Unmanned aerial vehicle AI image recognition grading method for expressway construction period

Abstract

The invention relates to the technical field of image recognition, in particular to an unmanned aerial vehicle AI image recognition grading method for a highway construction period, which comprises the following steps of: extracting a highway live-action three-dimensional point cloud from an unmanned aerial vehicle image, calling a project design building information model, and calculating to obtain a multi-modal model set under an aligned coordinate system; and segmenting the live-action three-dimensional point cloud in the multi-modal model set under the aligned coordinate system to obtain a semantic live-action component set. According to the method, the analysis of construction behavior logic and process rhythm is deepened, the collaborative compactness and workflow efficiency among the operation units can be quantitatively evaluated, and a process connection problem and a potential efficiency bottleneck which are difficult to find by traditional routing inspection are revealed; finally, the concrete geometric deviation judgment result and the abstract construction rhythm index are subjected to integrated operation, a global decision view considering the result quality and the process health is provided for managers, and the scientificity of project supervision is improved.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a drone AI image recognition and classification method for highway construction. Background Technology

[0002] Image recognition is a branch of artificial intelligence. Its core goal is to train computer systems to "understand" and extract meaningful information from digital images or videos.

[0003] Current image recognition technologies, when applied to engineering, focus on analyzing visual features within single frames or continuous images to identify the category or shape of specific objects. For example, image recognition can accurately identify highway bridge piers and even calculate their dimensions through 3D reconstruction. However, because it cannot determine whether the spatial location or principal axis direction of the pier conforms to design specifications during processing, the evaluation results remain at the level of existence confirmation, failing to delve into quantitative judgments of compliance. Minor but critical geometric deviations generated during construction are easily overlooked until they accumulate into significant quality hazards. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a drone AI image recognition and classification method for highway construction.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a drone AI image recognition and classification method for highway construction, comprising the following steps: Extract the real-scene 3D point cloud of the highway from the drone imagery, retrieve the building information model of the project design, calculate the multimodal model set in the aligned coordinate system, segment the real-scene 3D point cloud in the multimodal model set in the aligned coordinate system, and obtain the semantic real-scene component set. Based on the semantic real-world component set and the component information in the building information model, the existence of each component is compared one by one. Missing or extra components in the real-world are filtered out, and a component existence deviation list is established. Based on the component existence deviation list, for components that exist in the model, the Euclidean distance of the geometric center, the included angle of the principal axis direction, and the differences in length, width and height are calculated to generate hierarchical construction deviation quantitative data. From the continuous drone video stream during the highway construction period, key work units, including asphalt pavers, road rollers, and dump trucks, are detected and continuously tracked to form a time-series work element event sequence. Based on the order and logical relationship of each event in the time-series work element event sequence, nodes and directed edges are constructed to establish a dynamic construction workflow graph. Based on the dynamic construction workflow diagram, the collaborative work tightness of related work units is quantified to obtain a set of construction rhythm indicators. The hierarchical construction deviation quantification data is compared and judged with the preset construction specification tolerance standards. The set of construction rhythm indicators and the judgment results are integrated for calculation to generate a comprehensive construction compliance level.

[0006] Preferably, the step of obtaining the semantic real-world component set is as follows: Feature points are extracted from UAV images and back-projected to generate a real-world 3D point cloud. The topology of the project design building information model is loaded, and the center point of the bridge pier and the edge line of the roadbed are extracted from the real-world 3D point cloud. The key structural points of the building information model are obtained, and the coordinate naming rules and field formats are unified to form an aligned input feature set. Based on the aligned input feature set, a one-to-one mapping between the pier center point and the key points of the structure is established. The coordinate transformation matrix composed of rotation and displacement terms is calculated and applied to the real-world 3D point cloud and building information model. Abnormal matches are eliminated according to the key point residual threshold and the coordinate transformation matrix is ​​calculated again. When the residual change is lower than the preset ratio, the matrix is ​​solidified to form a multimodal model set under the aligned coordinate system. Based on the multimodal model set in the aligned coordinate system, the real-world 3D point cloud is segmented block by block according to normal distribution, planar continuity and boundary curvature in the aligned coordinate system. The geometric range and connection topology of the walls, beams, slabs and columns of the building information model are combined to assign categories. Isolated fragments without assigned categories are removed and overlapping boundary areas are merged. The component category label and the component number of each point are output to generate a semantic real-world component set.

[0007] Preferably, the step of obtaining the list of component existence deviations is as follows: Based on the semantic real-world component set and the component information in the building information model, the existence status is checked item by item according to component number and component type. Components not appearing in the semantic real-world component set are marked as missing, and component information not appearing in the building information model is marked as extra, thus generating a component existence deviation list.

[0008] Preferably, the steps for obtaining the hierarchical construction deviation quantification data are as follows: Based on the component existence deviation list, components that exist simultaneously in the semantic real-world component set and the building information model are selected, and their geometric center coordinates, principal axis direction, length, width and height dimensions are extracted. The Euclidean distance of the geometric center, the included angle of the principal axis direction and the length, width and height dimension differences are calculated for each pair to form a detailed list of Euclidean distance of the geometric center, included angle of the principal axis direction and length, width and height dimension differences. Based on the Euclidean distance from the geometric center, the included angle along the principal axis, and the details of length, width, and height differences, the connection topology of the component information in the building information model is called up to check whether the connection relationship between adjacent components conforms to the design. The difference values ​​are statistically analyzed and aggregated according to the component level to generate hierarchical construction deviation quantitative data.

[0009] Preferably, the step of obtaining the time-series job meta-event sequence is as follows: Decode the frame sequence of continuous UAV video streams from highway construction period, read and number the timestamp of each frame, locate the candidate vehicle area, and compare it item by item according to the aspect ratio of the rectangle, the number of tracks or tires, the opening and closing angle of the hopper, and the edge contour of the screed to confirm that it is an asphalt paver, a road roller and a dump truck. Connect the trajectory of the same target according to the distance between the center points of adjacent frames and the contour similarity, and statistically analyze the speed, hopper opening and closing, vibration mark and unloading gate status according to continuous time window to mark the operation status and generate a time-series operation meta-event sequence.

[0010] Preferably, the steps for obtaining the dynamic construction workflow map are as follows: Based on the time-series job event sequence, a node is created for each event, recording the equipment identifier, job status, and timestamp. The consistency of equipment identifiers and the adjacency of processes of adjacent events are checked in chronological order. If the conditions are met, a directed edge is established. The difference in timestamps between adjacent events is calculated and written into the edge weight. Pairs that do not meet the conditions are discarded, forming a set of nodes and directed edges. Based on the set of nodes and directed edges, the nodes are arranged hierarchically according to equipment category, and the nodes of the same equipment are sorted topologically according to timestamp. Duplicate directed edges are merged and edge weights are accumulated. Directed edges that cross the process sequence are deleted, and continuous directed edge paths according to timestamp are retained. The starting point index and ending point index of the path are registered to form a dynamic construction workflow graph.

[0011] Preferably, the steps for obtaining the construction rhythmicity index set are as follows: The event records of each work unit are read from the dynamic construction workflow diagram. Data segments with the event type of waiting events are filtered out, the start and end times are extracted and the duration is calculated. The average duration of all waiting events is calculated to obtain the average duration of waiting events. At the same time, the start and end points of key operation cycles are identified on the same time axis. The deviation of the duration of each cycle is calculated and the standard deviation is calculated to generate the average duration of waiting events and the standard deviation of key operation cycle time. Calculate the dynamic rhythm coupling degree between work units based on the average duration of the waiting events and the standard deviation of the critical task cycle time; Based on the dynamic rhythm coupling degree, combined with the average duration of waiting events and the standard deviation of critical operation cycle time, the dynamic rhythm coupling degree of each work unit is compared pairwise. The units are grouped by equipment category and the average value within each group is calculated. The average value of each group is then summarized with the average duration of waiting events and the standard deviation of critical operation cycle time to form a set of construction rhythm indicators.

[0012] Preferably, the steps for obtaining the comprehensive construction compliance level are as follows: The deviation value of each component is read from the hierarchical construction deviation quantification data, and compared with the tolerance standard of the construction specification item by item to determine whether it exceeds the tolerance limit. If it exceeds the limit, it is recorded as a non-conforming component and classified into serious defects, general defects and minor defects according to the severity level of the defect. The number of non-conforming components of each level and the total number of components are counted to generate a quality judgment result containing the defect classification number and the total number of components. Based on the quality judgment results, the average duration of waiting events, the average value of key operation cycle time, the standard deviation of key operation cycle time, and the average value of dynamic rhythm coupling degree in the construction rhythmic index set are extracted and organized into a comprehensive input parameter set. The overall construction compliance level is calculated based on the comprehensive input parameter set.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention extracts real-world 3D point clouds and aligns them with the building information model (BIM) of the project design to construct a multimodal model set. This creates a unified digital comparison space between the construction entity and the design blueprint, enabling comprehensive quantitative deviation analysis of the construction entity from macroscopic existence to microscopic geometric dimensions. It transforms the traditional quality control model, which relies on manual sampling, into a data-driven, traceable assessment. Furthermore, it detects and tracks key work units from continuous UAV imagery and constructs a dynamic construction workflow map based on the temporal operation status. This goes beyond simple object recognition, delving into the analysis of construction behavior logic and process rhythm. It can quantitatively assess the coordination tightness and workflow efficiency between work units, revealing process connection problems and potential efficiency bottlenecks that are difficult to detect through traditional inspections. Finally, it integrates the concrete geometric deviation judgment results with the abstract construction rhythm indicators, providing managers with a holistic decision-making perspective that considers both result quality and process health, thus improving the scientific nature of project supervision. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] Please see Figure 1 This invention provides a technical solution for AI image recognition and classification methods for drones during highway construction, comprising the following steps: Extract the real-scene 3D point cloud of the highway from the drone imagery, retrieve the building information model of the project design, calculate the multimodal model set in the aligned coordinate system, segment the real-scene 3D point cloud in the multimodal model set in the aligned coordinate system, and obtain the semantic real-scene component set. Based on the semantic real-world component set and the component information in the building information model, the existence of each component is compared one by one. Missing or extra components in the real-world are filtered out, and a component existence deviation list is established. Based on the component existence deviation list, for components that exist in the model, the Euclidean distance of the geometric center, the included angle of the principal axis direction, and the differences in length, width and height are calculated to generate hierarchical construction deviation quantitative data. From the continuous drone video stream during the highway construction period, key work units are detected and continuously tracked. These key work units include asphalt pavers, road rollers, and dump trucks, forming a time-series work element event sequence. Based on the sequence of events and logical relationships in the time-series work element event sequence, nodes and directed edges are constructed to establish a dynamic construction workflow graph. Based on the dynamic construction workflow diagram, the collaborative work tightness of related work units is quantified to obtain a set of construction rhythm indicators. The hierarchical construction deviation quantification data is compared and judged with the preset construction specification tolerance standards. The construction rhythm indicator set and the judgment results are integrated for calculation to generate a comprehensive construction compliance level.

[0017] The steps for obtaining the semantic real-world component set are as follows: Feature points are extracted from UAV images and back-projected to generate a real-world 3D point cloud. The topology of the project design building information model is loaded, and the center point of the bridge pier and the edge line of the roadbed are extracted from the real-world 3D point cloud. The key structural points of the building information model are obtained, and the coordinate naming rules and field formats are unified to form an aligned input feature set. Based on the aligned input feature set, a one-to-one mapping between the pier center point and the key points of the structure is established. The coordinate transformation matrix composed of rotation and displacement terms is calculated and applied to the real-world 3D point cloud and building information model. Abnormal matches are eliminated according to the key point residual threshold and the coordinate transformation matrix is ​​calculated again. When the residual change is lower than the preset ratio, the matrix is ​​solidified to form a multimodal model set under the aligned coordinate system. Based on the multimodal model set in the aligned coordinate system, the real-world 3D point cloud is segmented block by block according to normal distribution, planar continuity and boundary curvature in the aligned coordinate system. The geometric range and connection topology of the walls, beams, slabs and columns of the building information model are combined to assign categories. Isolated fragments without assigned categories are removed and overlapping boundary areas are merged. The component category label and the component number of each point are output to generate a semantic real-world component set.

[0018] Specifically, feature points are extracted from each image using a scale-invariant feature transform algorithm from UAV imagery. Based on multi-view geometry principles, motion reconstruction technology is used for camera pose estimation and sparse point cloud reconstruction. Subsequently, a multi-view stereo matching algorithm is employed to encrypt the sparse point cloud, generating a real-world 3D point cloud containing millions to tens of millions of points. Simultaneously, the project design building information model, stored in an industrial basic class format, is retrieved, and its file structure is parsed to load the geometry, spatial location, and connection topology of the components. Then, the real-world 3D point cloud is preprocessed, and a density-based spatial clustering algorithm is applied to identify and segment independent structures such as bridge piers. The geometric centers of these clusters are calculated as the pier center points, and the points on the roadbed edge are fitted using a random sampling consensus algorithm to extract the edge point set representing the roadbed orientation. For the building information model, the component vertices, center points, or specific reference points that are predefined as structural key points are directly queried and extracted. Finally, a unified data structure is established for all extracted pier center points, roadbed edge points, and structural key points. This structure includes fields such as source identifier, point type, and 3D coordinates, and the naming rules and data formats of all coordinates are unified. For example, the right-handed Cartesian coordinate system is used and the coordinate unit is meters, forming an aligned input feature set containing point cloud features and model features.

[0019] Based on the aligned input feature set obtained in the preceding steps, a fast nearest neighbor search algorithm based on k-dimensional trees is first used to find the nearest structural key point in three-dimensional space for each pier center point, constructing an initial one-to-one mapping candidate set. Then, a random sampling consensus algorithm is used to iteratively optimize this candidate set. In each iteration, three mapping pairs are randomly selected, and the rotation and translation terms required for a rigid body transformation are calculated based on these three point pairs, forming a temporary coordinate transformation matrix. This matrix is ​​applied to all structural key points in the building information model, and the Euclidean distance between the transformed point and its corresponding pier center point in the candidate set, i.e., the residual, is calculated. The matching point pairs are then divided into inner and outer points based on the key point residual threshold. This key point residual threshold is set based on statistical analysis of historical data from similar projects, using the mean of the registration error plus one standard deviation as a benchmark. For example, if historical data shows an average registration error of 0.05 meters and a standard deviation of 0.02 meters, then the key point residual threshold is set to 0.05 + 1 * 0.02 = For a distance of 0.07 meters, any mapping pair with a residual less than 0.07 meters is considered an interior point. After the iteration, the coordinate transformation matrix with the most interior points is selected as the current optimal solution. All exterior point pairs that are judged as abnormal matches are removed. Using all the remaining interior point pairs, a more accurate coordinate transformation matrix is ​​recalculated using singular value decomposition. The root mean square error of the residuals of all interior point pairs after the transformation is calculated and compared with the root mean square error calculated in the previous iteration. When the rate of change between the two errors is less than a preset ratio, such as less than 0.5%, the algorithm is considered to have converged and the current coordinate transformation matrix is ​​solidified. Finally, the solidified matrix is ​​applied to the original real-world 3D point cloud and building information model to unify them under the same coordinate system, forming a multimodal model set under the aligned coordinate system.

[0020] Based on the multimodal model set in the aligned coordinate system, the segmentation of the aligned real-world 3D point cloud adopts a region growing algorithm. First, the local surface normal and curvature of each point in the point cloud are calculated. During calculation, a set of neighboring points within a radius of 0.2 meters around each point is selected, and the normal vector is estimated using principal component analysis. Next, an unsegmented point is randomly selected from the point cloud as a seed point to initiate a growth process. This seed point and its neighboring points constitute the initial region. Then, the neighboring points of the region boundary are iteratively examined. If the angle between the normal of a neighboring point and the average normal of the points within the region is less than a preset normal angle threshold of 15 degrees, and the distance from this point to the plane fitted by the points within the region is less than a plane continuity threshold of 0.03 meters, and the boundary curvature of this point is less than a curvature threshold of 0.1, then this neighboring point is merged into the current region. This process continues until no new points can be added, completing the extraction of a segmentation block. This process is repeated until all points are assigned to a segmentation block. After segmentation, each segment is assigned a category to a component in the Building Information Model (BIM). Specifically, the centroid of each segment is calculated, and it is determined whether the centroid falls within the bounding box of a component in the BIM, such as a wall, beam, slab, or column. Simultaneously, the topological information of inter-component connections recorded in the BIM is used for verification. For example, if two adjacent segments are assigned to a beam and a column connected in the model, respectively, the confidence level of the assignment is increased. Isolated segments that cannot be assigned to any component bounding box and have fewer than 1000 points are marked as noise and removed. Finally, adjacent segments with the same component category label are checked. If their boundary regions overlap or the distance is less than 0.05 meters, these segments are merged into a single complete component instance. All points in each merged segment are assigned a unified component category label and a unique component number, generating the final semantic real-world component set.

[0021] The steps for obtaining the list of component existence deviations are as follows: Based on the semantic real-world component set and the component information in the building information model, the existence status is checked one by one according to the component number and component type. Components that do not appear in the semantic real-world component set are marked as missing, and component information that does not appear in the building information model is marked as extra, thus generating a component existence deviation list.

[0022] Specifically, based on the semantic real-world component set generated in the previous step and the component information in the Building Information Model (BIM) loaded from the project design files, each component in the BIM is first traversed to extract its unique component number and component type. This information is then stored as key-value pairs in a hash lookup table. For example, the information of a bridge pier component is stored as {'P-01', The process begins by processing each component in the semantic real-world component set, extracting its component number and type. This information is then used to query the previously constructed hash lookup table. If a matching entry is found, the component is marked as "coexisting" and removed from the hash lookup table. If a component exists in the semantic real-world component set, such as a temporary support structure erected on-site, but its number does not exist in the hash lookup table, it is marked as "extra" and its geometric range and component type are recorded. After traversing all components in the semantic real-world component set, all remaining entries in the hash lookup table—those that exist in the design model but cannot be matched in the real-world 3D point cloud—are marked as "missing." Finally, the verification results of all components, including those marked as "coexisting," "extra," and "missing," along with their corresponding numbers and types, are compiled into a structured list to generate a component existence deviation list.

[0023] The steps for obtaining hierarchical construction deviation quantification data are as follows: Based on the component existence deviation list, components that exist simultaneously in the semantic real-world component set and the building information model are selected. The geometric center coordinates, principal axis direction, length, width and height dimensions are extracted. The Euclidean distance of the geometric center, the included angle of the principal axis direction and the length, width and height dimension differences are calculated for each pair to form a detailed list of the Euclidean distance of the geometric center, the included angle of the principal axis direction and the length, width and height dimension differences. Based on the Euclidean distance from the geometric center, the included angle along the principal axis, and the details of length, width, and height differences, the connection topology of the component information in the building information model is called up to check whether the connection relationship between adjacent components conforms to the design. The difference values ​​are statistically analyzed and aggregated according to the component level to generate hierarchical construction deviation quantitative data.

[0024] Specifically, based on the component existence deviation list, all components marked as "coexisting" are first selected. A data pair is created for each such component, containing its instance in the semantic real-world component set and its corresponding design instance in the Building Information Model (BIM). Next, the geometric attributes of each component pair are extracted and calculated. For real-world components, their geometric center coordinates are obtained by calculating the average of the coordinates of all 3D point clouds constituting the component. Their principal axis directions are determined by applying principal component analysis (PCA) to these point cloud coordinates and selecting the eigenvector corresponding to the largest eigenvalue. Their length, width, and height dimensions are obtained by projecting all point clouds onto the principal axis coordinate system obtained from PCA and calculating the maximum span of the point set along each axis. For BIM... In the model, the geometric center coordinates, principal axis directions, and length, width, and height dimensions of the design components are directly read as predefined attributes. Then, the geometric deviations between the real-world components and the design components are calculated one by one. The Euclidean distance of the geometric center is obtained by calculating the square root of the sum of the squares of the differences in the three-dimensional coordinates of the two center points. The angle between the principal axis directions is obtained by calculating the dot product of the unit vectors of the two principal axis directions and then taking their inverse cosine value. The differences in length, width, and height dimensions are calculated by subtracting the corresponding dimensions of the design components from the dimensions of the real-world components. Finally, the number of each coexisting component, the Euclidean distance of the geometric center, the angle between the principal axis directions, and the dimensional differences in the length, width, and height directions are organized into records to form a detailed list of the Euclidean distance of the geometric center, the angle between the principal axis directions, and the differences in length, width, and height dimensions.

[0025] Based on the Euclidean distance from the geometric center, the included angle along the principal axes, and the detailed differences in length, width, and height, the connection topology data attached to the component information in the Building Information Model (BIM) is first retrieved. This data stores the connection relationships between components in the form of an adjacency list. For example, it records that "pier P-01" and "cap beam C-01" are connected at a specific connection surface. Subsequently, for each pair of components defined as adjacent in the topology, such as P-01 and C-01, their respective geometric deviation values ​​are retrieved from the difference details. Based on these deviation values, the actual connection relationship is checked to see if it meets the design requirements. This check is determined by a connection tolerance standard, which is formulated according to the national "Highway Engineering Quality Inspection and Evaluation Standard." For example, the standard stipulates that the relative positional deviation between the pier and the cap beam shall not exceed 20 mm. A threshold of 20 mm is set for the deviation of the connection position. If the projection of the vector difference between the geometric centers of two components onto the connection direction exceeds this threshold, the connection is deemed non-compliant with the design. Next, the number of non-compliant connections is counted, and statistical indicators of various differences for all components, such as mean and variance, are calculated. Data is then aggregated from bottom to top according to the component hierarchy defined in the Building Information Model (BIM), such as the hierarchy of "component-part-structural unit". During the aggregation process, a weighted average method is used to calculate the comprehensive deviation of the upper-level structure. The weighting coefficients take into account both the structural importance and volume of the components. Structural importance is classified into levels 1 to 5 by the design engineer based on the component's load-bearing capacity and failure consequences. Volume is directly obtained from the model. An example of weight calculation is: Weight = 0.6 * (Structural Importance Level / 5) + 0.4 * (Component Volume / Average Volume of Components in the Level). In this way, the detailed deviation values ​​of the bottom-level components are summarized level by level, generating hierarchical quantitative data of construction deviations reflecting different granularities of construction deviations from individual components to the entire structural unit.

[0026] The steps for obtaining the event sequence of a time-series job are as follows: Decode the frame sequence of continuous UAV video streams from highway construction period, read and number the timestamp of each frame, locate the candidate vehicle area, and compare it item by item according to the aspect ratio of the rectangle, the number of tracks or tires, the opening and closing angle of the hopper, and the edge contour of the screed to confirm that it is an asphalt paver, a road roller and a dump truck. Connect the trajectory of the same target according to the distance between the center points of adjacent frames and the contour similarity, and statistically analyze the speed, hopper opening and closing, vibration mark and unloading gate status according to continuous time window to mark the operation status and generate a time-series operation meta-event sequence.

[0027] Specifically, from continuous drone video streams during highway construction, the H.264 or H.265 encoded video streams are decoded into a sequence of PNG image frames at 25 frames per second using the FFmpeg library. Simultaneously, the precise timestamp from the metadata of each frame is read and assigned a consecutive number starting from 0. Then, a pre-trained YOLOv5 object detection model is applied to each frame to locate all potential vehicle candidate regions. This model is trained on a dataset containing over 50,000 drone aerial images labeled with asphalt pavers, road rollers, and dump trucks. During training... The input images are uniformly 640x640 pixels in size. Data augmentation techniques such as random cropping, brightness and contrast adjustment are employed. The loss function combines localization loss, confidence loss, and classification loss, and the Adam optimizer is used for 200 epochs of iterative training until the mean accuracy on the validation set no longer improves. During model inference, all bounding boxes with a confidence score higher than 0.7 are output. Subsequently, for each detected bounding box, a rule-based classifier is applied for secondary verification. This classifier compares each feature against a preset feature template; for example, the aspect ratio of the rectangle representing an asphalt paver is set to range from 2.5 to 4.0. The roller's value is 1.5 to 2.2. Simultaneously, the number and shape of tires or tracks are detected using Hough transform. The unique trapezoidal contour of the screed is extracted and matched using the Canny edge detection algorithm. The opening angle of the hopper is determined by analyzing the color changes of pixels in the hopper area. Only when a candidate region meets more than 70% of the feature rules of its corresponding category is it finally confirmed as that type of equipment. For the confirmed equipment, the DeepSORT algorithm is used for cross-frame tracking. This algorithm initializes a Kalman filter for each newly detected target to predict its position in the next frame and calculates the predicted position relative to all detected targets in the next frame. The intersection-union ratio and cosine similarity of appearance features between test frames constitute a cost matrix. Then, the Hungarian algorithm is used to solve for the optimal match, thereby connecting the same target in consecutive frames into a trajectory. Finally, within a sliding time window of 5 seconds, the instantaneous velocity of the target is calculated based on the position changes of trajectory points and timestamp differences. The unloading gate status is determined by analyzing the angle changes between the dump truck's cargo box and chassis. Based on fixed operation logic, such as marking the vibration mark as open when the road roller moves at low speed in the paved area, these state change points and their attributes are marked as specific operation events, generating a time-series operation meta-event sequence.

[0028] The steps to obtain a dynamic construction workflow diagram are as follows: Based on the sequence of time-series job events, a node is created for each event, recording the equipment identifier, job status, and timestamp. The consistency of equipment identifiers and the adjacency of processes of adjacent events are checked in chronological order. If the conditions are met, a directed edge is established. The difference in timestamps between adjacent events is calculated and written into the edge weight. Pairs that do not meet the conditions are discarded, forming a set of nodes and directed edges. Based on the set of nodes and directed edges, nodes are arranged hierarchically according to equipment category, nodes of the same equipment are topologically sorted according to timestamp, duplicate directed edges are merged and edge weights are accumulated, directed edges that cross the process sequence are deleted, and continuous directed edge paths according to timestamp are retained. The starting point index and ending point index of the path are registered to form a dynamic construction workflow graph.

[0029] Specifically, based on the time-series operation event sequence, each event record in the sequence is first traversed, such as "timestamp T1, equipment ID is paver 01, operation status is start paving," and an independent node is created for it in the graph data structure. This node contains three core attributes: "equipment identifier," "operation status," and "timestamp." After creating all nodes, they are sorted in ascending order by timestamp. Then, two time-adjacent event nodes are processed sequentially, denoted as event A and event B. When processing each pair of nodes, two checks are performed. The first is an equipment identifier consistency check, which determines whether the equipment identifiers of event A and event B are the same. If they are the same, the process adjacency check is performed. This check refers to a predefined single-equipment state transition diagram, which is formulated according to the construction process specifications and clarifies all possible state transitions of a single equipment. For example, for a paver, the next legal state after the "waiting" state can be "move" or "pave," but not "end." If the state of event A to the state of event B is a valid transition line in the diagram, the process is considered complete. If the transfer path is valid, the condition is met. The second check applies to cases where equipment identifiers are inconsistent. In this case, the process adjacency check is also performed, but a predefined cross-equipment process flowchart is used as a reference. This flowchart defines the logical order of different equipment working together. For example, "unloading of the dump truck is completed" is a valid preceding process of "pavers start paving". If the process transition from event A to event B conforms to the definition of the flowchart, the condition is met. Only when the condition in any of the above checks is met is a directed edge from A to B established between the corresponding nodes of event A and event B, and the difference between the timestamps of the two events is calculated and written as the weight of the edge. For example, if T2 minus T1 is 30 seconds, the edge weight is 30. For any adjacent event pair that does not meet the above equipment identifier consistency or process adjacency conditions, it is judged as logically discontinuous or noisy data and is directly discarded without establishing a connection. After traversing and processing all adjacent event pairs, a set of nodes and directed edges containing all event nodes and the logical relationships between them is formed.

[0030] Based on the set of nodes and directed edges, the graph is first logically visualized and laid out. According to the three equipment categories—asphalt pavers, road rollers, and dump trucks—all nodes are divided into three horizontal "swimlanes" or layers, ensuring that all equipment nodes of the same type are located at the same layer. Next, within each equipment category layer, nodes are grouped according to specific equipment identifiers, such as "Road Roller 01" and "Road Roller 02," and all nodes of each equipment are strictly sorted topologically according to timestamps to ensure that the operational event chain of each equipment is unidirectionally progressive in time. Subsequently, the directed edges in the graph are organized and optimized. It is checked whether there are multiple directed edges with identical start and end points. If so, these duplicate directed edges are merged into one, and its weight is set to the sum of the weights of all merged edges. Then, a predefined standard process sequence list is used, for example... For example, in the process of "unloading → paving → initial compaction → secondary compaction", all directed edges that cross equipment types in the diagram undergo a validity review. Those edges that skip intermediate steps are deleted. For instance, if a directed edge in the diagram points directly from the "unloading dump truck" node to the "secondary compaction roller" node, it is considered illogical and deleted because it skips the two necessary steps of "paving" and "initial compaction". This step retains only event paths that strictly follow the preset construction process and are formed continuously according to timestamps. Finally, all remaining continuous directed edge paths in the diagram are identified and registered. For each path representing a complete work cycle or process connection, a unique index of its starting node and a unique index of its ending node are recorded. Finally, all nodes and directed edges that have undergone layering, sorting, merging, and filtering are integrated to form a structured and logical dynamic construction workflow diagram.

[0031] The steps for obtaining the construction rhythm index set are as follows: The event records of each work unit are read from the dynamic construction workflow diagram. Data segments with the event type of waiting events are filtered out, the start and end times are extracted and the duration is calculated. The average duration of all waiting events is calculated to obtain the average duration of waiting events. At the same time, the start and end points of the critical operation cycle are identified on the same time axis. The deviation of the duration of each cycle is calculated and the standard deviation is calculated to generate the average duration of waiting events and the standard deviation of the critical operation cycle time. The dynamic rhythm coupling degree between work units is calculated based on the average duration of waiting events and the standard deviation of critical task cycle time. The calculation formula is as follows: ; in, Work unit With work unit The dynamic rhythm coupling degree between them Work unit and work unit The number of time indices that are simultaneously in a job state within the same time window. Work unit and work unit The total number of time indices that are in any job state within the same time window. Work unit Relative to work unit The standard deviation of the start time delay is obtained by extracting the time difference between multiple consecutive "job A ends - job B begins" events and calculating the standard deviation. This is the average cycle time of the critical tasks, obtained by averaging the cycle times of each critical task. This is the rhythm sensitivity coefficient, used to adjust the sensitivity of the exponential penalty term to rhythm fluctuations; Based on the dynamic rhythm coupling degree, combined with the average duration of waiting events and the standard deviation of critical operation cycle time, the dynamic rhythm coupling degree of each work unit is compared pairwise. The units are grouped by equipment category and the average value within each group is calculated. The average value of each group is then summarized with the average duration of waiting events and the standard deviation of critical operation cycle time to form a set of construction rhythm indicators.

[0032] Specifically, from the dynamic construction workflow graph generated in the aforementioned steps, the node set representing each work unit in the graph is traversed. For each work unit, its complete event record stream is read in timestamp order, and all nodes whose event types are marked as "waiting," "idle," or "paused" are selected. These are the data segments of waiting events. For each selected waiting event, the timestamp of its starting node is extracted as the starting point of the start and end times, and the timestamp of the first non-waiting event node that follows it is extracted as the ending point. The duration of the wait is obtained by calculating the difference between these two timestamps. All waiting durations of all work units are collected, and their average is calculated to obtain the average duration of waiting events. At the same time, in the entire On a unified timeline of the construction process, the start and end points of critical operation cycles are identified according to the predefined construction process flow. A critical operation cycle is defined as a complete sequence that begins with the event "the dump truck begins unloading material onto the paver" and ends with the event "the last roller completes the final compaction of the paved area". By searching and matching these specific start and end event patterns in the graph, all independent critical operation cycle instances are located, and the duration of each cycle is calculated. Then, the average duration of all cycles is calculated, and the duration of each cycle is subtracted from the average to obtain its respective deviation. Finally, the standard deviation of these deviation values ​​is calculated to generate data containing two core indicators: the average duration of waiting events and the standard deviation of critical operation cycle time.

[0033] In the formula for calculating dynamic rhythm coupling, the first part This is the Jaccard similarity coefficient, used to measure the degree of overlap in working time between two work units, i.e., spatial synergy. A higher value indicates a greater proportion of time spent working simultaneously, and more opportunities for collaborative work. (Part Two) It is an exponential penalty term used to measure the temporal stability, or temporal rhythm, of the process connection between two work units. This reflects the volatility of process handover time delays; the greater the volatility, the heavier the penalty. The exponential design ensures that even small fluctuations will affect the final score. The coefficient modulates the sensitivity of this punishment.

[0034] Work unit and work unit The number of time indices simultaneously in a job state within the same time window is obtained through the following steps: First, the total observation duration of the entire UAV video stream, for example, 3600 seconds, is discretized into a time index sequence with 1-second units, i.e., divided into 3600 time slices. Then, based on the time-series job meta-event sequence, the job units are traversed. and homework The event logs are processed, and for each time slice, it is determined whether the unit is in an "operating state" (e.g., paving, compaction, unloading) or a "non-operating state" (e.g., waiting, moving). A Boolean state vector of length 3600 is generated for each unit, where 1 represents operation and 0 represents non-operation. Finally, a bitwise AND operation is performed on these two Boolean vectors, and the number of values ​​of 1 in the resulting vector is counted, which is the time index of both being in an operating state simultaneously. For example, within a 3600-second observation window, by analyzing the event logs, the operating unit... The total working time of the (paver) is 1800 seconds, and the working unit... The total operating time of the (road roller) is 1500 seconds. By comparing its state vectors, it was found that the time spent in the same operating state by both was 1200 seconds. Therefore... The value is 1200.

[0035] Work unit and work unit The total number of time indices in any job state within the same time window; the steps to obtain this parameter are the same as... The acquisition process is similar, also based on the Boolean state vectors of the two work units generated above. However, here a bitwise OR operation is performed instead of a bitwise AND operation. The number of values ​​of 1 in the result vector after the bitwise OR operation is counted, which gives the total number of time indices where at least one unit is in the work state. This value reflects the total time span of the two work units participating in the construction activities. Using the example of the previous parameter, the work unit... The operation time is 1800 seconds, and the operation unit is... The individual task has a time limit of 1500 seconds, and the time they work simultaneously is 1200 seconds. According to set theory, the total number of time indices for each task in any given task state is the sum of their individual task times minus the time spent working simultaneously, which is 1800 + 1500 - 1200 = 2100 seconds. Therefore... The value is 2100.

[0036] Work unit Relative to work unit The standard deviation of the start-up time delay is obtained by first filtering out all work units that meet the criteria from the dynamic construction workflow diagram. A certain task event ends, and its subsequent event is a task unit. A series of consecutive event pairs starting with "a certain job event begins" represent a complete job handover. The timestamp difference between each such pair of events is extracted, i.e., the start time delay. This forms a time-delay sample sequence. For example, within the observation period, eight handovers from the end of a paver's operation to the start of a roller's operation were identified, and the recorded time delay sequence is {45, 50, 42, 55, 48, 52, 46, 49} seconds. Then, the standard deviation of this sequence is calculated. First, the average delay time is calculated as (45+50+42+55+48+52+46+49) / 8 = 48.375 seconds. Then, the variance is calculated as [(45-48.375)]. 2 + ... + (49-48.375) 2 ] / 8 ≈ 13.73, and finally, taking the square root gives the standard deviation. Therefore The value is approximately 3.71 seconds.

[0037] This represents the average critical task cycle time. For example, by analyzing the entire construction process, 15 complete critical task cycles were identified, with durations of {320, 335, 310, 325, 340, 315, 330, 322, 318, 333, 328, 312, 338, 324, 319} seconds. Adding these values ​​together and dividing by 15 yields the average critical task cycle time. The value is 325.27 seconds.

[0038] The rhythm sensitivity coefficient is used to adjust the sensitivity of the exponential penalty term to rhythm fluctuations. Its setting is based on the project's quality control requirements and tolerance for process continuity stability. The setting process is as follows: First, the project management team defines an "acceptable maximum delay fluctuation ratio" based on construction specifications and experience. For example, it defines that when the standard deviation of the start-up time delay reaches 5% of the average value of the critical operation cycle, the rhythm stability score should drop to a specific threshold, such as 0.8. Then, the solution is derived based on this setting. The value of is calculated using the following formula: By solving this equation, we can obtain ,Right now ,in It is the natural logarithm, and after calculation, ,therefore The higher the value of this coefficient, the greater the penalty for small fluctuations in the process connection time, and the higher the required rhythmicity.

[0039] Calculations based on parameters: Substitute the parameter values ​​obtained in the preceding steps into the dynamic rhythm coupling degree calculation formula: Work unit With work unit Dynamic rhythm coupling between The calculation process is as follows: ; ; This result indicates the work unit (Paver) and work unit The dynamic rhythm coupling degree between the (road rollers) is approximately 0.5431. This value comprehensively reflects the collaborative work efficiency of the two in time and space. According to the preset evaluation criteria, for example, the coupling degree range is divided into three levels: [0.8, 1.0] is high coupling, [0.5, 0.8) is medium coupling, and [0, 0.5) is low coupling. The result of 0.5431 means that the collaboration between the two work units is at a medium level. Although there is a certain time overlap, the stability of the process connection needs to be improved.

[0040] Based on the dynamic rhythm coupling values ​​between all work unit pairs calculated in the previous step, as well as the average duration of waiting events and the standard deviation of critical operation cycle time, all calculated dynamic rhythm coupling values ​​are first organized. For example, if there is 1 paver (P1), 2 rollers (R1, R2), and 3 dump trucks (T1, T2, T3) on site, then the calculated dynamic rhythm coupling values ​​are... , , , , , First, calculate the coupling degree values ​​for all relevant work unit pairs. Then, group these coupling degree values ​​according to equipment category. For example, the coupling degree within a road roller unit is... The coupling degree between the paver and the roller unit includes and The coupling degree between the dump truck and the paver unit includes , and Then, calculate the average coupling degree within each group. For a roller group, if there is only one roller, the average value within the group is 1. If there are multiple rollers, calculate the average coupling degree between all pairs of rollers. The average coupling degree between different equipment categories is also calculated. For example, the average coupling degree between the paver and roller group is calculated as (…). + Finally, the average coupling values ​​calculated from these groups are combined with the average duration of waiting events and the standard deviation of critical operation cycle time obtained earlier to form a structured dataset, which is the construction rhythm index set. It contains multiple quantitative indicators that reflect individual efficiency, intra-group collaboration efficiency, and inter-group collaboration efficiency.

[0041] The steps to obtain the comprehensive construction compliance level are as follows: The deviation value of each component is read from the hierarchical construction deviation quantification data, and compared with the tolerance standard of the construction specification item by item to determine whether it exceeds the tolerance limit. If it exceeds the limit, it is recorded as a non-conforming component and classified into serious defects, general defects and minor defects according to the severity level of the defect. The number of non-conforming components of each level and the total number of components are counted to generate a quality judgment result that includes the number of defect classifications and the total number of components. Based on the quality assessment results, the average duration of waiting events, the average value of critical operation cycle time, the standard deviation of critical operation cycle time, and the average value of dynamic rhythm coupling degree in the construction rhythm index set are extracted and organized into a comprehensive input parameter set. Based on the comprehensive input parameter set, the comprehensive construction compliance level is calculated using the following formula: ; in, To assess the overall construction compliance level, This is the sum of the number of non-conforming components at each severity level multiplied by their risk weights. For the first The risk weight coefficient corresponding to each severity level For the first The number of non-conforming components at each severity level. The total number of severity levels of the defects. The risk weighting coefficient is the highest level of severity. This represents the total number of components. The average duration of waiting events. This is the average cycle time of the critical task, used to normalize the time parameters. The standard deviation of the critical task cycle time. This represents the average value of the dynamic rhythm coupling degree. , , These are adjustment coefficients for waiting time, cyclic fluctuation, and cooperative coupling, used to balance the influence of different process indicators.

[0042] Specifically, the deviation values ​​of each component are read from the hierarchical construction deviation quantification data, including the Euclidean distance of the geometric center, the included angle of the principal axis, and the differences in length, width, and height dimensions. A construction specification tolerance standard comparison table, compiled according to the highway engineering quality inspection and evaluation standards and project design documents, is also retrieved. This table specifies clear tolerance limits for each deviation index of each type of component (such as piers, cap beams, and roadbeds). For example, the tolerance limit for the geometric center position deviation of a pier is ±20 mm, and the tolerance limit for the roadbed width deviation is -50 mm to +20 mm. Each item is then compared. If the measured position deviation of a pier component is -25 mm, its absolute value of 25 mm exceeds the 20 mm tolerance limit, and the component is recorded as a non-conforming component. For all components recorded as non-conforming, they are further classified according to a pre-set defect severity level classification standard. This standard comprehensively... Considering the degree of deviation exceeding the limit and the structural importance of the component, the structural importance of the component is extracted from the building information model and preset by the design engineer into a 1-5 level rating. The severity level of the defect is divided as follows: when the deviation value exceeds 200% of the tolerance limit, or for a component with a structural importance rating of level 4 or 5, the deviation value exceeds 150% of the tolerance limit, it is marked as "serious defect"; when the deviation value is between 150% and 200% of the tolerance limit, or for a component with a structural importance rating of level 3, the deviation value exceeds 120% of the tolerance limit, it is marked as "general defect"; all other cases exceeding the tolerance limit are marked as "minor defect". After completing the judgment and classification of all components, the number of unqualified components classified as serious defect, general defect and minor defect is counted respectively, and the total number of components in the project is counted at the same time, generating a quality judgment result that includes the number of defective components of each level and the total number of components.

[0043] Based on the quality assessment results generated in the previous step, the following four values ​​are first extracted: the number of critically defective components, the number of generally defective components, the number of slightly defective components, and the total number of components. These values ​​are then stored in a temporary data structure. Next, the construction rhythmic index set generated in the previous step is retrieved and accessed. This index set is a structured data record containing multiple process monitoring indicators. From this set, the values ​​of three indicators—"average duration of waiting events," "average time of critical operation cycles," and "standard deviation of critical operation cycles"—are precisely extracted. Subsequently, the dynamic rhythmic coupling degree data in the construction rhythmic index set is processed. This data is a series of coupling degree values ​​stored in the form of work unit pairs, for example (paver 01, roller). The coupling degree of machine 01 is 0.75, and the coupling degree of paver 01 and roller 02 is 0.68, etc. By calculating the average value of all these independent dynamic rhythm coupling degree values, a comprehensive index that can represent the tightness of multi-equipment collaborative operation on the entire construction site is obtained, namely the average value of dynamic rhythm coupling degree. Finally, the four defect quantity-related values ​​extracted from the quality judgment results, the three time and rhythm-related index values ​​extracted from the construction rhythmic index set, and the calculated average value of dynamic rhythm coupling degree, a total of eight parameters, are integrated into a new, unified data container. Each parameter is given a clear identifier and organized into a flat key-value pair set to form a comprehensive input parameter set.

[0044] The comprehensive construction compliance level calculation formula maps a linear combination of multiple process indicators (waiting time, cyclical fluctuations, and synergistic coupling degree) to an efficiency coefficient between 0 and 1. This design can smoothly quantify the rhythm and synergy of the construction process, where waiting time and cyclical fluctuations are considered negative factors, and synergistic coupling degree is considered a positive factor, which is adjusted by a coefficient. , , To balance their respective importance.

[0045] The sum of the number of non-conforming components at each severity level multiplied by their risk weights represents a comprehensive risk measure for construction quality. To determine the total number of defect severity levels, defects are classified into three levels: critical defects, minor defects, and slight defects. , For the first The number of non-conforming components at each severity level, for example, the number of critical defects as determined by statistics. Number of general defects Number of minor defects indivual, For the first The risk weight coefficients corresponding to the severity levels are set based on the Analytic Hierarchy Process (AHP). Three senior highway engineering experts were organized to conduct pairwise comparisons and scoring of the relative importance of the three defect levels, constructing a judgment matrix. The normalized eigenvector corresponding to the largest eigenvalue was then calculated to determine the weights. The experts unanimously agreed that the importance of "severe defect" was between "strongly important" and "extremely important" (8) compared to "moderate defect," and "extremely important" compared to "minor defect" (9). "Moderate defect" was considered "moderately important" compared to "minor defect" (3). The weight vectors obtained after comprehensive calculation were rounded and proportionally adjusted to finally set the risk weight coefficient for severe defects. Risk weighting coefficient for general defects Risk weighting coefficient for minor defects .

[0046] The risk weighting coefficient for the most severe level is taken from... The maximum value in the above is used to normalize the quality risk, according to the aforementioned The highest severity level is "critical defect," with a corresponding risk weighting coefficient of 10. .

[0047] The total number of components represents the overall scale of the project and is used to calculate the defect rate. For example, the total number of components included in the highway section inspected this time is... indivual.

[0048] The average waiting time for waiting events reflects the ineffective time costs during construction. Extracted from the construction rhythm indicators generated in the preceding steps, and based on the analysis of UAV video streams, the average waiting time for all critical work units is calculated as follows: Second.

[0049] This is the average time of the critical operation cycle. This parameter serves as a benchmark time for measuring construction rhythm, and is also extracted from the set of construction rhythm indicators. The average time is calculated by identifying and statistically analyzing the duration of all critical operation cycles. Second.

[0050] The standard deviation of the critical operation cycle time quantifies the stability of the construction rhythm; a smaller standard deviation indicates a more stable rhythm. This value is extracted from a set of construction rhythm indicators and calculated to be: Second.

[0051] The average value of the dynamic rhythm coupling degree is used to comprehensively evaluate the collaborative work efficiency among different work units. It is extracted from the set of construction rhythm indicators and obtained by calculating the average value of the dynamic rhythm coupling degree of all work unit pairs. .

[0052] , , These are adjustment coefficients for waiting time, cyclical fluctuations, and collaborative coupling degree. These coefficients are used to balance the relative influence of different process indicators in the comprehensive evaluation. Their setting is based on regression analysis of historical data from similar projects, with the final project cost and schedule achievement rate as dependent variables, and these three process indicators as independent variables. A multiple linear regression model is established, and the obtained standardized regression coefficients can be used as a reference for adjustment coefficients. Fine-tuning is then performed based on expert experience. To reflect the high importance placed on the stability of construction rhythm in project management, the next priority is collaborative efficiency, and finally, the control of scattered waiting time. As the most important factor, Secondly, Minimal impact, final setting , , These coefficients indicate that, in the evaluation, the volatility of critical operation cycle time is considered the most important influencing factor, followed by the degree of synergy and coupling, and finally the waiting time.

[0053] Calculations based on parameters: First, calculate the cumulative result of the weighted number of defective components: ; Calculate the quality score items: ; The index portion of the efficiency score in the calculation process: ; ; Efficiency score items in the calculation process: ; Calculate the final overall construction compliance level: ; The results indicate that the overall construction compliance score for this highway construction period is 0.7014. This score is a dimensionless value between 0 and 1, which can be interpreted according to the preset grading standards. For example, [0.9, 1.0] is excellent, [0.75, 0.9) is good, [0.6, 0.75) is qualified, and below 0.6 is unqualified. Therefore, a score of 0.7014 indicates that the overall construction compliance is at the "qualified" level. The analysis and calculation process shows that the quality score (0.9931) is very high, indicating that there are few defects and low risks in the quality of the physical project. However, the process efficiency score (0.7062) is relatively low, which pulls the total score down to the qualified range. This indicates that although the final output quality is controllable, the organizational efficiency of the construction process, especially the stability of the rhythm and the coordination between equipment, is the main area of ​​weakness, and there is considerable room for improvement.

[0054] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for AI image recognition grading of unmanned aerial vehicles oriented to highway construction period, characterized in that, The method comprises the following steps: extracting highway real scene three-dimensional point cloud from unmanned aerial vehicle image, calling project design building information model, calculating multi-modal model set under aligned coordinate system, segmenting real scene three-dimensional point cloud in the multi-modal model set under the aligned coordinate system, and obtaining semantic real scene component set; comparing the semantic real scene component set and the component information in the building information model one by one to determine whether the components exist, screening missing or excessive components in the real scene, establishing a component existence deviation list, calculating the Euclidean distance of the geometric center, the included angle of the main axis direction and the length-width-height size difference for the components commonly existing in the model according to the component existence deviation list, and generating hierarchical construction deviation quantitative data; detecting and continuously tracking key operation units from continuous highway construction period unmanned aerial vehicle image video stream, the key operation units including asphalt paver, road roller and earthmoving truck, forming a time sequence operation element event sequence, constructing nodes and directed edges according to the sequence and logical association of each event in the time sequence operation element event sequence, and establishing a dynamic construction workflow atlas; quantifying the coordination work tightness of the associated operation units according to the dynamic construction workflow atlas to obtain a construction rhythm index set, comparing the hierarchical construction deviation quantitative data with the preset construction specification tolerance standard to determine, and integrating the construction rhythm index set and the determination result to operate and generate a comprehensive construction compliance level.

2. The expressway construction period-oriented drone AI image recognition grading method of claim 1, wherein The obtaining step of the semantic real scene component set is: extracting feature points from unmanned aerial vehicle image and back projecting to generate real scene three-dimensional point cloud, calling project design building information model to load topology, extracting real scene three-dimensional point cloud bridge pier center point and roadbed boundary line, obtaining building information model structure key points, unifying coordinate naming rules and field formats, and forming aligned input feature set; according to the aligned input feature set, establishing one-to-one mapping of bridge pier center point and structure key point, calculating coordinate transformation matrix composed of rotation term and displacement term and applying to real scene three-dimensional point cloud and building information model, removing abnormal matching according to key point residual error threshold and calculating coordinate transformation matrix again, solidifying the matrix when the residual error change is lower than the preset ratio, and forming multi-modal model set under aligned coordinate system; according to the multi-modal model set under the aligned coordinate system, segmenting real scene three-dimensional point cloud in the aligned coordinate system by normal distribution, plane continuity and boundary curvature, assigning categories in combination with the geometric range and connection topology of the walls, beams, plates and columns of the building information model, removing isolated fragments without assigned categories and merging boundary overlapping areas, outputting the component category label and the component number of each point, and generating semantic real scene component set. 3.The expressway construction period-oriented drone AI image recognition grading method of claim 1, wherein The obtaining step of the component existence deviation list is: according to the semantic real scene component set and the component information in the building information model, checking the existence state according to the component number and the component type, marking the components not appearing in the semantic real scene component set as missing, marking the components not appearing in the component information in the building information model as excessive, and generating a component existence deviation list. 4.The expressway construction period-oriented drone AI image recognition grading method of claim 1, wherein The obtaining step of the hierarchical construction deviation quantitative data is: According to the component existence deviation list, components existing in the semanticized real scene component set and the building information model at the same time are screened, geometric center coordinates, main axis direction, length, width and height dimensions are extracted, Euclidean distance of geometric centers, included angle of main axis directions and length, width and height dimension differences are calculated pair by pair, and Euclidean distance of geometric centers, included angle of main axis directions and length, width and height dimension difference details are formed; According to the Euclidean distance of geometric centers, the included angle of the main axis direction and the length, width and height dimension difference details, the connection topology of the component information in the building information model is called, whether the connection relationship of the adjacent components conforms to the design is checked, the difference value is counted and aggregated according to the component level, and hierarchical construction deviation quantitative data is generated. 5.The expressway construction period-oriented drone AI image recognition grading method of claim 1, wherein The acquisition step of the time sequence operation element event sequence is: The frame sequence is decoded from the continuous highway construction period unmanned aerial vehicle image video stream, the time stamp of each frame is read and numbered, the vehicle candidate area is located, and each item is compared according to the rectangular length-width ratio, the number of tracks or tires, the opening angle of the hopper, and the edge profile of the screed, confirming that it is an asphalt paver, a road roller and a dump truck, connecting the trajectories of the same target according to the center point distance and the contour similarity of adjacent frames, counting the speed, hopper opening and closing, vibration mark and unloading door state according to the continuous time window, marking the operation state, and generating the time sequence operation element event sequence. 6.The expressway construction period-oriented drone AI image recognition grading method of claim 1, wherein The acquisition step of the dynamic construction workflow graph is: According to the time sequence operation element event sequence, a node is created for each event, the device identifier, operation state and time stamp are recorded, the device identifier consistency and process adjacency of adjacent events are checked according to the event sequence, the directed edge is established if the condition is met, the time stamp difference of adjacent events is calculated and written into the edge weight, and the pairing that does not meet the condition is discarded, forming a node and a directed edge set; According to the node and directed edge set, the nodes are arranged in layers according to the device category, the nodes of the same device are topologically sorted according to the time stamp, repeated directed edges are merged and the edge weight is accumulated, directed edges that cross the process order are deleted, continuous directed edge paths according to the time stamp are retained, the start index and end index of the path are registered, and the dynamic construction workflow graph is formed. 7.The expressway construction period-oriented drone AI image recognition grading method of claim 1, wherein The acquisition step of the construction rhythm index set is: From the event records of each operation unit in the dynamic construction workflow graph, the data segment of the waiting type event is screened, the start and end time is extracted and the duration is calculated, the duration mean value of all waiting type events is counted, the average length of the waiting type event is obtained, and the start and end points of the key operation cycle are identified on the same time axis, the duration of each cycle is calculated, the standard deviation is calculated, and the average length of the waiting type event and the standard deviation of the key operation cycle time are generated; According to the average length of the waiting type event and the standard deviation of the key operation cycle time, the dynamic rhythm coupling degree between operation units is calculated; According to the dynamic rhythm coupling degree, the average length of the waiting type event and the standard deviation of the key operation cycle time, the dynamic rhythm coupling degree of each operation unit is compared pair by pair, the average value in the group is calculated according to the device category, and the average value of each group is summarized with the average length of the waiting type event and the standard deviation of the key operation cycle time, forming the construction rhythm index set. 8.The expressway construction period-oriented drone AI image recognition grading method of claim 1, wherein The obtaining step of the comprehensive construction compliance level is: Read the deviation value of each component from the hierarchical construction deviation quantification data, compare it with the construction specification tolerance standard item by item, judge whether it exceeds the tolerance limit value, if it exceeds, record it as unqualified component and divide it into serious defect, general defect and slight defect according to the defect severity level, count the number of unqualified components of each level and the total number of components, and generate the quality judgment result containing the defect classification number and the total number of components; According to the quality judgment result, extract the average length of the waiting type event, the average value of the key operation cycle time, the standard deviation of the key operation cycle time and the average value of the dynamic rhythm coupling degree in the construction rhythm index set, and arrange them into a comprehensive input parameter set; According to the comprehensive input parameter set, calculate the comprehensive construction compliance level.

Images