty is far greater than that of other features; in unobstructed areas, the importance of passable manifold features is more prominent.

[0072] A multi-head attention layer is constructed, and attention weights are calculated for different feature dimensions to achieve dynamic weight allocation, ensuring that the fused features can accurately reflect the key security risks in the current scenario.

[0073] In this embodiment, three attention heads are set up to focus on collision risk, kinematic feasibility, and perception uncertainty, respectively. Each attention head independently calculates the importance weight of the corresponding feature, and then integrates them into the final weight through an attention fusion mechanism to adapt to the dynamic changes in the smart construction site scenario.

[0074] The system employs a three-tiered attention head system. The first attention head focuses on the collision risk dimension, specifically the collision risk value of the spatiotemporally bounding volume features, to detect potential collisions between devices and obstacles, and between devices themselves. The second attention head focuses on the kinematic feasibility dimension, emphasizing the feasibility index of the traversable manifold features to ensure the kinematic feasibility of the path. The third attention head focuses on the perception uncertainty dimension, specifically the risk weights of perception uncertainty features to mitigate safety hazards caused by unreliable perception. These three attention heads compute in parallel without interference, ensuring that the importance of each feature dimension is accurately captured and preventing any single feature from being overlooked.

[0075] Specifically, the three types of features are encoded into a unified feature vector sequence. Where N is the spatial location or the total number of elements in the feature map, and d is the feature dimension. Subsequently, a learnable linear projection matrix is ​​defined for each attention head h (h=1,2,3). For head h, the query, key, and value are calculated separately, and the attention weights for that head are calculated using scaled dot product attention. The outputs of the three heads are concatenated and then fused through a linear projection layer to obtain the final fused feature weight matrix. This weight matrix is ​​then multiplied element-wise with the original features F to obtain a weighted comprehensive feature representation.

[0076] Furthermore, dynamic weight calculation rules are formulated based on the risk characteristics of different scenarios in smart construction sites to ensure that the weight allocation aligns with actual safety needs, rather than remaining fixed. For example, if the perception uncertainty level of a certain area is 5 (completely unreliable), it indicates that the perception results in that area cannot accurately reflect the actual situation, and the perception risk needs to be focused on. Therefore, the uncertainty feature weight for this area is set to 0.6 (highest), the bounding volume feature weight is set to 0.3 (to assist in judging collision risk), and the manifold feature weight is set to 0.1 (lowest, because the kinematic feasibility reference value is reduced due to unreliable perception). If a certain area has no obstacles (such as the main road passage of the construction site), the collision risk is extremely low, the importance of the bounding volume feature decreases significantly, and the weight is set to 0.1. The passable manifold feature (reflecting the feasibility of equipment movement) becomes the core, with a weight set to 0.8, and the uncertainty feature weight is set to 0.1 (perception is reliable in unobstructed scenarios, but the reference value is low).

[0077] In addition, weight allocation rules for other typical scenarios can be added: for moderate dust areas (uncertainty level 3), the weight allocation is 0.4 for uncertainty feature, 0.4 for bounding volume feature, and 0.2 for manifold feature; for work areas (with collaborative equipment), the weight allocation is 0.4 for bounding volume feature, 0.3 for manifold feature, and 0.3 for uncertainty feature, to ensure that the weight allocation in different scenarios is scientific and reasonable and accurately matches the key risks of the scenario.

[0078] S222. By combining the spatiotemporal bounding volume features, the traversable manifold features, the perceptual uncertainty feature map, and the corresponding attention weights, a comprehensive security potential field feature is obtained.

[0079] The three types of features are weighted and integrated according to attention weight to calculate the safety potential field value of each spatial location: Potential field value = collision risk value × risk weight + (1 - kinematic feasibility) × 0.5. This formula is designed in accordance with the core requirements of smart construction sites that prioritize safety while taking into account feasibility, and also incorporates the dynamic adjustment logic of attention weight.

[0080] The collision risk value of the spatiotemporal bounding body is multiplied by the risk weight of perceived uncertainty. This considers both the magnitude of the collision risk and the reliability of the perception results. If the collision risk is high and the perception is unreliable (high risk weight), the calculation result of this part is significantly increased, highlighting the safety hazard. Secondly, (1 - kinematic feasibility) represents the degree of kinematic infeasibility. The lower the feasibility, the larger this value. Multiplying it by 0.5 (fixed weight, ensuring the core position of kinematic feasibility) reflects the influence of kinematic constraints on the safety potential field. Finally, the two results are added together to obtain the safety potential field value for each spatial location. The potential field value ranges from 0 to 2 (collision risk value 0-1, risk weight 1.0-2.0, product range 0-2; (1 - kinematic feasibility) 0-1, multiplied by 0.5 range 0-0.5, overall potential field value range 0-2.5). The higher the potential field value, the higher the safety risk of that spatial location, and the more it needs to be avoided.

[0081] Furthermore, the potential field values ​​are mapped to a three-dimensional safety potential field (high potential field value = high risk, must be avoided; low potential field value = low risk, priority passage), forming a comprehensive safety potential field characteristic: Based on the spatial coordinate system of the construction site BIM model, the safety potential field value of each spatial location is mapped to the three-dimensional space according to the coordinate correspondence to construct a three-dimensional safety potential field. Secondly, the potential field values ​​are subjected to hierarchical visualization processing to facilitate subsequent path selection: low-risk areas (potential field value 0-0.8), marked in green, correspond to areas with no collisions, kinematic feasibility, and reliable perception, and are prioritized as equipment passage paths; medium-risk areas (potential field value 0.8-1.5), marked in yellow, correspond to areas with slight collision risk, moderate kinematic feasibility, or low perception uncertainty, and require cautious passage; high-risk areas (potential field value above 1.5), marked in red, correspond to areas with high collision risk, kinematic infeasibility, or completely unreliable perception, and are prohibited from passage.

[0082] The resulting comprehensive safety potential field characteristics integrate the core information of three types of constraints: geometry, kinematics, and perception. Through visualization and hierarchical classification, they intuitively present the distribution of safety risks in the construction site scenario.

[0083] S3. Based on spatiotemporal collaborative constraints, homotopy-class sampling is used to optimize the semantic map of the construction site and the features of the passable manifold to obtain the initial collision-free path cluster features.

[0084] The method of obtaining initial collision-free path cluster features based on spatiotemporal collaborative constraints and using homotopy-class sampling optimization to process the semantic map of the construction site and the features of the accessible manifold includes the following steps: S31. Based on spatiotemporal collaborative constraints, the semantic map of the construction site is integrated with the traversable manifold to obtain a feasible region in the spatiotemporal dimension.

[0085] First, semantic annotation is performed on the BIM model, embedding construction process attributes and spatial constraint information into the model, transforming it from a static geometric model into a constrained semantic model.

[0086] Specifically, by combining the construction process flow on the site (such as earthwork excavation → formwork erection → rebar tying → concrete pouring → curing), the built-in attribute editing function of the BIM model is used to add exclusive process attribute labels to each work area and component in the model. The labels need to clearly specify three core information: time interval + process type + allowed equipment type, so as to achieve the binding of spatial area - process - time.

[0087] For example, a building foundation area is marked with the time period t3-t8: rebar tying process, only rebar trucks / cranes are allowed to enter, clearly indicating the type of work and the equipment that can pass through during this time period; for areas with overlapping work, the layered process attributes need to be marked, such as the time period t8-t12 in the same area being the concrete pouring process, only pouring trucks and concrete placing machines are allowed to enter, to avoid conflicts between equipment of different processes.

[0088] At the same time, the annotation information must be strictly synchronized with the subsequent schedule to ensure that the time interval of the process attributes is consistent with the schedule, so as to provide a unified semantic benchmark for subsequent timeline mapping.

[0089] Furthermore, smart construction sites have narrow operating areas and diverse equipment types (large excavators, small forklifts, heavy-duty concrete pouring vehicles, etc.). Spatial constraints directly determine whether equipment can pass through and operate safely. Therefore, key spatial constraint parameters related to equipment path planning are accurately extracted from the BIM model.

[0090] Specifically, the spatial measurement function of the BIM model is used to extract the boundary coordinates of each work area and the width of equipment access passages (based on the dimensions of commonly used equipment on the construction site, the minimum width of the site passage is determined to be approximately 3m to ensure the smooth passage of large excavators, dump trucks, and other equipment); for areas such as temporary access roads and work platforms, load-bearing limit parameters are extracted based on the construction design documents (such as the load-bearing capacity of temporary access roads). To prevent road collapses when heavy concrete pouring vehicles and cranes pass through; in addition, it is necessary to extract the spatial coordinates of obstacles (such as fences, building components, and temporary facilities) to provide spatial reference for subsequent constraint feature selection, and ensure that the marked spatial constraints fully cover the core needs of equipment passage and operation.

[0091] Furthermore, the schedule is transformed into a standardized three-dimensional mapping table, and the timeline is adjusted in combination with the characteristics of the construction site processes to ensure that the time constraints fit the actual construction logic and achieve a deep integration of process, time and space.

[0092] First, the core information in the Gantt chart is analyzed, including the start and end times of each process, the corresponding work area, and the work content. Management information irrelevant to equipment path planning (such as personnel scheduling and material arrival times) is removed. Then, a three-dimensional mapping table is constructed according to the principles of timeline segmentation, spatial area correspondence, and process type association. Each entry in the table contains four core elements: time interval, spatial area, process type, and permitted equipment type, ensuring clear information that can be directly used for path planning constraints. For example, in the t0-t2 time period: Area A, earthwork excavation process, excavators and dump trucks are allowed to enter, clearly defining the operational requirements and accessible equipment for Area A during this time period; in the t2-t5 time period: Area A, formwork construction process, forklifts and woodworking equipment are allowed to enter, achieving precise matching of equipment types after process switching. The mapping table must be sorted chronologically, with no overlapping or omissions in time intervals, ensuring clear access constraints for multiple devices in different time periods and avoiding time conflicts.

[0093] Introducing process constraints to modify the timeline: For processes requiring continuous operation (such as concrete pouring and asphalt paving), the corresponding area in the timeline is marked as uninterrupted and prohibited from entry by other unrelated equipment to ensure continuous operation and avoid construction quality problems caused by equipment interference; For exclusive processes (such as rebar tying and formwork removal), the time intervals of each process are clearly marked to ensure that the times of different exclusive processes in the same area do not overlap, and a buffer period (such as a 30-minute process switching interval) is set in the timeline for equipment withdrawal and site cleanup to avoid collisions between equipment during the switching of different processes; For emergency processes (such as emergency repairs and pouring before the initial setting of concrete), priority access is marked in the timeline to provide a time basis for subsequent dynamic right-of-way allocation.

[0094] Next, the dispersed constraint information is transformed into standardized spatiotemporal collaborative constraint features, which are then filtered in conjunction with equipment characteristics to ensure that the constraint features are targeted and feasible, providing top-level process constraints for subsequent path planning, including: The mapping table is transformed into spatiotemporal constraint features: Centering on time intervals, spatial regions, and permitted equipment types / quantities, the 3D mapping table undergoes feature extraction and standardization to generate spatiotemporal collaborative constraint features directly applicable to path planning. Each constraint feature entry includes a specific time range (accurate to the minute, adapting to the real-time requirements of construction site equipment operations), clear spatial region coordinates (based on the spatial coordinate system of the BIM model, accurately locating the region), permitted equipment types (determining equipment models and functions based on process requirements, such as limiting excavators to medium-sized excavators to prevent large equipment from entering narrow areas), and equipment quantity restrictions (reasonably setting the number of equipment based on the size of the area and operational efficiency, such as allowing excavators in area A during time period t0-t2). Taiwan, dump truck (To avoid overly dense equipment leading to traffic congestion and collision risks).

[0095] At the same time, a process priority label is added to each constraint feature (such as the priority of the pouring process being higher than that of the earthwork excavation process) to provide a process-level weight reference for subsequent dynamic right-of-way allocation, ensuring that the constraint features not only meet process requirements but also adapt to the actual scenario of multi-device collaboration.

[0096] Furthermore, by associating the spatiotemporal bounding volume features, invalid entries in the constraint features that indicate the equipment size exceeds the area's passage width are removed: the spatiotemporal co-constraint features only consider process, time, and space constraints, without considering the equipment's own geometric dimensions. This may result in invalid constraints that allow the equipment to enter, but the equipment size exceeds the area's passage width. Therefore, it is necessary to combine the equipment's spatiotemporal bounding volume features for adaptation and screening.

[0097] Specifically, the spatiotemporal bounding geometry of each piece of equipment (such as the overall width and height of an excavator, and the turning radius of a dump truck) is extracted and compared with the corresponding area access width and spatial height in the constraint features. If the equipment bounding dimensions exceed the area access width (e.g., a large excavator with a width of 3.5m exceeds the minimum area passage width of 3m), the constraint feature is marked as invalid and removed. If the equipment dimensions conform to the area spatial constraints, the constraint feature is retained to ensure its practical feasibility. This screening process ensures that the constraint features match the process, time, and space, while also taking into account the geometric characteristics of the equipment itself, avoiding path planning errors caused by invalid constraints.

[0098] The semantic map of the construction site includes spatial semantic labels such as work area, passage area, and obstacle area, which can only provide static spatial classification information; while the traversable manifold features focus on the kinematic feasible range of the equipment itself and are not related to site semantics and construction process constraints. Deep integration of the two can achieve the unification of spatial semantic constraints, equipment kinematic constraints and time process constraints.

[0099] Specifically, the traversable manifold features are projected onto the spatial coordinate system of the semantic map to mark the semantic regions in the map that are traversable by the device: First, the spatial coordinate system of the semantic map and the traversable manifolds is unified (using the UTM coordinate system commonly used in construction sites to ensure accurate spatial correspondence). The features of the traversable manifolds (four-dimensional spatiotemporal features) are projected point by point into the semantic map. Combined with the regional annotations of the semantic map, the validity of the traversable manifolds is classified and marked. As the main passage area for construction equipment, all traversable manifolds within the passage area are marked as valid, ensuring that equipment can move freely along the passage. The traversable manifolds in the work area must strictly match the requirements of the current process. For example, when a certain area is in the rebar binding process, only traversable manifolds that match the motion characteristics of rebar trucks and cranes are marked as valid, and manifold areas corresponding to equipment that is not in this process, such as pouring trucks and excavators, are excluded to avoid invalid path sampling. For example, the traversable manifolds in the passage area can adapt to the motion trajectories of all types of equipment, while the traversable manifolds in the work area only adapt to equipment allowed by the current process, achieving accurate binding between manifold features and site semantics.

[0100] Next, combining spatiotemporal collaborative constraint features, the accessibility of the time dimension is marked on the map: the accessibility manifold feature itself already contains time dimension information, but without combining the time constraint of construction process, there may be a situation where the manifold of a certain spatial area is valid, but the equipment is not allowed to enter at the current time. Therefore, it is necessary to associate the spatiotemporal collaborative constraint features to mark the time validity of the accessibility manifold in each area of ​​the map.

[0101] Specifically, the relationships between time intervals and process types in the spatiotemporal collaborative constraint features are extracted and mapped to corresponding regions in the semantic map, clarifying the effective time range of the traversable manifolds in each region. Manifold regions exceeding the time range are marked as invalid to avoid sampling paths that conflict with the process in terms of time. For example, at time t4, region B is in the formwork support process, allowing only forklifts and woodworking equipment to enter. In this case, the traversable manifolds corresponding to forklifts and woodworking equipment in this region need to be marked as valid at time t4, while the manifolds corresponding to excavators and dump trucks are invalid at time t4, ensuring that the temporal validity of path sampling is consistent with the process constraints.

[0102] S32. By using homotopy partitioning and semantic priority sampling, the feasible region is divided into multiple topologically equivalent path families.

[0103] Based on the start / end point of the operation, the semantic map is divided into multiple homotopy classes: First, the start point (such as the equipment parking area, material stacking area) and end point (such as the center of the operation area, unloading area) of each device are defined. Using the start point and end point as fixed benchmarks, combined with the distribution of obstacles and the boundaries of the area in the semantic map, the passable area is divided into multiple homotopy classes using a topology partitioning algorithm.

[0104] Each homotopy class corresponds to a set of topologically equivalent paths. This means that paths in this class do not cross obstacles, do not exceed semantic constraints, and have a consistent overall orientation (topological structure), differing only in detailed path details. For example, equipment traveling from the parking area (starting point) to the basic work area (ending point) can be divided into 2-3 homotopy classes based on the passageway orientation. Each homotopy class corresponds to a core orientation (such as via the east passageway or the west passageway), ensuring that paths within each homotopy class conform to the site's spatial semantic constraints and avoiding invalid paths that cross fences or obstacles.

[0105] Furthermore, a semantic-first sampling strategy is adopted, prioritizing the sampling of path nodes within the channel areas and permitted areas of the process marked in the semantic map to reduce invalid sampling. Specifically, during sampling, the effective semantic regions (channel areas and work areas) corresponding to the current time and current process are first screened out. Path nodes are selected within these areas according to the principle of uniform sampling, and then the nodes are connected to form the initial sampling path. For obstacle areas and areas not permitted by the current process, they are skipped directly without sampling, reducing the generation of invalid paths from the source. At the same time, combined with the traversable manifold features, it is ensured that all sampled path nodes fall within the kinematically feasible range of the equipment, avoiding the generation of kinematically infeasible paths, further improving sampling efficiency and path quality.

[0106] In another embodiment, in a two-dimensional spatiotemporal (XYT) or three-dimensional spatiotemporal (XYZT) raster map that integrates semantic and temporal constraints, a homotopy class partitioning algorithm based on "Persistent Homology" is used: The feasible region and obstacles (including the spatiotemporal trajectories of dynamic obstacles) are modeled as a cavity complex. Then, the homology groups of this complex at different spatial scales are calculated. By analyzing the "birth" and "death" times of homology classes (e.g., loops corresponding to one-dimensional homology classes), stable homotopy classes representing different "detour methods" are extracted. Each homotopy class is represented by a representative basepoint path. The path families are filtered using kinematic feasibility verification and path optimization. Semantic priority sampling refers to sampling path nodes with higher probability density on the grid of the "channel area" and the "allowed area" of the current process within the channel corresponding to each homotopy class (e.g., using the RRT* algorithm). Kinematic feasibility verification involves using the constructed "flattened kinematic model" to inversely solve the time series of the flattened output variables (e.g., solving for polynomial coefficients) for each sampled path, and checking whether the solution satisfies the equipment's maximum speed, acceleration, and joint angle constraints; paths that do not meet these constraints are discarded.

[0107] S33. Filter the path family using kinematic feasibility verification and path optimization to obtain the initial collision-free path cluster features.

[0108] In this embodiment, the kinematic feasibility of each sampling path is calculated and invalid paths are eliminated: based on the traversable manifold features, the kinematic feasibility of each sampling path is verified one by one, with the core verification being whether the path conforms to the kinematic constraints of the device.

[0109] Specifically, the spatial coordinates and time information of each node on the sampling path are extracted and substituted into the features of the passable manifold to determine whether the motion state of the equipment corresponding to the node (such as joint angle and chassis steering angle) is within the feasible range of the passable manifold. At the same time, the overall motion trajectory of the path is verified to meet the mechanical performance requirements of the equipment by combining the dynamic parameters of the equipment (such as the maximum rotation angle of the joint and the minimum steering radius of the chassis).

[0110] For example, if a sampling path requires the excavator boom lifting angle to exceed the maximum limit of 75°, or requires the tracked excavator to turn in place, it is determined to be a kinematically infeasible path and is eliminated; only sampling paths that are kinematically feasible and conform to the mechanical performance of the equipment are retained to ensure the feasibility of subsequent path planning.

[0111] After kinematic feasibility screening, multiple effective paths will still be retained in each homotopy class. Further optimization screening is needed to retain the optimal path in each homotopy class and reduce the number of paths. The optimization screening adopts a multi-objective evaluation strategy, combining the core needs of multi-equipment collaborative operation on the construction site. Three core evaluation indicators are selected: path length, energy consumption cost, and collaborative adaptability, to comprehensively score the effective paths in each homotopy class. Among them, path length is given priority (shortening operation time), followed by energy consumption cost (reducing construction site operation costs), and collaborative adaptability is secondary (facilitating cooperation with other equipment).

[0112] Based on the comprehensive score, the highest-scoring optimal path in each homotopy class is retained, such as the shortest path, the path with the lowest energy consumption, or the path with the best cooperative fit. The optimal paths of all homotopy classes are aggregated to form the initial collision-free path cluster features. Each cluster corresponds to the optimal path of a homotopy class, which ensures both the diversity of paths (covering different topological orientations) and the high quality of paths (collision-free and kinematically feasible), providing a reliable candidate path basis for subsequent topological homotopy elimination strategies.

[0113] S4. The initial collision-free path cluster features and dynamic right-of-way features are processed by topological homotopy elimination to obtain the preferred cooperative path features.

[0114] In one embodiment, analyzing the dynamic right-of-way characteristics includes the following steps: S411. Treat each device as a game participant and construct a payoff function.

[0115] Each piece of equipment is treated as an independent game subject, and the operation priority (e.g., concrete pouring truck > excavator > dump truck) is the initial payout weight.

[0116] Based on the core needs of smart construction site construction technology, the equipment priority classification standard is clarified - based on the importance of the process and the urgency of the operation, priority is given to ensuring the operational rights of key process equipment. The specific priority order is: pouring truck (core process equipment, affecting construction progress nodes) > crane (key material transfer equipment) > excavator (earthwork excavation foundation equipment) > dump truck (waste material transfer auxiliary equipment) > forklift (small material transfer equipment).

[0117] Then, each priority is assigned a corresponding initial revenue weight (1.0 for concrete pouring truck, 0.9 for crane, 0.8 for excavator, 0.7 for dump truck, and 0.6 for forklift). The weight is directly included in the revenue calculation to ensure that high-priority equipment has a better decision tendency in the game. At the same time, the game decision scope of each equipment is limited to the spatiotemporal area covered by its own operation path to avoid ineffective cross-regional games and improve decision efficiency.

[0118] Furthermore, a revenue function is constructed: Revenue = Collaboration Revenue (e.g., increased efficiency from two excavators working together) - Collision Costs (e.g., losses from equipment collisions) - Waiting Costs (e.g., time lost while waiting for low-priority equipment). Based on the actual operational scenarios of smart construction sites, each component of the revenue function is quantitatively defined to ensure the function is calculable and adaptable to different scenarios.

[0119] Synergy benefits are quantified based on the coordination relationship between equipment. For example, when two excavators work together to excavate, efficiency increases by 30%, and the corresponding synergy benefit = benefit per unit time of single equipment × 0.3 × working time. Collision costs are quantified based on equipment value, maintenance costs, and downtime losses. For example, if an excavator collides with a concrete pouring truck, the collision cost = equipment maintenance cost + loss due to process delay during downtime (calculated as cost per minute of process delay × delay duration). Waiting costs are quantified based on equipment priority and waiting time. When low-priority equipment waits for high-priority equipment, the waiting cost = equipment operating cost per unit time × waiting time × priority weight (the lower the priority, the higher the weight of the waiting cost; for example, when a dump truck waits for a concrete pouring truck, the weight of the waiting cost is 1.2).

[0120] S412. The potential field function is obtained by characterizing the repulsive and attractive forces between devices through the potential field function.

[0121] Based on the spatiotemporal enclosing characteristics of the equipment, the repulsive force between the equipment is calculated (the higher the overlap of the enclosing bodies, the greater the repulsive force). The core function of the repulsive potential field is to avoid collisions between equipment. Combining the dynamic operating characteristics of the equipment on the construction site, the repulsive force is calculated using a two-factor approach: enclosing body overlap and movement speed, satisfying: Repulsive force = k1 × enclosing body overlap × relative speed of equipment (k1 is the repulsion coefficient, set according to the size of the equipment on the construction site; for large equipment, k1 = 5.0, and for small equipment, k1 = 3.0).

[0122] The overlap of the enclosing bodies is calculated using spatiotemporal enclosing body characteristics, specifically the ratio of the overlapping volume of two enclosing bodies to the volume of the smaller enclosing body (overlap ranges from 0 to 1; when overlap = 1, there is complete overlap, and the repulsive force reaches its maximum). The relative velocity of the equipment is calculated using GNSS / IMU positioning data; the higher the relative velocity, the greater the repulsive force, thus preventing collisions with high-speed moving equipment. Simultaneously, for temporary obstacles on the construction site (such as piles of rebar and formwork), their enclosing bodies are treated as fixed equipment, and the repulsive force between the equipment and the obstacle is calculated to ensure that equipment collision avoidance covers all potential collision targets.

[0123] Based on the spatiotemporal coordination constraint characteristics, the attractive force between cooperating equipment is calculated (e.g., the attractive potential field between a concrete pouring truck and a concrete placing boom increases as the distance decreases). The core function of the attractive potential field is to guide the efficient cooperation of cooperating equipment. Considering the site's process coordination requirements, the attractive force is calculated only for cooperating equipment pairs marked in the spatiotemporal coordination constraint characteristics (e.g., concrete pouring truck and concrete placing boom, excavator and dump truck). The attractive force between non-cooperating equipment is set to 0 to avoid path deviations caused by ineffective attraction. The attractive force satisfies: Attractive Force = k2 × Cooperation Weight × (1 / Distance between Equipment) (k2 is the attraction coefficient, k2 = 4.0 for cooperating equipment pairs; the coordination weight is set according to the importance of process coordination, e.g., the coordination weight between a concrete pouring truck and a concrete placing boom = 0.9, and the coordination weight between an excavator and a dump truck = 0.7). The distance between equipment is calculated based on the centroid coordinates of the spatiotemporal bounding volume. The smaller the distance, the greater the attractive force, guiding cooperating equipment to approach quickly and improving operational efficiency.

[0124] Furthermore, an urgency-based adjustment potential field is introduced (e.g., the weight of the attraction potential field for the pouring truck is increased as the pouring process approaches its initial setting time). In smart construction sites, some processes have strict time constraints (e.g., the initial setting time of concrete pouring, emergency repairs), requiring the urgency-based adjustment potential field to prioritize the right-of-way for equipment performing urgent tasks. The urgency of a task is quantified based on the remaining time and importance of the task, divided into 5 levels (1-5, with level 5 being the highest urgency). For example, if the remaining time for a pouring process is less than 30 minutes, the urgency level is 5. The urgency adjustment coefficient is calculated as 1 + 0.2 × urgency level. This coefficient is multiplied by the attraction potential field and the payoff function weights to strengthen the potential field for emergency equipment. For instance, when the urgency level of a pouring truck is 5, the adjustment coefficient is 2.0, doubling both its attraction potential field and payoff weights, ensuring it obtains the highest priority in the game and preventing delays due to right-of-way conflicts.

[0125] S413. Combining the revenue function and the potential field function, solve for the dynamic passage right characteristics.

[0126] In this embodiment, the Nash equilibrium game model is used to obtain the passage priority of each device in each spatiotemporal region. The core of Nash equilibrium is that each device's decision is the optimal response to the decisions of other devices. Combining the dynamic operation characteristics of multiple devices in a smart construction site, an iterative solution method is used to solve the Nash equilibrium, adapting to real-time changes in device states. First, based on the initial revenue weights and potential functions of each device, the initial decision (pass / wait) of each device in the current spatiotemporal region is calculated.

[0127] Secondly, the decision of each device is iteratively updated until the decisions of all devices no longer change (reaching Nash equilibrium). At this point, the decision of each device is the optimal decision.

[0128] Finally, based on the optimal decision and the results of the benefit function calculation, a passage priority is assigned to each device (priority range 1-10, with 10 being the highest). The higher the priority, the more priority the device has in passing through that spatiotemporal region.

[0129] During the solution process, it iterates every 100ms (consistent with the update cycle of the bounding volume tree) to ensure that the access priority can adapt to changes in equipment status and work scenario in real time, and avoid access right conflicts caused by decision lag.

[0130] Furthermore, the priority is transformed into dynamic right-of-way characteristics: with time + spatial area + device ID + access priority as the core, such as (t5, area C, pouring truck 1, priority 9) and (t5, area C, excavator 2, priority 3).

[0131] Standardization can also be performed: the time dimension is accurate to milliseconds (e.g., t5=10:05:00.000), keeping it synchronized with the equipment positioning time sequence data; the spatial area adopts the spatial coordinate system of the BIM model, clearly defining the coordinates of the area boundaries (e.g., the coordinate range of area C is X∈[100,120]m, Y∈[50,70]m, Z∈[0,5]m), ensuring accurate spatial positioning; the equipment ID uses a unique identifier (e.g., pouring truck 1, excavator 2), corresponding one-to-one with the equipment status data; the access priority is dynamically adjusted based on the urgency of the operation and the result of the benefit function, such as when the urgency of the pouring truck operation increases, the priority is increased from level 9 to level 10 in real time.

[0132] In other embodiments, dynamic access rights features can be associated with spatiotemporal collaborative constraint features to ensure that access priorities are consistent with process requirements. For example, if only pouring trucks are allowed to pass through a certain area at time t5, then the access priority of other equipment in that area is set to level 1 (lowest) to avoid invalid access decisions.

[0133] The process of obtaining preferred cooperative path features by topological homotopy removal of the initial collision-free path cluster features and dynamic right-of-way features includes the following steps: S421. Based on dynamic right-of-way characteristics, identify a conflict-free subset of paths from the initial collision-free path cluster.

[0134] First, traverse each path in the initial collision-free path cluster and extract the complete spatiotemporal information of each path—splitting the path by timestamp (keeping the time precision consistent with the dynamic right-of-way feature, accurate to milliseconds) and clarifying the spatial area coordinates occupied by each path at each timestamp (based on the BIM model UTM coordinate system, accurately corresponding to the specific area of ​​the construction site).

[0135] Subsequently, by associating the dynamic right-of-way feature, the highest priority device type and ID corresponding to the spatiotemporal region (a certain timestamp + a certain spatial region) are queried. If the priority of the device corresponding to the current path is lower than that of the highest priority device in the spatiotemporal region, and the two overlap in the spatiotemporal region (i.e., two or more devices need to pass through the same time and the same spatial region), it is marked as a conflicting path. For example, path A corresponds to an excavator (priority 0.8), which occupies region D (coordinate range X∈[150,160]m, Y∈[80,90]m) at time t6 (10:06:00.000). However, the dynamic right-of-way feature shows that the highest priority device in region D at time t6 is a concrete pouring truck (priority 1.0), and the planned path of the concrete pouring truck also needs to occupy region D at time t6. At this time, path A and the path of the concrete pouring truck form a right-of-way conflict, and path A is marked as a conflicting path. If the highest priority device in the spatial region occupied by a path at all timestamps is the device corresponding to the path, or no other device needs to occupy the spatiotemporal region, it is determined to be a non-conflicting path.

[0136] After completing the conflict detection for all paths, a list of conflicting paths is established. All conflicting paths in the list are removed at once to prevent them from entering the subsequent screening process. For conflict-free paths, a path subset is formed by summarizing them. At the same time, the device ID, access priority, and corresponding spatiotemporal occupancy information of each path are recorded to ensure that all paths in the path subset have the highest or unique access rights in their respective access time and spatial area, without any access rights conflicts. This provides a high-quality path foundation for subsequent security potential field value assessment.

[0137] S422. Evaluate the safety potential field value of the path subset based on the comprehensive safety potential field characteristics to obtain low-risk paths.

[0138] First, each path in the path subset is split into all spatial nodes on the path by a fixed step size (step size = 0.5m, consistent with the spatial resolution of the integrated security potential field feature). Each node corresponds to a spatial location in the integrated security potential field feature.

[0139] Subsequently, the safety potential value corresponding to each node is extracted (by querying the 3D safety potential field to ensure that the node coordinates and the potential field coordinates are accurately matched); finally, the arithmetic mean of the potential field values ​​of all nodes on the path is calculated, which is the average safety potential value of the path.

[0140] During the calculation, abnormal nodes on the path (such as nodes with abnormal fluctuations in potential field value due to sensor noise) need to be excluded, and the median is used to replace outliers to ensure that the average potential field value can truly reflect the overall safety risk level of the path. For example, if a path contains 100 nodes and the potential field value of each node is distributed between 0.3 and 0.9, after removing two outliers (1.8 and 2.0), the mean of the remaining 98 nodes is calculated to obtain the average safety potential field value of the path.

[0141] Furthermore, considering the safety requirements of multi-device collaboration in smart construction sites, and after verification using extensive on-site measurement data, a threshold of 0.7 for the average safety potential field value was set. The average safety potential field value of each path within the path subset was compared to the threshold of 0.7: if the average potential field value > 0.7, it indicates a high overall safety risk for the path, potentially traversing yellow or red risk zones, and it is therefore excluded; if the average potential field value < 0.7, it is excluded. This indicates that the overall safety risk of this path is low, as it mainly traverses the green low-risk area, and therefore this path should be retained. Simultaneously, the average potential field value of the retained path is recorded to provide auxiliary reference for subsequent ranking of collaborative benefits.

[0142] S423. Rank the low-risk paths using synergistic benefits to generate preferred synergistic path features.

[0143] The remaining low-risk paths are sorted by synergistic benefits (such as operational compatibility with other equipment), and the top N paths are retained (N = number of equipment × 2): The benefits of collaboration mainly include three core indicators: operational coordination degree, process connection efficiency, and overall energy saving, satisfying the following formula: Collaboration Benefit = 0.5 × Operational Coordination Degree + 0.3 × Process Connection Efficiency + 0.2 × Energy Saving Rate. Among these, operational coordination degree... The calculation method is as follows: for the current device path Path with another cooperating device Calculate the minimum safe distance between the two paths in space and time. Over time, ,in To establish a safe distance, The attenuation coefficient is used in this formula to ensure a high score when the ideal cooperative distance is maintained between paths. Process connection efficiency. The calculation method is as follows: ,in This is the planned completion time for the tasks along this route. This refers to the planned start time for subsequent related processes. Energy saving rate. The calculation method is as follows: ,in This is the baseline energy consumption for this type of task, based on historical data statistics. This is the estimated energy consumption for this route based on the route length, gradient, and number of starts and stops.

[0144] Next, the collaborative benefit value of each remaining path is calculated and sorted from highest to lowest. Finally, the number of paths to retain, N (N = number of equipment × 2), is determined based on the number of equipment on the construction site. The top N paths are retained—ensuring optimal path collaboration while maintaining a certain level of path diversity to prevent subsequent trajectory optimization from failing to adapt to real-time operational changes due to a single path. For example, if there are 3 pieces of equipment on the construction site (1 concrete pouring truck, 1 excavator, and 1 dump truck), then N = 3 × 2 = 6, and the top 6 paths with the highest collaborative benefits are retained.

[0145] Furthermore, the path features are associated with dynamic right-of-way and safety potential field features to generate optimal collaborative path features (path coordinates + right-of-way priority + safety potential field value).

[0146] Specifically, the core features of each retained path are extracted: path coordinates (three-dimensional spatial coordinates sorted by timestamp, accurate to 0.1m, adaptable to equipment control precision), the device ID and right-of-way priority corresponding to the path (extracted from dynamic right-of-way features, clarifying the right-of-way priority of each path in each spatiotemporal region), the average safety potential field value of the path, and the potential field value of each node (extracted from comprehensive safety potential field features, clarifying the safety risks at each location of the path). These features are integrated to form standardized preferred collaborative path features. Each feature entry includes device ID, path coordinate sequence, timestamp sequence, right-of-way priority sequence, average safety potential field value, and node potential field value sequence, which not only clarifies the spatial direction and temporal arrangement of the path, but also marks its right-of-way priority and safety risks, realizing a three-dimensional association between path-collaboration-security.

[0147] S5. Combining the dynamic right-of-way characteristics, the comprehensive security potential field characteristics, and the preferred collaborative path characteristics, multi-device collaborative path planning is completed using spatiotemporal corridor secondary planning.

[0148] In one embodiment, the step of combining the dynamic right-of-way features, the comprehensive security potential field features, and the preferred cooperative path features to complete multi-device cooperative path planning using spatiotemporal corridor secondary planning includes the following steps: S51. Taking the optimal collaborative path as the core, and combining the geometric characteristics of equipment, perception uncertainty and dynamic right-of-way constraints, a spatiotemporal corridor is constructed.

[0149] Centered on the preferred collaborative path, a spatial corridor is formed by extending along the path normal (width = equipment enclosure size + safety distance, safety distance = perceived uncertainty level × 0.3m): First, based on the centerline coordinate sequence of the preferred collaborative path, the direction and spatial distribution of the path are clarified. Then, the expansion rules of the spatial corridor are determined—expanding bidirectionally along the normal direction of the path centerline (the direction perpendicular to the path direction), with the expansion width consisting of the equipment enclosure size plus a dynamic safety distance, to ensure that the equipment will not collide with surrounding obstacles or other equipment during its movement.

[0150] The dimensions of the equipment enclosure are based on the maximum length, width, and height of the equipment's spatiotemporal enclosure (e.g., if the excavator's overall width is 3m, 3m is used as the baseline for the enclosure dimensions). The dynamic safety distance is calculated using the perceived uncertainty feature map, satisfying the following: Safety distance = Perceived uncertainty level × 0.3m. For uncertainty level 0 (no uncertainty), the safety distance is 0.3m (basic safety buffer); for level 5 (completely unreliable), the safety distance is 1.5m (maximum safety buffer), avoiding collision risks caused by perceived unreliability. For example, when an excavator (enclosure width 3m) travels in an area with uncertainty level 2, the spatial corridor width is 3m + (2 × 0.3m) = 3.6m, extending 1.8m in each direction along the path normal to form sufficient safety buffer space. Simultaneously, the spatial corridor needs to be associated with the comprehensive safety potential field characteristics. If a section of the path crosses a medium-to-high-risk area, an additional 0.5m safety distance is added to further enhance safety.

[0151] Furthermore, a spatiotemporal corridor is formed by extending along the time axis: the time range of the corridor = the time interval during which the right-of-way is allowed, and the speed constraint = the maximum speed of the equipment dynamics. The core of the spatiotemporal corridor is spatiotemporal binding, which requires precise association between the spatial corridor and the time dimension to ensure that the equipment can travel along the spatial corridor within the specified time and avoid conflicts in the right-of-way in the time dimension.

[0152] Specifically, the time range is directly derived from the dynamic right-of-way characteristics, i.e., the allowed passage time intervals for the equipment corresponding to the path in each spatiotemporal region. This is integrated to form the overall time range of the spatiotemporal corridor, and the time precision is consistent with the dynamic right-of-way characteristics (accurate to milliseconds) to ensure that the passage time of other equipment does not overlap. For example, if the right-of-way time interval for the pouring truck corresponding to the preferred collaborative path in region E is t10-t15, then the spatiotemporal corridor time range corresponding to this path is t10-t15, and the pouring truck is prohibited from entering the spatial corridor outside this time range. Speed ​​constraints are set based on equipment dynamic parameters, combined with the operating characteristics of construction site equipment, to determine the maximum travel speed of various types of equipment (e.g., the maximum speed of an excavator). Maximum speed of the concrete pouring truck This serves as the time axis velocity boundary of the spacetime corridor, ensuring that the speed of the equipment does not exceed its own dynamic limits when moving within the corridor, thus avoiding mechanical damage or collision risks caused by excessive speed. Simultaneously, a 500ms buffer time needs to be reserved for the speed transition during equipment startup and braking, ensuring smooth trajectory connections.

[0153] S52. Based on the spatiotemporal corridor, solve the multi-device collaborative path using the objective function and constraints of multi-objective quadratic programming.

[0154] A multi-objective quadratic programming objective function is constructed, incorporating energy consumption, trajectory curvature, and tracking error into a unified optimization system.

[0155] Specifically, the objective function satisfies: Energy consumption cost Curvature cost Tracking error cost), where: Energy consumption cost: Based on the hydraulic / motor energy consumption model of the equipment, the total energy consumption of the trajectory is calculated. Combining the power characteristics of construction machinery in smart construction sites, energy consumption models for hydraulic equipment (such as excavators and loaders) and motor equipment (such as small forklifts) are constructed respectively. The energy consumption model of hydraulic equipment combines the pressure, flow rate and operation time of the hydraulic system to quantify, satisfying: Energy consumption cost = ∫(hydraulic system pressure × flow rate × time) dt (the integration interval is the total time of the trajectory); the energy consumption model of motor equipment combines the power of the motor and the operation time to quantify, satisfying: Energy consumption cost = rated power of the motor × operation time × load factor (the load factor is set according to the operation intensity of the equipment, with a value of 0.6-1.0).

[0156] Curvature cost: the sum of squares of trajectory curvature (minimizing curvature for smoothness). Trajectory curvature directly determines the smoothness of equipment movement. Greater curvature leads to more abrupt equipment turning, increasing the risk of mechanical wear, operational instability, and even collisions. Therefore, curvature cost is calculated by summing the squares of curvature at all sampling points on the trajectory, satisfying: Curvature cost (curvature i) (i is the sequence number of the trajectory sampling point, and the sampling step size is consistent with the spatial corridor step size, set to 0.5m). By minimizing the sum of squares of curvature, the trajectory can be made as smooth as possible, reducing violent movements of equipment joints and chassis, extending the service life of the equipment, improving the stability of equipment control, and ensuring that the trajectory can be tracked in real time.

[0157] Tracking error cost: The sum of squared positional deviations between the trajectory and the optimal collaborative path. The optimal collaborative path is the best collaborative path selected through multiple rounds of screening. Excessive tracking error can cause equipment to deviate from the planned route, potentially leading to right-of-way conflicts, collision risks, or process deviations. The tracking error cost is calculated as follows: extract the spatial coordinates of each sampling point on the optimized trajectory, compare them with the coordinates of the corresponding timestamp on the optimal collaborative path, calculate the positional deviation (Euclidean distance), and then sum the squares of all deviation values. By minimizing the tracking error cost, it ensures that the optimized trajectory closely matches the optimal collaborative path, guaranteeing the consistency of multi-device collaborative operations.

[0158] Weights (set to weights in construction site scenario) , , (Prioritizing low energy consumption). The weighting is strictly aligned with the operational needs of smart construction sites—construction machinery on construction sites has high energy consumption (e.g., excavators can consume 10-20L of fuel per hour), and low energy consumption is key to reducing operating costs; therefore, energy consumption cost is given the highest weight. The smoothness of the trajectory directly affects the equipment's lifespan and operational stability, thus assigning a cost weight to curvature. Tracking error needs to be controlled within a reasonable range, but its priority is lower than energy consumption and smoothness, and it is given weight. Meanwhile, the weights can be dynamically adjusted according to the construction site operation scenario. For example, in high-precision operation scenarios such as pouring concrete, the weights can be adjusted accordingly. Adjusted to 0.3. Adjust to 0.4 to ensure tracking accuracy is prioritized; for extensive operation scenarios such as earthwork excavation, the original weight can be maintained, and energy consumption can be controlled first.

[0159] Furthermore, embedded constraints are applied: the trajectory must be within the spatiotemporal corridor (geometric constraints), and the device's acceleration / angular velocity must be specified. Matching maximum value (dynamic constraint) and right-of-way priority (cooperative constraint): Geometric constraints: All sampling points on the optimized trajectory must fall within the spatiotemporal corridor and be at least 0.1m away from the corridor boundary (with a small buffer). If a sampling point exceeds the corridor range, it is considered a violation of the constraints and needs to be re-optimized to ensure that the trajectory has no risk of collision.

[0160] Dynamic constraints: Based on the equipment's dynamic parameters, determine the maximum acceleration, maximum deceleration, and maximum angular velocity of joints for various types of equipment (e.g., the maximum acceleration of an excavator chassis). Maximum angular velocity of the boom During the optimization process, it is necessary to ensure that the acceleration and angular velocity of the trajectory do not exceed the set thresholds to avoid equipment damage caused by exceeding the mechanical performance of the equipment; at the same time, speed continuity constraints should be added to ensure that the trajectory speed transitions smoothly without sudden changes to avoid impacts during equipment start-up and braking. Collaborative constraints: The timing of the trajectory must be fully matched with the dynamic right-of-way characteristics. If the passage time of a certain trajectory exceeds the allowed passage range of the corresponding area, or the passage priority is lower than that of other devices, it is considered a violation of the constraint. The timing parameters of the trajectory need to be adjusted (such as slowing down the speed or adjusting the passage order) to ensure that no right-of-way conflict occurs and to ensure smooth collaborative operation of multiple devices.

[0161] Then, the objective function is solved using a fast QP solver (such as OSQP) to obtain a smooth trajectory. During the solution process, the objective function and constraints are transformed into a standard QP problem form, satisfying: Where x is the optimization variable (coordinates, velocity, and acceleration of the trajectory sampling point), P and q are the objective function coefficient matrices, and A and b are the constraint condition coefficient matrices; the trajectory obtained after solving must satisfy all constraints and minimize the objective function value, that is, achieve the optimization objectives of low energy consumption, high smoothness, and high precision.

[0162] Simultaneously, the smoothness of the solved trajectory is checked. If the curvature of a certain segment of the trajectory changes abruptly (e.g., curvature value > 0.5m), the smoothness is checked. -1 If the weights or constraints need to be readjusted, the solution needs to be applied again to ensure that the trajectory is smooth and traceable.

[0163] In other embodiments, the trajectory is updated in real time at a 10ms cycle. In smart construction site scenarios, equipment status (such as position and speed), perceived environment (such as dust concentration and obstacle location), and right-of-way priority may all change in real time. If the trajectory is not updated for a long time, it may become out of sync with the actual scene, leading to collision risks. Therefore, the trajectory update cycle is set to 10ms to match the equipment control cycle: real-time equipment status data (GNSS / IMU positioning, encoder data), perception uncertainty feature map, and dynamic right-of-way features are collected every 10ms to update the boundary parameters of the spatiotemporal corridor and the coefficients of the objective function; then, the OSQP solver is used to resolve the problem to obtain a real-time trajectory adapted to the current scene; the updated trajectory is then sent to the equipment control system to achieve real-time trajectory tracking.

[0164] Please see Figure 2In an embodiment, to efficiently execute the multi-device cooperative path planning method considering geometric and kinematic constraints provided by the present invention, the present invention also provides a multi-device cooperative path planning system considering geometric and kinematic constraints, comprising: an input device 1, an output device 2, a processor 3, and a memory 4, wherein the input device 1, output device 2, processor 3, and memory 4 are interconnected, and the memory 4 stores program instructions for executing the steps of the multi-device cooperative path planning method considering geometric and kinematic constraints. The multi-device cooperative path planning system considering geometric and kinematic constraints of the present invention has a compact structure and stable performance, and can stably execute the multi-device cooperative path planning method considering geometric and kinematic constraints of the present invention, further improving the overall applicability and practical application capability of the present invention.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the present invention.

Claims

1. A multi-device cooperative path planning method considering geometric and kinematic constraints, characterized in that, Includes the following steps: Dynamic oriented bounding volume tree processing is performed on multi-source point cloud data to obtain the spatiotemporal bounding volume features of equipment or obstacles. By using differential flat mapping to map the motion data of engineering machinery, passable manifold features are obtained. By integrating the spatiotemporal bounding volume features, the traversable manifold features, and the perceived uncertainty feature map, a comprehensive security potential field feature is obtained. Based on spatiotemporal collaborative constraints, homotopy-class sampling optimization is used to process the semantic map of the construction site and the features of the accessible manifold to obtain the initial collision-free path cluster features. The initial collision-free path cluster features and dynamic right-of-way features are processed by topological homotopy elimination to obtain the preferred cooperative path features; Combining the dynamic right-of-way characteristics, the comprehensive security potential field characteristics, and the preferred collaborative path characteristics, multi-device collaborative path planning is completed using spatiotemporal corridor secondary planning.

2. The multi-device cooperative path planning method considering geometric and kinematic constraints according to claim 1, characterized in that, The process of performing dynamic directed bounding volume tree processing on multi-source point cloud data to obtain the spatiotemporal bounding volume features of devices or obstacles includes the following steps: Register multi-source point cloud data and construct directed bounding volumes according to device structure for different regions; By associating temporal point clouds with spatiotemporal attributes, a spatiotemporal bounding volume that can characterize the dynamic features of equipment or obstacles is formed; Construct a hierarchical bounding volume tree and obtain bounding volume features with spatiotemporal attributes through a dynamic update mechanism.

3. The multi-device cooperative path planning method considering geometric and kinematic constraints according to claim 1, characterized in that, The process of obtaining traversable manifold features by mapping the motion data of engineering machinery using differential flat mapping includes the following steps: Based on the specific characteristics of engineering machinery, we construct categorized, constrained, flattened kinematic models. Based on real-time motion data of the device, the traversable manifold in the spatiotemporal dimension is characterized by the flattened kinematic model. The dimensionality of the accessible manifold is reduced to obtain the features of the accessible manifold.

4. The multi-device cooperative path planning method considering geometric and kinematic constraints according to claim 1, characterized in that, The process of fusing the spatiotemporal bounding volume features, the traversable manifold features, and the perceived uncertainty feature map to obtain the comprehensive safe potential field features includes the following steps: Construct a multi-head attention layer and calculate attention weights for different feature dimensions; By combining the spatiotemporal bounding volume features, the traversable manifold features, the perceptual uncertainty feature map, and the corresponding attention weights, a comprehensive security potential field feature is obtained.

5. The multi-device cooperative path planning method considering geometric and kinematic constraints according to claim 4, characterized in that, Constructing the perceived uncertainty feature map includes the following steps: A visual radar fusion network is used to generate fused feature maps that combine geometric accuracy and semantic integrity. The perceptual uncertainty of the fused feature map is quantified based on the Monte Carlo dropout strategy; A perceptual uncertainty feature map is generated through the perceptual uncertainty.

6. The multi-device cooperative path planning method considering geometric and kinematic constraints according to claim 1, characterized in that, The method of obtaining initial collision-free path cluster features based on spatiotemporal collaborative constraints and using homotopy-class sampling optimization to process the semantic map of the construction site and the features of the accessible manifold includes the following steps: Based on spatiotemporal collaborative constraints, the semantic map of the construction site is integrated with the traversable manifold to obtain a feasible area in the spatiotemporal dimension; By using homotopy partitioning and semantic priority sampling, the feasible region is divided into multiple topologically equivalent path families; The path family is screened using kinematic feasibility verification and path optimization to obtain the initial collision-free path cluster features.

7. The multi-device cooperative path planning method considering geometric and kinematic constraints according to claim 1, characterized in that, The process of obtaining preferred cooperative path features by topological homotopy removal of the initial collision-free path cluster features and dynamic right-of-way features includes the following steps: Based on dynamic right-of-way characteristics, a conflict-free subset of paths is identified from the initial collision-free path cluster; Based on the comprehensive safety potential field characteristics, the safety potential field values ​​of the path subset are evaluated to obtain low-risk paths; The low-risk paths are ranked using synergistic benefits to generate preferred synergistic path features.

8. The multi-device cooperative path planning method considering geometric and kinematic constraints according to claim 7, characterized in that, The analysis of the dynamic right-of-way characteristics includes the following steps: Treat each device as a game participant and construct a payoff function; The potential field function is obtained by characterizing the repulsive and attractive forces between devices. By combining the aforementioned payoff function and the aforementioned potential field function, the dynamic passage right characteristics are solved.

9. The multi-device cooperative path planning method considering geometric and kinematic constraints according to claim 1, characterized in that, The method of combining the dynamic right-of-way characteristics, the comprehensive security potential field characteristics, and the preferred cooperative path characteristics to complete multi-device cooperative path planning using spatiotemporal corridor secondary planning includes the following steps: Based on the optimization of collaborative paths, and combined with equipment geometric characteristics, perception uncertainty and dynamic right-of-way constraints, a spatiotemporal corridor is constructed. Based on the aforementioned spatiotemporal corridor, the multi-device collaborative path is solved using a multi-objective quadratic programming objective function and constraints.

10. A multi-device cooperative path planning system considering geometric and kinematic constraints, characterized in that, The multi-device collaborative path planning system considering geometric and kinematic constraints includes: an input device, an output device, a processor, and a memory, wherein the input device, output device, processor, and memory are interconnected, and the memory stores program instructions for executing the multi-device collaborative path planning method considering geometric and kinematic constraints as described in any one of claims 1-9.

Images