Search

CN-121999439-A - Efficient three-dimensional target detection method for point cloud data of intelligent construction scene

CN121999439ACN 121999439 ACN121999439 ACN 121999439ACN-121999439-A

Abstract

The invention discloses a point cloud data high-efficiency three-dimensional target detection method for an intelligent construction scene, which belongs to the field of engineering management and comprises the specific steps of periodically collecting and preprocessing point cloud data of a construction site, establishing a continuous time frame sequence, simultaneously constructing a construction twin model corresponding to the construction site, II, aligning the point cloud data with the semantic topology of the construction twin model, identifying abnormal areas of the construction site, and analyzing evolution tracks of objects of the construction site in real time.

Inventors

  • XU YE
  • SUN ZHENG

Assignees

  • 孙正

Dates

Publication Date
20260508
Application Date
20260210

Claims (7)

  1. 1. The efficient three-dimensional target detection method for the point cloud data of the intelligent construction scene is characterized by comprising the following specific steps of: I. periodically collecting and preprocessing construction site point cloud data, establishing a continuous time frame sequence, and simultaneously constructing a construction twin model corresponding to a construction site; II. Aligning the point cloud data with the semantic topology of the construction twin model, identifying the abnormal region of the construction site, and analyzing the evolution track of each object of the construction site in real time; III, dividing the multi-scale subgraph of the construction site based on an evolution track analysis result, and reconstructing the shielded object in the point cloud data according to the divided subgraph; IV, generating a three-dimensional target candidate frame based on the reconstructed high-integrity point cloud data, and carrying out preliminary prediction on each item of information of the target; And V, constructing a construction knowledge graph, verifying and correcting a preliminary prediction result through the construction knowledge graph, and dynamically adjusting a detection threshold and a target priority according to the current construction stage.
  2. 2. The method for efficient three-dimensional object detection of point cloud data for intelligent construction scenes according to claim 1, wherein the specific steps of periodically collecting and preprocessing the point cloud data of the construction site in the step I are as follows: S1.1, synchronously scanning a scene by utilizing a plurality of laser radars with different spatial positions and complementary visual angles on a construction site, respectively generating one frame of original point cloud data in the same acquisition period by each laser radar, and recording corresponding acquisition time stamps for each frame of point cloud data; S1.2, mapping point cloud data acquired by each laser radar into a public world coordinate system of a unified construction scene through rigidity transformation, and respectively calculating neighborhood average distances between each point cloud data with unified coordinates and adjacent point cloud data in a preset range; And S1.3, if the average distance of the existing field is higher than a preset threshold value, judging the corresponding point cloud data as outlier noise points, removing each detected outlier noise point, performing voxel downsampling on the denoised point cloud data, fitting a ground model, calculating the distance between the point cloud data and the ground, and separating the ground points in the point cloud data.
  3. 3. The method for efficient three-dimensional target detection of point cloud data for intelligent construction scenes according to claim 2, wherein the specific steps of constructing a construction twinning model corresponding to a construction site in step I are as follows: S2.1, analyzing a BIM file of a construction site into parameterized representation of component level, extracting each item of basic information of a component, meanwhile, disassembling complex curved surfaces of each component into a plurality of groups of parameter units, establishing semantic association among the components to generate a construction twin model comprising an indexable data table and a quick query interface, and mapping a construction plan into a normalized progress scalar and a stage interval; S2.2, extracting visible intervals of each component according to the mapped construction plan, simultaneously establishing a global progress scalar for searching states of each component at any time point based on the completed and non-started states in construction, extracting a plurality of groups of semantic anchor points from parameterized BIM geometry, obtaining corresponding observation anchor points on site through manual or automatic calibration, and obtaining rigid transformation between a construction twin model and site actual measurement coordinates under a world coordinate system by adopting rigid registration based on each semantic anchor point and the observation anchor points; S2.3, coding expected visibility, position and state of each component into a probability format to obtain position priori distribution and shape priori probability of each component, and recording confidence coefficient of each priori, when each laser radar obtains field observation, carrying out linear-Gaussian approximate update on each priori in a construction twin model based on measurement uncertainty to obtain corresponding posterior distribution corrected by observation, and recording residual error and confidence coefficient change; S2.4, a group of disturbance pattern generators are established in the construction twin model, when residual errors of the posterior distribution after observation and correction are higher than a preset threshold value, a plurality of groups of disturbance samples are generated through the disturbance pattern generators, each group of disturbance samples are generated to conduct Monte Carlo sampling on each prior, prior observation distribution under different disturbance samples is estimated, and real-time adjustment is conducted on each posterior distribution based on the updated prior observation distribution.
  4. 4. The method for efficient three-dimensional object detection of point cloud data for intelligent construction scenes according to claim 3, wherein the specific steps of aligning the point cloud data with the semantic topology of the construction twinning model in the step II and identifying the abnormal region of the construction site are as follows: S3.1, extracting a plurality of groups of key points from real-time point cloud data by a spatial clustering method to serve as graph nodes, taking geometrical adjacency relations among points as edges, establishing a point cloud graph, decomposing each component in a construction twin model into corresponding semantic units, combining the corresponding semantic units into another group of graph nodes, simultaneously taking semantic dependency relations of graph node parts as edges, establishing a corresponding twin graph, and then respectively generating adjacency matrixes corresponding to the point cloud graph and the twin graph; s3.2, respectively constructing corresponding original feature vectors for each node of the point cloud image and the twin image, respectively inputting the original feature vectors of each group of nodes into two groups of independent encoder networks, mapping each node feature vector into a corresponding embedded vector in an embedded vector space, respectively running multi-layer message transmission on each point cloud image and the twin image, aggregating local neighborhood information into node embedding, and establishing corresponding context information; S3.3, mapping the context of each node on the twin graph to a point cloud node space through linear projection, calculating the similarity between each point cloud node and all the twin nodes to establish a soft pairing similarity matrix, and distributing corresponding attention weights for projection embedding of each node in the twin graph based on the similarity matrix; And S3.4, carrying out weighted summation on projection embedding of each node of the twin map based on each distributed attention weight to generate expected embedding of a corresponding node in the point cloud map under the twin priori, then calculating the difference degree between the expected embedding and the actual embedding of the point cloud, converting each calculated group of difference degrees into a corresponding attention mask, then calculating the normalized difference score of each region of the current construction site by combining each attention mask based on uncertainty estimation of the twin priori, and marking the construction region as an abnormal region and feeding back constructors to carry out verification if the difference score of the construction region is higher than a preset threshold value.
  5. 5. The method for efficient three-dimensional object detection of point cloud data for intelligent construction scenes according to claim 4, wherein the specific steps of analyzing the evolution track of each object in the construction site in real time in step II are as follows: S4.1, sequentially slicing continuously acquired multi-frame point cloud data according to time to form a time window with preset length, carrying out space voxelization on the point cloud data in each time window, converting each frame of point cloud data into regularized voxel grids, stacking each voxel grid along a time dimension to form standardized space-time points cloud mass, storing additional channels of each voxel grid, and then taking the space-time points cloud mass as input, and adopting three-dimensional time sequence convolution with multiple scales to parallelly extract evolution perception features on local space-time neighbors in a branching mode; S4.2, carrying out batch normalization and activation on extracted evolution perception features to generate corresponding space-time feature map samples, carrying out low-pass and high-pass separation on each space-time feature map sample on a time axis, taking stable and constant components in the time direction after separation as static geometric features, taking difference components as dynamic evolution features, respectively normalizing and independently reducing dimensions of the two types of features, carrying out multi-time aggregation on each dynamic evolution feature, and obtaining evolution track codes of each spatial position of a construction site; S4.3, calculating point cloud optical flow of each voxel grid between adjacent time frames, establishing an optical flow field corresponding to a construction site, carrying out reverse or forward reprojection on static geometric features and dynamic evolution features of each time frame according to the optical flow field to a coordinate system of the next time so as to obtain the features after flow alignment, then calculating consistency of the reprojection features and actual features of a target frame, and adjusting track coding and attention weight based on calculation results.
  6. 6. The method for efficient three-dimensional object detection of point cloud data for intelligent construction scenes according to claim 5, wherein the specific steps of performing multi-scale sub-division on a construction site based on the analysis result of the evolution track in step III are as follows: S5.1, carrying out linear weighting and normalization processing on the multi-channel characteristics obtained by each voxel grid in the construction stage of the space-time points cloud mass to generate final node representation of the corresponding voxel grid, simultaneously recording the three-dimensional position and the observation view angle set of the voxel grid, then calculating the space distance and the feature similarity of each pair of node representations in the three-dimensional position space, and generating the corresponding candidate edge weight based on the calculated space distance and the feature similarity; s5.2, calculating a corresponding space visibility score for each node representation according to the laser radar pose set and the visibility record of the node and combining with sight shielding judgment, projecting a local neighborhood represented by each node under each laser radar view angle into an imaging plane of the view angle, evaluating whether the projection covers an expected geometric area or not, and if not, re-adjusting projection parameters and re-projecting; S5.3, respectively carrying out weighted average on the local projection coverage rate of each node under all view angles to obtain view angle integrity scores of corresponding nodes, and dividing the node representation into a global sub-graph node set when the spatial visibility score and the integrity score of the node representation are higher than a preset threshold value, otherwise, the node representation is a local sub-graph candidate set; And S5.4, reserving edges with candidate edge weights higher than a preset threshold value between node representations in the global sub-graph and the local sub-graph, adding long-distance context edges into the global sub-graph, collecting additional edges with common occurrence of visual angles, adding the additional edges to the local sub-graph, respectively calculating descriptors corresponding to the global sub-graph and the local sub-graph, carrying out connectivity inspection on each global sub-graph and the local sub-graph, adding bridging edges to adjacent sub-graphs based on inspection results for merging, and then calculating statistics of each sub-graph, and recording.
  7. 7. The method for efficient three-dimensional object detection of point cloud data for intelligent construction of a scene according to claim 6, wherein the specific steps of preliminary predicting each item of information of the object in step IV are as follows: S6.1, aggregating the reconstructed point cloud data according to a space voxel grid to generate a group of candidate seed points, collecting characteristics of neighbor points of each candidate seed point corresponding to the scale based on preset different scales, carrying out statistical pooling processing on each scale to generate corresponding scale vectors, splicing the multi-scale vectors, and acquiring the characteristic vectors of the candidate seed points through a group of MLPs; s6.2, calculating the displacement pointed to the potential target center based on the current candidate seed point position through feature learning based on the feature vector of each candidate seed point to acquire one or more voting vectors, clustering votes of all candidate seed points to acquire a plurality of groups of candidate centers, acquiring seed features belonging to the clusters based on each candidate center, and carrying out pooling treatment; S6.3, respectively predicting target size vectors and an initial orientation representation of each candidate center by using regression branches, simultaneously outputting corresponding category confidence coefficients, carrying out parameterization on the positions of the candidate centers, the regression sizes and the orientations to obtain three-dimensional candidate boundary frame representations, predicting aggregate feature categories by using another path of classified branches, acquiring category probability distribution by using softma functions, and calculating comprehensive confidence scores of each candidate boundary frame; And S6.4, calculating the coverage rate of point cloud data in each candidate boundary frame and the average distance between points and a fitting surface, calculating final frame confidence score based on each geometric index, sorting each candidate boundary frame in a descending order according to the final frame confidence score, calculating NMS overlapping degree of each candidate boundary frame based on the fine-granularity three-dimensional box body intersection ratio, sequentially reserving and eliminating other candidate boundary frames with the NMS overlapping degree exceeding a threshold value of the reserved candidate boundary frames, and outputting frame parameters, prediction categories, confidence scores, uncertainty vectors and index summaries of each candidate boundary frame.

Description

Efficient three-dimensional target detection method for point cloud data of intelligent construction scene Technical Field The invention relates to the field of engineering management, in particular to a point cloud data efficient three-dimensional target detection method for intelligent construction scenes. Background Along with the continuous improvement of the level of new infrastructure construction and building industrialization, the intelligent demands of the construction site on safety supervision, progress management and quality control are increasingly increased. The three-dimensional laser scanning and point cloud sensing technology has become an important environmental sensing means in intelligent construction scenes due to the advantages of high precision, non-contact, all weather and the like. By carrying out three-dimensional target detection on the construction site point cloud data, automatic identification and state evaluation of components, equipment, personnel and temporary facilities can be realized, and basic support is provided for digital management of the construction process. However, the construction scene has complex characteristics significantly different from the autopilot or indoor scene. On one hand, construction objects continuously evolve along with engineering progress, the form, position and semantic state of a component show obvious stepwise changes, and on the other hand, a large number of shielding, stacking, noise interference and unstructured environments exist on site, so that the point cloud data is strong in sparsity and insufficient in integrity. In addition, the traditional three-dimensional target detection method generally assumes that a scene structure is relatively stable, lacks modeling capability for construction flow and engineering knowledge, and is difficult to fully utilize construction priori information, so that detection accuracy and robustness are limited in a complex dynamic environment. In order to solve the problems, it is important to develop a point cloud data efficient three-dimensional target detection method for intelligent construction scenes. The existing point cloud data efficient three-dimensional target detection method is high in occurrence probability of false detection and unreasonable engineering semantic results, poor in detection stability of slowly moving or stepwise changing targets, incapable of avoiding the problems of jump and omission easily caused by single-frame detection, capable of reducing the identification accuracy of shielded components in a complex construction environment and incapable of improving the overall detection efficiency while guaranteeing the accuracy, and therefore, the point cloud data efficient three-dimensional target detection method for intelligent construction scenes is provided. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides a point cloud data efficient three-dimensional target detection method for an intelligent construction scene. In order to achieve the above purpose, the present invention adopts the following technical scheme: A point cloud data efficient three-dimensional target detection method for an intelligent construction scene comprises the following specific steps: I. periodically collecting and preprocessing construction site point cloud data, establishing a continuous time frame sequence, and simultaneously constructing a construction twin model corresponding to a construction site; II. Aligning the point cloud data with the semantic topology of the construction twin model, identifying the abnormal region of the construction site, and analyzing the evolution track of each object of the construction site in real time; III, dividing the multi-scale subgraph of the construction site based on an evolution track analysis result, and reconstructing the shielded object in the point cloud data according to the divided subgraph; IV, generating a three-dimensional target candidate frame based on the reconstructed high-integrity point cloud data, and carrying out preliminary prediction on each item of information of the target; And V, constructing a construction knowledge graph, verifying and correcting a preliminary prediction result through the construction knowledge graph, and dynamically adjusting a detection threshold and a target priority according to the current construction stage. As a further scheme of the invention, the specific steps of periodically collecting and preprocessing the construction site point cloud data in the step I are as follows: S1.1, synchronously scanning a scene by utilizing a plurality of laser radars with different spatial positions and complementary visual angles on a construction site, respectively generating one frame of original point cloud data in the same acquisition period by each laser radar, and recording corresponding acquisition time stamps for each frame of point cloud data; S1.2, mapping point cloud d