Image Recognition-Based Construction Monitoring Method and System for Prefabricated Green Buildings

Abstract

This invention discloses a method and system for monitoring the construction of prefabricated green buildings based on image recognition. The system acquires standard image data through sensors; a lightweight multi-task neural network (LMTN) is constructed to analyze the standard image data; an improved Hessian matrix ridge detection algorithm and a U-Net segmentation model are combined to identify concrete cracks and exposed rebar defects, yielding image recognition results; stereoscopic images of the construction site are acquired, and a disparity map is generated using an improved SGM algorithm, which is then converted into 3D point cloud data; the volume and distribution density of building materials are calculated based on the 3D point cloud data, and the construction load is estimated by combining material property data to generate a load distribution map; a construction risk index is calculated based on the image recognition results and the load distribution map, and a graded early warning mechanism is triggered when the risk index exceeds a preset threshold. This effectively reduces safety risks and improves monitoring efficiency.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method and system for monitoring the construction of prefabricated green buildings based on image recognition. Background Technology

[0002] With the large-scale development of prefabricated green buildings, the demand for safety management and quality supervision during construction is becoming increasingly urgent. Traditional monitoring relies on manual inspections, which suffers from low efficiency, numerous blind spots, and delayed risk response, making it difficult to adapt to the complex scenarios of component hoisting and multiple overlapping processes in prefabricated construction. Although existing image recognition technology is being gradually applied, image acquisition is easily affected by construction dust and sudden changes in lighting, resulting in poor data quality; single-task recognition models require parallel deployment of multiple systems, consuming large resources and lacking coordination; load calculations rely heavily on empirical formulas, and the accuracy and predictive power of combining them with 3D scenes are insufficient. All of these factors contribute to the low efficiency of building construction monitoring. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing a method and system for monitoring the construction of prefabricated green buildings based on image recognition.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned image recognition-based prefabricated green building construction monitoring method:

[0005] Time-series data, product image data, and material property data are collected through a sensor network around the mold. The collected data is preprocessed to obtain multimodal data.

[0006] Multi-view images and video images of the prefabricated green building construction area are acquired by sensors. The improved DHE dynamic histogram equalization algorithm is used to enhance the images. After eliminating uneven lighting and dust interference, invalid frames are removed to obtain standard image data.

[0007] A lightweight multi-task neural network (LMTN) is constructed to analyze the standard image data. The YOLOv8 model is used to detect the wearing status of safety helmets and reflective clothing of construction workers and their violations. The improved Hessian matrix ridge detection algorithm and U-Net segmentation model are combined to identify concrete cracks and exposed steel bars, and the image recognition results are obtained.

[0008] The system collects stereoscopic images of the construction site, generates disparity maps using an improved SGM algorithm, and converts them into 3D point cloud data. Based on the 3D point cloud data, it calculates the volume and distribution density of building materials, estimates the construction load by combining material property data, and generates a load distribution map.

[0009] The construction risk index is calculated based on the image recognition results and load distribution map. When the risk index exceeds a preset threshold, a graded early warning mechanism is triggered.

[0010] Furthermore, in the above-mentioned image recognition-based prefabricated green building construction monitoring method, the step of acquiring multi-view images and video images of the prefabricated green building construction area through sensors, enhancing the images using an improved DHE dynamic histogram equalization algorithm, eliminating uneven lighting and dust interference, and then discarding invalid frames to obtain standard image data, including:

[0011] Multi-view images and video images of the prefabricated green building construction area are acquired through sensors;

[0012] Based on the traditional DHE dynamic histogram equalization, an adaptive block division strategy is introduced. The image is divided into 16 sub-blocks according to the average gray value. After Gaussian filtering preprocessing, histogram equalization is performed to obtain the improved DHE dynamic histogram equalization algorithm.

[0013] The color image is converted to YUV format using an improved DHE dynamic histogram equalization algorithm. The improved DHE processing is performed only on the luminance channel, while the chrominance channel remains unchanged to avoid color distortion. After contrast stretching optimization, it is converted to RGB format to obtain an enhanced image.

[0014] Furthermore, in the above-mentioned image recognition-based prefabricated green building construction monitoring method, the step of acquiring multi-view images and video images of the prefabricated green building construction area through sensors, enhancing the images using an improved DHE dynamic histogram equalization algorithm, eliminating uneven lighting and dust interference, and then discarding invalid frames to obtain standard image data, including:

[0015] An adaptive brightness correction algorithm is used to calculate the average brightness of each sub-block of the enhanced image. Brightness compensation is performed on areas that are 60% lower than the global average. Fine dust noise is removed by morphological opening operation. Then, an adaptive threshold segmentation algorithm is used to separate dust-occluded areas and perform local repair to obtain the corrected image.

[0016] Invalid frames are removed from the corrected image. The image after removing invalid frames is then standardized and the pixel size is adjusted to generate standard image data.

[0017] Furthermore, in the above-mentioned image recognition-based prefabricated green building construction monitoring method, the construction of an LMTN lightweight multi-task neural network to analyze the standard image data, and the detection of construction workers' safety helmet and reflective clothing wearing status and violations through the YOLOv8 model, includes:

[0018] The standard image is input into the improved YOLOv8 model, and candidate boxes are generated through anchor box clustering. Duplicate boxes are filtered out by nonmaximum suppression, and the target detection results are output, which include at least the personnel location coordinates, the wearing status of safety helmets and reflective clothing, the category of violation, and the confidence level.

[0019] Furthermore, in the above-mentioned image recognition-based prefabricated green building construction monitoring method, the method of combining the improved Hessian matrix ridge detection algorithm with the U-Net segmentation model to identify concrete cracks and exposed rebar defects, and obtaining image recognition results, includes:

[0020] An improved Hessian matrix ridge detection algorithm is used to perform Gaussian blur noise reduction on the image, calculate the Hessian matrix and extract feature values; the improvement lies in introducing an adaptive threshold, repairing fracture cracks through a ridge connection algorithm, and finally outputting the length, width and location information of the cracks.

[0021] The coordinates of the crack and exposed rebar identification results are calibrated, duplicate labeled areas are removed, and a total image recognition report containing defect type, location, size, and severity is generated.

[0022] Furthermore, in the above-mentioned image recognition-based prefabricated green building construction monitoring method, the acquisition of stereoscopic images of the construction site, the generation of a disparity map using an improved SGM algorithm, and the conversion into three-dimensional point cloud data; the calculation of building material volume and distribution density based on the three-dimensional point cloud data, the estimation of construction load by combining material property data, and the generation of a load distribution map, including:

[0023] The system acquires stereoscopic images of the construction site and calculates the initial cost by combining Census transform with grayscale difference based on the traditional SGM algorithm. In the aggregation step, a dynamic aggregation window is introduced to reduce disparity error in edge areas, and a disparity map is obtained through disparity refinement.

[0024] Based on the intrinsic and extrinsic parameters of the binocular camera, the disparity map is converted into three-dimensional point cloud data using the principle of triangulation.

[0025] Furthermore, in the above-mentioned image recognition-based prefabricated green building construction monitoring method, the acquisition of stereoscopic images of the construction site, the generation of a disparity map using an improved SGM algorithm, and the conversion into three-dimensional point cloud data; the calculation of building material volume and distribution density based on the three-dimensional point cloud data, the estimation of construction load by combining material property data, and the generation of a load distribution map, including:

[0026] A point cloud segmentation algorithm based on region growing is adopted. According to the surface roughness and color characteristics of building materials, the point cloud is segmented into different types of building materials. The minimum bounding volume is calculated by the convex hull algorithm for the segmented point cloud. Combined with point cloud density correction, the actual volume is obtained.

[0027] The construction area is divided into grid cells. The number of building material point clouds in each grid cell is counted. The building material distribution density of each grid cell is calculated by combining the grid area and the point cloud single point area; a density heat map is generated.

[0028] The density parameters of various building materials are collected, and the weight of a single component is calculated by combining the volume. The total construction load of each grid is calculated by superposition method, and a load distribution map is generated.

[0029] Furthermore, in the image recognition-based prefabricated green building construction monitoring system, the prefabricated green building construction monitoring system includes the following modules:

[0030] The image data acquisition module is used to acquire multi-view images and video images of the construction area of ​​prefabricated green buildings through sensors. The improved DHE dynamic histogram equalization algorithm is used to enhance the images, eliminate uneven lighting and dust interference, and remove invalid frames to obtain standard image data.

[0031] The construction image recognition module is used to construct a lightweight multi-task neural network (LMTN) to analyze the standard image data. It uses the YOLOv8 model to detect the wearing status of safety helmets and reflective clothing of construction workers and their violations. It combines the improved Hessian matrix ridge detection algorithm with the U-Net segmentation model to identify concrete cracks and exposed steel bars, and obtains the image recognition results.

[0032] The load distribution calculation module is used to acquire stereoscopic images of the construction site, generate disparity maps through an improved SGM algorithm, and convert them into three-dimensional point cloud data; calculate the volume and distribution density of building materials based on the three-dimensional point cloud data, estimate the construction load by combining material property data, and generate a load distribution map.

[0033] The construction graded early warning module is used to calculate the construction risk index based on the image recognition results and load distribution map. When the risk index exceeds a preset threshold, the graded early warning mechanism is triggered.

[0034] Furthermore, in the image recognition-based prefabricated green building construction monitoring system, the image data acquisition module includes the following sub-modules:

[0035] The calculation submodule is used to calculate the average brightness of each sub-block of the enhanced image using an adaptive brightness correction algorithm, perform brightness compensation on areas that are 60% lower than the global average, remove fine dust noise through morphological opening operation, and then use an adaptive threshold segmentation algorithm to separate dust-occluded areas and perform local repair to obtain the corrected image.

[0036] The removal submodule is used to remove invalid frames from the corrected image, standardize the image after removing invalid frames, adjust the pixel size, and generate standard image data.

[0037] Furthermore, in the image recognition-based prefabricated green building construction monitoring system, the construction image recognition module includes the following sub-modules:

[0038] The output submodule is used to input standard images into the improved YOLOv8 model, generate candidate boxes through anchor box clustering, filter duplicate boxes through non-maximum suppression, and output target detection results, which include at least the personnel location coordinates, the wearing status of safety helmets and reflective clothing, the category of violation, and the confidence level.

[0039] Its beneficial effects lie in achieving a synergistic improvement in monitoring accuracy, efficiency, and green goals through end-to-end technological innovation. The improved DHE algorithm, combined with targeted interference processing, enhances image sharpness, providing high-quality data for subsequent identification and solving the data acquisition challenges in construction environments. The lightweight LMTN network integrates dual-task branches, reducing the number of parameters, increasing inference speed, and improving the accuracy of identification features such as safety helmet wearing. Crack detection error is ≤0.2mm, enabling multi-dimensional monitoring within a single system and reducing equipment deployment costs. The improved SGM algorithm generates a 3D point cloud density of 100 points / ㎡, with a building material volume calculation error of ≤3%, and accurately marks areas exceeding standards on the load distribution map, providing data support for green construction and improving building material utilization. The tiered early warning mechanism, combined with dynamic thresholds and specific trigger conditions, shortens the early warning response time to the second level, effectively reducing safety risks and improving monitoring efficiency. Attached Figure Description

[0040] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0041] Figure 1 This is a schematic diagram of the first embodiment of the image recognition-based prefabricated green building construction monitoring method in this invention.

[0042] Figure 2 This is a schematic diagram of the second embodiment of the image recognition-based prefabricated green building construction monitoring method in this invention.

[0043] Figure 3 This is a schematic diagram of the first embodiment of the prefabricated green building construction monitoring system based on image recognition in this invention. Detailed Implementation

[0044] 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.

[0045] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0046] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, the image recognition-based construction monitoring method for prefabricated green buildings includes the following steps:

[0047] Step 101: Acquire multi-view images and video images of the prefabricated green building construction area through sensors, enhance the images using the improved DHE dynamic histogram equalization algorithm, eliminate uneven lighting and dust interference, and remove invalid frames to obtain standard image data.

[0048] Specifically, in this embodiment, multi-view images and video images of the prefabricated green building construction area are acquired through sensors. Based on traditional DHE dynamic histogram equalization, an adaptive block-segmentation strategy is introduced, dividing the image into 16 sub-blocks according to the average grayscale value. After Gaussian filtering preprocessing, histogram equalization is performed, resulting in an improved DHE dynamic histogram equalization algorithm. The improved DHE dynamic histogram equalization algorithm is used to convert the color image to YUV format, performing improved DHE processing only on the luminance channel while keeping the chroma channel unchanged to avoid color distortion. After contrast stretching optimization, it is converted to RGB format to obtain an enhanced image. An adaptive luminance correction algorithm is used to calculate the average luminance of each sub-block in the enhanced image. Luminance compensation is performed on areas below 60% of the global average. Fine dust noise is removed using morphological opening operations. Then, an adaptive threshold segmentation algorithm is used to separate dust-occluded areas and perform local repair to obtain a corrected image. Invalid frames in the corrected image are removed. The image after removing invalid frames is standardized, and the pixel size is adjusted to generate standard image data.

[0049] Sensor selection is based on the scale of the construction area. When a single area is ≥500㎡, additional equipment and scene requirements are needed. A 4K high-definition industrial camera with a frame rate of 25-30fps and an adjustable lens focal length of 8-24mm is selected as the main sensor, paired with a low-light infrared camera with automatic switching to night mode and a light sensitivity of ≤0.001lux to achieve all-day and all-night data acquisition. For key areas such as high-altitude operation areas and component hoisting areas, fisheye cameras are additionally deployed with a 180° field of view to cover blind spots.

[0050] The camera placement adopts a three-dimensional cross-location + key area densification principle. One main camera is placed on the top of the tower cranes around the construction area, at least 20m above the ground, at the top of the scaffolding poles, every 30m, and at the entrance of the construction passage, to ensure that the field of view overlap rate of adjacent equipment is ≥30%. Infrared cameras are installed at the same locations as the main cameras, and fisheye cameras focus on the hoisting radius within 15m and dangerous areas such as stairwells and edge openings.

[0051] Data transmission and storage are achieved via industrial Ethernet with a bandwidth of ≥100Mbps. Real-time data is transmitted to local edge computing nodes, using an edge caching + cloud backup mode. Edge nodes store nearly 72 hours of raw data with a hard drive capacity of ≥10TB. Effective data is automatically compressed at midnight every day with a compression ratio of 10:1 before being uploaded to the cloud server.

[0052] The key improvements to the algorithm are as follows: Based on the traditional DHE dynamic histogram equalization, an adaptive block division strategy is introduced. The image is divided into 4×4=16 sub-blocks according to the average gray value. Areas with uniform illumination are merged into 8×8 sub-blocks, and areas with abrupt changes in illumination are split into 2×2 sub-blocks. A noise suppression module is added, which uses Gaussian filtering with a standard deviation σ=0.8-1.2. The preprocessing is dynamically adjusted according to the dust concentration before histogram equalization is performed.

[0053] The enhanced operation process first converts the color image to YUV format, performing improved DHE processing only on the luminance and Y channels, while keeping the chroma, U, and V channels unchanged to avoid color distortion. After processing, it is further optimized by contrast stretching, with a stretching range of 1.2-1.5 times, and finally converted back to RGB format to obtain the enhanced image.

[0054] For targeted interference handling, an adaptive brightness correction algorithm is used to address uneven illumination. After calculating the average brightness of each sub-block in the image, brightness compensation is applied to areas that are 60% lower than the global average, with a compensation coefficient of 0.8-1.2. For dust interference, fine dust noise is removed by morphological opening operations and a 3×3 rectangular operator as the structuring element. Then, adaptive threshold segmentation is used, with the threshold dynamically calculated by the Otsu algorithm to separate dust-occluded areas and perform local repair.

[0055] The criteria for judging invalid frames are set with three rejection indicators. A frame is judged to be invalid if any one of them is met: ① Sharpness indicator: The image sharpness value is calculated by the Laplacian operator. When the sharpness value is <20, it is judged as a blurry frame; ② Percentage of valid targets: When the proportion of core targets such as construction workers and components in the image is <5%, it is judged as a meaningless frame; ③ Environmental interference: When the image signal-to-noise ratio is <15dB due to weather such as heavy rain or fog, or when the lens is blocked by an area ≥30%, it is judged as an interference frame.

[0056] The preprocessing results output standardizes the images after removing invalid frames, uniformly adjusting them to a size of 1280×720 pixels, storing them in the RGB color space, generating a standard image dataset, and adding metadata tags such as acquisition time, location, and device number.

[0057] Step 102: Construct a lightweight multi-task neural network (LMTN) to analyze standard image data. Use the YOLOv8 model to detect the wearing status of safety helmets and reflective clothing of construction workers and their violations. Combine the improved Hessian matrix ridge detection algorithm with the U-Net segmentation model to identify concrete cracks and exposed steel bars, and obtain image recognition results.

[0058] Specifically, in this embodiment, a standard image is input into an improved YOLOv8 model. Candidate boxes are generated through anchor box clustering, and duplicate boxes are filtered out using non-maximum suppression. The output target detection results include at least the personnel's location coordinates, the wearing status of safety helmets and reflective vests, the type of violation, and the confidence level. An improved Hessian matrix ridge detection algorithm is used to perform Gaussian blur noise reduction on the image, calculate the Hessian matrix, and extract feature values. The improvement lies in introducing an adaptive threshold and repairing fractures and cracks through a ridge connection algorithm. Finally, the length, width, and location information of the cracks are output. The coordinates of the crack and exposed rebar identification results are calibrated, duplicate labeled areas are removed, and a total image recognition report containing the defect type, location, size, and severity is generated.

[0059] The overall architecture design adopts a shared feature extraction + multi-task branch structure. The input layer receives the standard image data output from step one; the shared feature extraction layer uses MobileNetV3 as the backbone network, reduces the number of parameters through depthwise separable convolution (≤5 million parameters), and adds an attention mechanism in the middle layer of the network, with the ECA attention module enhancing the feature extraction of key regions; the subsequent branches are: object detection branch, branches for personnel wearing and illegal behavior and defect segmentation branches, and branches for concrete cracks and exposed rebar.

[0060] Lightweight optimization measures include channel pruning in the shared feature layer (pruning rate of 30% to remove redundant feature channels), quantization processing (8-bit quantization to reduce computational complexity), introducing depthwise separable convolutions to replace traditional convolutions in the branch network, and using global average pooling to replace fully connected layers, thereby improving the model inference speed to ≥30fps to meet real-time detection requirements.

[0061] Model improvements and training were based on improvements to the YOLOv8-nano version. SPPF was added to the neck network, and the spatial pyramid pooling fusion module was used to improve the multi-scale feature fusion capability. A dataset containing 100,000 labeled images was constructed, covering safety helmets under different lighting and angles, red, yellow, and blue reflective clothing, orange-red and fluorescent green samples, and samples of eight types of violations such as not wearing safety helmets, climbing scaffolding, and throwing objects from heights. Mosaic data augmentation was used, with a scaling range of 0.8-1.2 times and a rotation angle of ±15°. Training was carried out in three stages: a warm-up stage with a learning rate of 1e-4, a training stage with a learning rate of 1e-5, and a fine-tuning stage with a learning rate of 1e-6. Training stopped when the loss function value was ≤0.05.

[0062] The real-time detection process inputs a standard image into an improved YOLOv8 model. First, anchor box clustering is performed (9 clusters). Candidate boxes are generated based on the target size statistics of the dataset. After non-maximum suppression and an IOU threshold of 0.5, duplicate boxes are filtered out. The output target detection results include: personnel location coordinates (pixel-level accuracy), helmet / reflective vest wearing status (accuracy ≥ 95%), violation category and confidence level (confidence level ≥ 0.8 is considered valid identification).

[0063] For concrete crack identification, an improved Hessian matrix ridge detection algorithm is adopted. First, the image is Gaussian blurred and denoised with σ=1.0. Then, the Hessian matrix is ​​calculated and feature values ​​are extracted. The improvement lies in the introduction of an adaptive threshold, which dynamically adjusts the ridge response threshold according to the complexity of the concrete surface texture, ranging from 0.1 to 0.3. Through the ridge connection algorithm, the fracture crack is repaired with a minimum distance of 0.5 pixels. Finally, the length of the crack is output with an error ≤5mm, the width with an error ≤0.2mm, and the location information.

[0064] For the identification of exposed rebar defects, an improved U-Net segmentation model is used. Residual connections are added to the encoder, and ResNet18 is used as the basic module to solve the gradient vanishing problem. The decoder uses a combination of transposed convolution and pixel shuffling to improve segmentation accuracy. A dataset containing 50,000 concrete component images is constructed, and the exposed rebar areas are labeled. The Dice loss function is used during training, with a segmentation intersection-union ratio (IoU) of ≥0.85. The identification process is as follows: after the input image is segmented by the model, morphological closing operations are performed, and 5×5 structuring elements are used to fill the holes to calculate the area and exposure rate of the exposed rebar.

[0065] The recognition results are fused, and the coordinates of the crack and exposed rebar recognition results are calibrated. Based on the metadata of the collected points, duplicate marked areas are removed, and a total image recognition report is generated that includes the defect type, location, size, and severity. Cracks with a width > 0.3 mm or exposed rebar area > 0.01 m² are judged as serious defects.

[0066] Step 103: Acquire stereoscopic images of the construction site, generate a disparity map using the improved SGM algorithm, and convert it into 3D point cloud data; calculate the volume and distribution density of building materials based on the 3D point cloud data, estimate the construction load by combining material property data, and generate a load distribution map; specifically, in this embodiment, stereoscopic images of the construction site are acquired, and the initial cost is calculated by combining Census transform with grayscale difference based on the traditional SGM algorithm; a dynamic aggregation window is introduced in the aggregation step to reduce disparity error in edge areas, and a disparity map is obtained through disparity refinement; based on the intrinsic and extrinsic parameters of the binocular camera, the disparity map is converted into 3D point cloud data through the principle of triangulation.

[0067] The stereo acquisition configuration uses a binocular industrial camera with a baseline distance of 500mm and a calibration error of ≤0.1 pixels. It is deployed at the same point as the main camera in step one to ensure consistent acquisition perspective. During acquisition, the binocular camera is triggered synchronously to obtain left and right views of the same scene. The image resolution is consistent with the standard image, 1280×720, and the acquisition frequency is synchronized with the main camera, 25fps.

[0068] An improved SGM algorithm for generating disparity maps is proposed. Based on the traditional SGM algorithm, the cost calculation step is optimized by using Census transform and a 7×7 window size combined with grayscale difference to calculate the initial cost. A dynamic aggregation window is introduced in the aggregation step, with a window size of 5×5 in complex texture areas and 11×11 in flat areas to reduce disparity error in edge areas. Through disparity refinement and left-right consistency checks, interpolation is used to repair when the percentage of inconsistent pixels is <5% to obtain the final disparity map. The disparity range is 0-256 pixels, and the error is ≤1 pixel.

[0069] Point cloud conversion: Based on the intrinsic parameters of the binocular camera, focal length, principal point coordinates and extrinsic parameters, and baseline distance, the disparity map is converted into three-dimensional point cloud data through the principle of triangulation, with a point cloud density of ≥100 points / ㎡; homogeneous coordinate transformation is used to convert the point cloud data into a construction coordinate system, with the construction site benchmark point as the origin, the X-axis parallel to the construction passage, the Y-axis perpendicular to the building wall, and the Z-axis in the vertical direction.

[0070] Point cloud preprocessing first involves statistical filtering, setting K=50 neighboring points to remove noise points with a distance mean greater than 2 standard deviations to eliminate environmental noise; then, voxel mesh downsampling is used, with voxel sizes of 5mm×5mm×5mm to reduce data volume; finally, the RANSAC algorithm is used to segment the ground and background, retaining target point clouds such as construction components and building materials.

[0071] The volume calculation of building materials adopts a point cloud segmentation algorithm based on region growing. According to the surface roughness of building materials (0.5-1.0μm for concrete components, 1.0-2.0μm for steel bars) and color characteristics, the color information of RGB point clouds is used to assist in segmentation, dividing the point cloud into different types of building materials, such as precast concrete components, steel bars, and formwork. The minimum bounding volume is calculated by the convex hull algorithm for the segmented point cloud, combined with point cloud density correction. The correction coefficient is 0.95-0.98, which is adjusted according to the shape complexity of the building materials to obtain the actual volume with an error of ≤3%.

[0072] For density calculation, the construction area is divided into 1m×1m grid units. The number of building material point clouds in each grid is counted. Combining the grid area and the point cloud single-point area, the building material distribution density of each grid is calculated in kg / m². A density heat map is generated, and high-density areas with a density exceeding 120% of the design limit are marked.

[0073] Construction load estimation and distribution map generation: Density parameters of various building materials are collected, such as concrete (2400 kg / m³) and steel reinforcement (7850 kg / m³), and the weight of a single component is calculated based on volume. The weight of construction workers is considered (75 kg / person), and the load on construction machinery is calculated as 1.2 times the rated load of the equipment. The total construction load of each grid is calculated using the superposition method. A load distribution map is generated using Matlab, marking load values, design limits, and areas exceeding the limits. The percentage exceeding the limits is calculated as: (Actual load - Design limit / Design limit) × 100%.

[0074] Step 104: Calculate the construction risk index based on the image recognition results and load distribution map. When the risk index exceeds the preset threshold, trigger the graded early warning mechanism.

[0075] Specifically, this embodiment employs a point cloud segmentation algorithm based on region growing. According to the surface roughness and color characteristics of building materials, the point cloud is segmented into different types of building materials. The minimum bounding volume is calculated using a convex hull algorithm on the segmented point cloud, and the actual volume is obtained by combining this with point cloud density correction. The construction area is divided into grid cells, and the number of building material point clouds in each grid cell is counted. The building material distribution density of each grid cell is calculated by combining the grid area and the point cloud single-point dot product. A density heatmap is generated. Density parameters of various building materials are collected, and the weight of a single component is calculated by combining this with the volume. The total construction load of each grid cell is calculated using an overlay method, generating a load distribution map.

[0076] The risk indicator system is constructed by selecting four primary indicators: personnel safety indicators (weight 0.4), defect risk indicators (weight 0.3), load exceeding standards indicators (weight 0.2), and environmental interference indicators (weight 0.1). Each primary indicator is further divided into secondary sub-indicators. For example, personnel safety indicators include compliance rates for wearing safety helmets, compliance rates for wearing reflective clothing, and the incidence rate of violations. Defect risk indicators include the number of serious defects and the density of defect distribution.

[0077] The indicators are quantified and weighted. Each secondary indicator is standardized and min-max normalization is used to convert the indicator values ​​into a 0-1 range. The weight of each sub-indicator is determined by the Analytic Hierarchy Process (AHP). The risk index is calculated as follows: Risk Index = Σ, weight of primary indicator × Σ, standardized value of secondary indicator × weight of sub-indicator. The calculation result is mapped to a score of 0-100, with higher scores indicating higher risk.

[0078] The warning thresholds are set according to the "Technical Standard for Construction Safety of Prefabricated Concrete Buildings" (JGJ355-2015) and the actual risk tolerance of the project. Three levels of warning thresholds are set: Level 1 warning, low risk, risk index 60-70; Level 2 warning, medium risk, 70-85; Level 3 warning, high risk, ≥85. At the same time, special warning thresholds are set. If the load exceeding the standard in a single area is greater than 50% or there are 3 or more serious defects, the Level 3 warning will be triggered directly.

[0079] The early warning triggering and response process is as follows: When the risk index reaches the corresponding threshold, the system automatically triggers an early warning. ① Level 1 warning: The local edge node issues an audible and visual alert, a buzzer sounds, and an indicator light flashes yellow. The system also pushes a warning message to the construction team leader's mobile app, including the risk area and preliminary cause. ② Level 2 warning: The on-site alarm continuously sounds, and the indicator light flashes red. The system pushes a message to the project manager and safety officer, requiring them to arrive on-site for verification within 15 minutes. ③ Level 3 warning: The system initiates an emergency shutdown procedure, suspends construction machinery in the relevant area, and simultaneously pushes an early warning report to the construction unit and supervision unit, notifying the emergency response team to handle the situation within 30 minutes. At the same time, the system automatically saves on-site images, point cloud data, and load data as traceability evidence.

[0080] Early warning feedback and iteration: After the early warning is handled, the safety officer enters the handling results into the system. The system updates the risk model parameters and adjusts the indicator weights based on the handling effect, such as when the risk index drops below the threshold, to improve the accuracy of subsequent risk assessments.

[0081] Its beneficial effects lie in achieving a synergistic improvement in monitoring accuracy, efficiency, and green goals through end-to-end technological innovation. The improved DHE algorithm, combined with targeted interference processing, enhances image sharpness, providing high-quality data for subsequent identification and solving the data acquisition challenges in construction environments. The lightweight LMTN network integrates dual-task branches, reducing the number of parameters, increasing inference speed, and improving the accuracy of identification features such as safety helmet wearing. Crack detection error is ≤0.2mm, enabling multi-dimensional monitoring within a single system and reducing equipment deployment costs. The improved SGM algorithm generates a 3D point cloud density of 100 points / ㎡, with a building material volume calculation error of ≤3%, and accurately marks areas exceeding standards on the load distribution map, providing data support for green construction and improving building material utilization. The tiered early warning mechanism, combined with dynamic thresholds and specific trigger conditions, shortens the early warning response time to the second level, effectively reducing safety risks and improving monitoring efficiency.

[0082] Please see Figure 2 In the image recognition-based construction monitoring method for prefabricated green buildings, multi-view images and video images of the prefabricated green building construction area are acquired through sensors. An improved DHE dynamic histogram equalization algorithm is used to enhance the images, eliminating uneven lighting and dust interference, and then invalid frames are removed to obtain standard image data. The process includes the following steps:

[0083] Step 201: Acquire multi-view images and video images of the prefabricated green building construction area using sensors;

[0084] Step 202: Based on the traditional DHE dynamic histogram equalization, an adaptive block strategy is introduced. The image is divided into 16 sub-blocks according to the average gray value. After Gaussian filtering preprocessing, histogram equalization is performed to obtain the improved DHE dynamic histogram equalization algorithm.

[0085] Step 203: Convert the color image to YUV format using the improved DHE dynamic histogram equalization algorithm. Perform the improved DHE processing only on the luminance channel, while keeping the chrominance channel unchanged to avoid color distortion. After contrast stretching optimization, convert it to RGB format to obtain an enhanced image.

[0086] The above describes embodiments of the image recognition-based prefabricated green building construction monitoring method of the present invention. Please refer to [link / reference]. Figure 3 In the image recognition-based prefabricated green building construction monitoring system, the prefabricated green building construction monitoring system includes the following modules;

[0087] The image data acquisition module is used to acquire multi-view images and video images of the construction area of ​​prefabricated green buildings through sensors. The improved DHE dynamic histogram equalization algorithm is used to enhance the images, eliminate uneven lighting and dust interference, and remove invalid frames to obtain standard image data.

[0088] The construction image recognition module is used to construct a lightweight multi-task neural network (LMTN) to analyze standard image data. It uses the YOLOv8 model to detect the wearing status of safety helmets and reflective clothing of construction workers and their violations. It combines the improved Hessian matrix ridge detection algorithm with the U-Net segmentation model to identify concrete cracks and exposed steel bars, and obtains image recognition results.

[0089] The load distribution calculation module is used to acquire stereoscopic images of the construction site, generate disparity maps through an improved SGM algorithm, and convert them into three-dimensional point cloud data; calculate the volume and distribution density of building materials based on the three-dimensional point cloud data, estimate the construction load by combining material property data, and generate a load distribution map.

[0090] The construction risk level early warning module is used to calculate the construction risk index based on image recognition results and load distribution map. When the risk index exceeds the preset threshold, the risk level early warning mechanism is triggered.

[0091] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended technical solutions and their equivalents.

Claims

1. A method for monitoring the construction of prefabricated green buildings based on image recognition, characterized in that, The prefabricated green building construction monitoring method includes the following steps: Multi-view images and video images of the prefabricated green building construction area are acquired by sensors. An improved DHE dynamic histogram equalization algorithm is used to enhance the images, eliminate uneven lighting and dust interference, and remove invalid frames to obtain standard image data. In this process, an adaptive block strategy is introduced on the basis of traditional DHE dynamic histogram equalization. The image is divided into 16 sub-blocks according to the average gray value of the image. After Gaussian filtering preprocessing, histogram equalization is performed to obtain the improved DHE dynamic histogram equalization algorithm. A lightweight multi-task neural network (LMTN) is constructed to analyze the standard image data. The YOLOv8 model is used to detect the wearing status of safety helmets and reflective vests, as well as violations by construction workers. An improved Hessian matrix ridge detection algorithm and a U-Net segmentation model are combined to identify concrete cracks and exposed rebar defects, yielding image recognition results. Specifically, the standard image is input into the improved YOLOv8 model, which generates candidate boxes through anchor box clustering. Duplicate boxes are filtered using non-maximum suppression, and the target detection results are output, including personnel location coordinates, the wearing status of safety helmets and reflective vests, violation categories, and confidence levels. The improved Hessian matrix ridge detection algorithm is used to perform Gaussian blur denoising on the image, calculate the Hessian matrix, and extract feature values. The improvement lies in introducing an adaptive threshold and using a ridge connection algorithm to repair fractures, ultimately outputting the length, width, and location information of the cracks. The improved YOLOv8 model is based on the YOLOv8-nano version, with SPPF added to the neck network and a spatial pyramid pooling fusion module to enhance multi-scale feature fusion capabilities. The process involves acquiring stereoscopic images of the construction site, and using Census transform combined with grayscale difference to calculate the initial cost based on the traditional SGM algorithm. An improved SGM algorithm is obtained by introducing a dynamic aggregation window in the aggregation step to reduce disparity error in edge regions. A disparity map is generated using the improved SGM algorithm and converted into 3D point cloud data. Based on the 3D point cloud data, the volume and distribution density of building materials are calculated, and the construction load is estimated by combining material property data, generating a load distribution map. The construction risk index is calculated based on the image recognition results and load distribution map. When the risk index exceeds a preset threshold, a graded early warning mechanism is triggered.

2. The image recognition-based construction monitoring method for prefabricated green buildings as described in claim 1, characterized in that, The process involves acquiring multi-view images and video images of the prefabricated green building construction area via sensors, enhancing the images using an improved DHE dynamic histogram equalization algorithm, eliminating uneven lighting and dust interference, and discarding invalid frames to obtain standard image data, including: The color image is converted to YUV format using an improved DHE dynamic histogram equalization algorithm. The improved DHE processing is performed only on the luminance channel, while the chrominance channel remains unchanged to avoid color distortion. After contrast stretching optimization, it is converted to RGB format to obtain an enhanced image.

3. The image recognition-based construction monitoring method for prefabricated green buildings as described in claim 2, characterized in that, The process involves acquiring multi-view images and video images of the prefabricated green building construction area via sensors, enhancing the images using an improved DHE dynamic histogram equalization algorithm, eliminating uneven lighting and dust interference, and discarding invalid frames to obtain standard image data, including: An adaptive brightness correction algorithm is used to calculate the average brightness of each sub-block of the enhanced image. Brightness compensation is performed on areas that are 60% lower than the global average. Fine dust noise is removed by morphological opening operation. Then, an adaptive threshold segmentation algorithm is used to separate dust-occluded areas and perform local repair to obtain the corrected image. Invalid frames are removed from the corrected image. The image after removing invalid frames is then standardized and the pixel size is adjusted to generate standard image data.

4. The image recognition-based construction monitoring method for prefabricated green buildings as described in claim 1, characterized in that, The constructed LMTN lightweight multi-task neural network analyzes the standard image data and uses a YOLOv8 model to detect the wearing status of safety helmets and reflective vests of construction workers and their violations, including: The system employs a shared feature extraction + multi-task branch structure. The input layer receives the standard image data output from step one. The shared feature extraction layer uses MobileNetV3 as the backbone network, reduces the number of parameters (≤5 million) through depthwise separable convolutions, and adds an attention mechanism in the middle layer of the network. The ECA attention module enhances the feature extraction of key regions. Subsequent branches are: an object detection branch for personnel wearing and illegal behavior, and a defect segmentation branch for concrete cracks and exposed rebar. Lightweight optimization measures include channel pruning in the shared feature layer (pruning rate of 30% to remove redundant feature channels), quantization processing (8-bit quantization to reduce computational complexity), introducing depthwise separable convolutions to replace traditional convolutions in the branch network, and using global average pooling to replace fully connected layers, thereby improving the model inference speed to ≥30fps to meet real-time detection requirements. Model improvements and training were based on improvements to YOLOv8-nano, including the addition of SPPF to the neck network and the enhancement of multi-scale feature fusion capabilities through a spatial pyramid pooling fusion module. A dataset of 100,000 labeled images was constructed, covering safety helmets under different lighting and angles, red, yellow, and blue reflective vests, orange-red and fluorescent green samples, and eight types of violations including not wearing safety helmets, climbing scaffolding, and throwing objects from heights. Mosaic data augmentation was employed, with scaling ranging from 0.8 to 1.2 times and rotation angles of ±15°. Training was conducted in three stages, with a learning rate of 1e during the warm-up stage. -4 Training phase 1e -5 Fine-tuning stage 1e -6 Training continues until the loss function value is ≤0.

05.

5. The image recognition-based construction monitoring method for prefabricated green buildings as described in claim 1, characterized in that, The improved Hessian matrix ridge detection algorithm combined with the U-Net segmentation model identifies concrete cracks and exposed rebar defects, yielding image recognition results, including: The coordinates of the crack and exposed rebar identification results are calibrated, duplicate labeled areas are removed, and a total image recognition report containing defect type, location, size, and severity is generated.

6. The image recognition-based construction monitoring method for prefabricated green buildings as described in claim 1, characterized in that, The collected stereoscopic images of the construction site are used to generate disparity maps using the improved SGM algorithm and then converted into three-dimensional point cloud data. The volume and distribution density of building materials are calculated based on 3D point cloud data. Construction loads are estimated by combining material property data, and a load distribution map is generated, including: Based on the intrinsic and extrinsic parameters of the binocular camera, the disparity map is converted into three-dimensional point cloud data using the principle of triangulation.

7. The image recognition-based construction monitoring method for prefabricated green buildings as described in claim 6, characterized in that, The collected stereoscopic images of the construction site are used to generate disparity maps using the improved SGM algorithm and then converted into three-dimensional point cloud data. The volume and distribution density of building materials are calculated based on 3D point cloud data. Construction loads are estimated by combining material property data, and a load distribution map is generated, including: A point cloud segmentation algorithm based on region growing is adopted. According to the surface roughness and color characteristics of building materials, the point cloud is segmented into different types of building materials. The minimum bounding volume is calculated by the convex hull algorithm for the segmented point cloud. Combined with point cloud density correction, the actual volume is obtained. The construction area is divided into grid cells. The number of building material point clouds in each grid cell is counted. The building material distribution density of each grid cell is calculated by combining the grid area and the point cloud single point area; a density heat map is generated. The density parameters of various building materials are collected, and the weight of a single component is calculated by combining the volume. The total construction load of each grid is calculated by superposition method, and a load distribution map is generated.

8. A prefabricated green building construction monitoring system based on image recognition, used to implement the prefabricated green building construction monitoring method based on image recognition as described in claim 1, characterized in that, The prefabricated green building construction monitoring system includes the following modules: The image data acquisition module is used to acquire multi-view images and video images of the construction area of ​​prefabricated green buildings through sensors. The improved DHE dynamic histogram equalization algorithm is used to enhance the images, eliminate uneven lighting and dust interference, and remove invalid frames to obtain standard image data. The construction image recognition module is used to construct a lightweight multi-task neural network (LMTN) to analyze the standard image data. It uses the YOLOv8 model to detect the wearing status of safety helmets and reflective clothing of construction workers and their violations. It combines the improved Hessian matrix ridge detection algorithm with the U-Net segmentation model to identify concrete cracks and exposed steel bars, and obtains the image recognition results. The load distribution calculation module is used to acquire stereoscopic images of the construction site, generate disparity maps through an improved SGM algorithm, and convert them into three-dimensional point cloud data; calculate the volume and distribution density of building materials based on the three-dimensional point cloud data, estimate the construction load by combining material property data, and generate a load distribution map. The construction graded early warning module is used to calculate the construction risk index based on the image recognition results and load distribution map. When the risk index exceeds a preset threshold, the graded early warning mechanism is triggered.

9. The prefabricated green building construction monitoring system based on image recognition as described in claim 8, characterized in that, The image data acquisition module includes the following sub-modules: The calculation submodule is used to calculate the average brightness of each sub-block of the enhanced image using an adaptive brightness correction algorithm, perform brightness compensation on areas that are 60% lower than the global average, remove fine dust noise through morphological opening operation, and then use an adaptive threshold segmentation algorithm to separate dust-occluded areas and perform local repair to obtain the corrected image. The removal submodule is used to remove invalid frames from the corrected image, standardize the image after removing invalid frames, adjust the pixel size, and generate standard image data.

10. The prefabricated green building construction monitoring system based on image recognition as described in claim 8, characterized in that, The construction image recognition module includes the following sub-modules: The output submodule is used to input standard images into the improved YOLOv8 model, generate candidate boxes through anchor box clustering, filter duplicate boxes through non-maximum suppression, and output target detection results, which include at least the personnel location coordinates, the wearing status of safety helmets and reflective clothing, the category of violation, and the confidence level.

Images