Real-time high-precision monitoring method and system for safety equipment in complex construction environment

Abstract

The invention provides a real-time high-precision monitoring method and system for safety equipment in a complex construction environment, and relates to the technical field of target detection, and the method comprises the steps: obtaining a construction site construction picture, and carrying out the preprocessing; constructing a construction site safety detection model based on the amplitude parameter, the self-adaptive reduction deformation convolutional layer and the self-adaptive amplification deformation convolutional layer; training a construction site safety detection model through the preprocessed construction site construction picture; obtaining a to-be-detected construction site work picture, and obtaining feature maps of different scales through the trained construction site safety detection model; sampling the feature maps of different scales through an ARoI pooling method; and respectively inputting the feature maps of different scales after the ARoI pooling into a decoupling head of a construction site safety detection model for identification so as to realize real-time monitoring of safety working equipment. Through the improved construction site safety detection model, high-precision real-time detection of safety equipment in a complex construction environment is realized.

Description

Technical Field

[0001] The present application relates to the field of target detection technology, and in particular to a real-time high-precision monitoring method and system for safety equipment in a complex construction environment. Background Art

[0002] With the steady growth of China's national economy and the continuous advancement of the development strategy in the central and western regions, highway construction sites have shown a rapid growth trend. However, behind this growth, the traditional construction site management method has exposed a series of problems, such as difficult supervision, complicated personnel management, frequent safety accidents and lack of effective safety early warning mechanism. These problems not only lead to poor coordination among the various regulatory parties, but also slow down the construction progress, weaken the project traceability, and bring potential safety risks.

[0003] Developed countries such as the United States, Canada and Japan have long begun to develop smart construction sites, and have explored and formed development models with their own characteristics based on their respective geographical and topographical features. Although China started late in the development of smart construction sites, in recent years, driven by the policy of artificial intelligence, digital and intelligent construction has rapidly become a development trend in the industry. By integrating advanced technologies and management methods, smart construction sites have significantly improved the safety management level of construction projects, effectively reduced the incidence of safety accidents, and gradually become an indispensable part of construction projects. Despite this, the development of smart construction sites still faces challenges. The environment of the construction site is complex and changeable, the work of personnel is staggered, and there is a general lack of safety awareness, making the construction area a place full of safety hazards. At present, construction site supervision still relies on traditional manual supervision, which has shortcomings such as low efficiency, slow response, and insufficient prevention. At the same time, problems such as shortage of supervisors, heavy management tasks, inability to effectively curb violations, untimely accident handling, and difficulty in obtaining evidence are also very prominent.

[0004] Chinese patent CN117975372A relates to a construction site safety detection system and method based on a combination of YOLOv8 and Transformer encoders, which directly deploys the system model on a cloud platform (computer). While ensuring recognition accuracy, it is unable to take into account computing efficiency. Summary of the invention

[0005] The purpose of the present invention is to provide a real-time high-precision monitoring method and system for safety equipment in a complex construction environment in order to solve the problem that existing construction site safety detection methods cannot take into account both detection accuracy and detection efficiency.

[0006] The above-mentioned purpose of the present application is achieved through the following technical solutions:

[0007] S1: Obtain construction site pictures and perform preprocessing;

[0008] S2: Based on the amplitude parameter, the adaptive downscaling deformation convolution layer and the adaptive upscaling deformation convolution layer, a construction site safety detection model is constructed; the construction site safety detection model includes: a convolutional neural network, a feature fusion and detection network;

[0009] S3: Train the construction site safety detection model through the pre-processed construction site images;

[0010] S4: Obtain on-site operation pictures of the construction site to be detected, and obtain feature maps of different scales through the trained construction site safety detection model;

[0011] S5: Sample feature maps of different scales through the ARoI pooling method;

[0012] S6: The feature maps of different scales after ARoI pooling are input into the decoupling head of the construction site safety detection model for identification, so as to realize real-time monitoring of safety working equipment.

[0013] Optionally, step S1 includes: the preprocessing includes: mosaic enhancement, mixed enhancement, random dithering, color dithering and size adjustment.

[0014] Optionally, step S3 includes:

[0015] The convolutional neural network includes: a first convolutional layer, a second convolutional layer, an adaptive shrinking and deforming convolutional layer, a fourth convolutional layer, and an adaptive enlarging and deforming convolutional layer;

[0016] The first convolution layer, the second convolution layer, the adaptive downscaling deformation convolution layer, the fourth convolution layer, and the adaptive upscaling deformation convolution layer are connected in sequence.

[0017] Optionally, in the adaptive downscaling deformation convolution layer and the adaptive upscaling deformation convolution layer, the sampling points of the original convolution kernel are augmented by using the offset, as follows:

[0018] During the training process of the construction site safety detection model, the coordinate points of the sampling point pixels of the adaptive downscaling deformation convolution layer and the adaptive upscaling deformation convolution layer are transformed as follows:

[0019] X(P i )=∑w(p xi )*x(p x0 +p n +Δp xi ) (Formula 4)

[0020] Y(P i )=∑w(p yi )*y(p y0 +p n +Δp yi ) (Formula 5)

[0021] Where (X, Y) represents the position coordinates of the sampling point pixel after the change, P i =(p xi , p yi ) is the coordinate point of the sampling point pixel; w represents the weight that needs to be trained during the transformation process; p n is the offset of each sampling point in the convolution kernel relative to the center position, and the offset is {(-1, -1), (-1, 0), ..., (0, 1), (1, 1)}; Δp xi Indicates the adaptive offset in the x direction, Δp yi Indicates the adaptive offset in the y direction; p x0 represents the original x coordinate; p y0 Represents the original y coordinate.

[0022] Optionally, the construction site safety detection model further includes:

[0023] In each convolution process of the construction site safety detection model, an adaptive amplitude parameter is added;

[0024] The specific formula for adding adaptive amplitude parameters to the adaptive shrinking deformation convolution layer and adaptive enlargement deformation convolution layer of the construction site safety detection model is as follows:

[0025] X(P i )=∑w(p xi )*x(p x0 +p n +Δp xi )*Δk (Formula 6)

[0026] Y(P i )=∑w(p yi )*y(p y0 +p n +Δp yi )*Δk (Formula 7)

[0027] Where Δk is the amplitude parameter that needs to be learned.

[0028] Step S3 includes:

[0029] Adaptively enlarge the feature map F output by the deformed convolutional layer in the convolutional neural network P5 , perform maximum pooling operations of different scales to obtain multi-scale feature maps;

[0030] Combine the multi-scale feature map with the feature map F output by the fourth convolutional layer in the convolutional neural network P4 Perform splicing to obtain feature map F1;

[0031] The feature map F1 is outputted by the adaptive shrinking and deformation convolution layer in the convolutional neural network after one upsampling and C2f layer. P3 Fusion is performed to obtain feature map T1;

[0032] After the feature map T1 passes through a convolution layer once and is fused with the feature map F1, it passes through a C2f layer once to obtain the feature map T2;

[0033] After the feature map T2 is fused with the feature map T3 through a convolution layer, a C2f layer is used to obtain the feature map T3;

[0034] The feature maps T1, T2, and T3 are feature maps of different scales output by the construction site safety detection model;

[0035] When the feature maps are fused through the fully connected layer, the amplitude parameter and the scaling ratio r are used to adjust the fusion process. The calculation process is as follows:

[0036] s=∑r*w*f(z i ) (Formula 8)

[0037] Where r is the scaling ratio, w is the weight of each operation, and f(z i ) is the current feature layer; s represents the output result.

[0038] A real-time high-precision monitoring system for safety equipment in complex construction environments, the system comprising: a high-resolution camera, an embedded device, a touch display, and a wireless transmission module;

[0039] The high-resolution camera is connected to the embedded device through a wireless transmission module; the touch display is connected to the embedded development board;

[0040] High-resolution cameras are used to obtain on-site operation pictures of the construction site to be inspected;

[0041] The embedded device is used to monitor the on-site operation pictures of the construction site to be inspected in real time through the deployed construction site safety detection model and generate detection results;

[0042] The wireless transmission module is used to transmit image data in real time;

[0043] The touch display is used to visualize the test results.

[0044] A computer-readable storage medium stores instructions, and when the instructions are executed, a real-time high-precision monitoring method for safety equipment in a complex construction environment is executed.

[0045] The beneficial effects of the technical solution provided by this application are:

[0046] 1. By adaptively reducing the deformation convolution layer and adaptively enlarging the deformation convolution layer, a construction site safety detection model is constructed to improve the detection accuracy. It can effectively realize real-time monitoring of workers and safety work equipment such as safety helmets, reflective vests and safety ropes at the construction site, thereby effectively reducing the risk of accidents.

[0047] 2. After adding the amplitude parameter in the convolution process of the model, the network is allowed to weight each input according to the contextual information, suppressing the impact of task-irrelevant or noise, thereby better understanding the overall context and improving the robustness and generalization ability of the network. This improvement can effectively solve the problems of background blur and external object occlusion in construction scenes. At the same time, the integration of information between different spatial scales and channels helps the network better understand the structure and association of objects, thereby improving the accuracy of object recognition and positioning, especially in the face of multi-scale multi-target detection such as construction sites.

[0048] 3. Adopt ARoI (Adaptive Region of Interest) pooling method. By canceling the quantization method, the sampling points are determined by using similar offsets, and then the real candidate boxes are obtained by averaging. This method transforms the entire feature aggregation process into a continuous operation, thereby reducing the generation of errors and improving the detection accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present application will be further described below with reference to the accompanying drawings and embodiments, in which:

[0050] Figure 1 is a step diagram in an embodiment of the present application;

[0051] Figure 2 is a structural diagram of a construction site safety detection model in an embodiment of the present application;

[0052] Figure 3 It is a structural diagram of a common convolution module of a construction site safety detection model in an embodiment of the present application;

[0053] Figure 4 It is an effect diagram of the ASC convolution kernel and the ACC convolution kernel in the embodiment of the present application;

[0054] Figure 5 : is an application effect diagram of a convolution kernel for adaptively detecting target size in an embodiment of the present application;

[0055] Figure 6 It is a floating point process diagram in an embodiment of the present application;

[0056] Figure 7 It is a process diagram of AroI pooling in an embodiment of the present application;

[0057] Figure 8 It is a system module diagram in an embodiment of the present application. DETAILED DESCRIPTION

[0058] In order to have a clearer understanding of the technical features, purposes and effects of the present application, the specific implementation methods of the present application are now described in detail with reference to the accompanying drawings.

[0059] The embodiments of the present application provide a method for real-time and high-precision monitoring of safety equipment in a complex construction environment.

[0060] Please refer to Figure 1 , Figure 1 This is a step diagram of a method for real-time high-precision monitoring of safety equipment in a complex construction environment in an embodiment of the present application, including:

[0061] S1: Obtain construction site pictures and perform preprocessing;

[0062] S2: Based on the amplitude parameter, the adaptive downscaling deformation convolution layer and the adaptive upscaling deformation convolution layer, a construction site safety detection model is constructed; the construction site safety detection model includes: a convolutional neural network, a feature fusion and detection network;

[0063] S3: Train the construction site safety detection model through the pre-processed construction site images;

[0064] S4: Obtain on-site operation pictures of the construction site to be detected, and obtain feature maps of different scales through the trained construction site safety detection model;

[0065] S5: Sample feature maps of different scales through the ARoI pooling method;

[0066] Specifically, after fusion, the feature map will be regressed and predicted. However, when the model regresses to generate candidate boxes, the boundary coordinates are usually quantized to integer point values, which will cause a certain deviation between the generated candidate boxes and the initially regressed positions. This deviation may affect the accuracy of detection or segmentation. Based on the idea of ​​the offset in the above improvement, this paper designs an ARoI (Adaptive Region of Interest) pooling method. The general pooling method converts feature maps of any size into feature vectors of fixed size. Two quantizations are used in the implementation process, which will cause errors in the detection of small-scale targets. At the construction site, the relative distances between different targets and the camera are different, and some targets will become very small in the picture. The general pooling method will increase the probability of false detection or missed detection of such situations.

[0067] The ARoI pooling method first divides the target candidate box into equal parts, generally into 2x2 regions. After the division, the corner points of each region are not necessarily on the pixel points. Different from the traditional method of plasticization, this method chooses to retain such corner points and converts them into floating points to achieve the purpose of preserving real information. The floating point conversion process is as follows: Figure 6 On this basis, if it is a normal pooling operation, the pixels corresponding to the four points in each box will be averaged, but ARoI will first perform the adaptive deformation convolution on the sampling points in each box, obtain pixels closer to the target through bilinear interpolation, and then average the updated sampling points to obtain the final candidate box that is more suitable for the target size. Figure 7 This adaptive pooling method uses bilinear interpolation to ensure that the feature points inside each region of interest are accurately aligned in the output feature map, retaining the information in the region of interest more accurately, reducing the information loss introduced in the pooling process, and helping to improve the system's detection accuracy for small targets or targets with fine structures in construction scenes.

[0068] S6: The feature maps of different scales after ARoI pooling are input into the decoupling head of the construction site safety detection model for identification, so as to realize real-time monitoring of safety working equipment.

[0069] Specifically, the structure of the decoupling head is as follows Figure 8 As shown, it includes four 3×3 convolutional layers and two 1×1 convolutional layers. In order to improve the detection accuracy, the WIOU (Weighted Intersection over Union) loss function is added to the regression branch of the detection head. The main task of the regression head is to calculate the position offset between the predicted box and the real box.

[0070] Specifically, the regression head first calculates the four coordinate offsets between the predicted box and the true box, and then inputs these offsets into the regression head for loss calculation. Finally, the regression head outputs a four-dimensional vector, which represents the upper left corner coordinates (x1, y1) and the lower right corner coordinates (x2, y2) of the target box. Next, after a series of convolution operations, the classification head generates a classifier output tensor. The value at each position of the tensor represents the probability distribution of the candidate box belonging to different categories. In order to ensure the accuracy of the final detection result, the maximum value suppression (NMS) method is applied in the last step. Maximum value suppression suppresses the candidate boxes with higher overlap and retains the candidate boxes with the highest confidence, thereby screening out the final detection results. The deep, middle, and shallow prediction results (i.e., feature maps T1, T2, and T3) are processed and used to detect the target object and draw the target box.

[0071] Step S1 includes: the preprocessing includes: mosaic enhancement, mixed enhancement, random dithering, color dithering and size adjustment.

[0072] Specifically, the data is expanded in real time online through data augmentation methods, and then the processed image is adjusted to the fixed size required by the model while maintaining the original aspect ratio.

[0073] Step S3 includes:

[0074] The convolutional neural network includes: a first convolutional layer, a second convolutional layer, an adaptive shrinking and deforming convolutional layer, a fourth convolutional layer, and an adaptive enlarging and deforming convolutional layer;

[0075] The first convolution layer, the second convolution layer, the adaptive downscaling deformation convolution layer, the fourth convolution layer, and the adaptive upscaling deformation convolution layer are connected in sequence.

[0076] Specifically, Figure 3 The common convolution module of the construction site safety detection model shown in the figure performs multiple feature extractions on the construction site pictures. The convolution layer in each convolution module uses a convolution kernel with a step size of 2 to perform downsampling operations to reduce the size of the feature map and increase the number of channels. The feature extraction obtains a total of five scale feature maps. The convolution operations of the first convolution layer, the second convolution layer, and the fourth convolution layer are as follows:

[0077] F P1 =Conv 3×3 (Input) (Formula 1)

[0078] F P2 =Conv 3×3 (F P1 ) (Formula 2)

[0079] F P4 =Conv 3×3 (F P3 ) (Formula 3)

[0080] Where Conv represents the convolution operation, and F represents the feature map obtained after the convolution operation;

[0081] In the third convolution, the original convolution kernel is replaced by an adaptive shrinkage convolution kernel (AdaptiveShrinkage Convolution, ASC) for convolution, and in the fifth convolution, the ordinary convolution operation is replaced by an adaptive amplification convolution kernel (AdaptiveAmplification Convolution, ACC) for convolution.

[0082] In the adaptive downscaling deformation convolution layer and the adaptive upscaling deformation convolution layer, the sampling points of the original convolution kernel are augmented by using the offset, as follows:

[0083] During the training process of the construction site safety detection model, the coordinate points of the sampling point pixels of the adaptive downscaling deformation convolution layer and the adaptive upscaling deformation convolution layer are transformed as follows:

[0084] X(P i )=∑w(p xi )*x(p x0 +p n +Δp xi ) (Formula 4)

[0085] Y(P i )=∑w(p yi )*y(p y0 +p n +Δp yi ) (Formula 5)

[0086] Where (X, Y) represents the position coordinates of the sampling point pixel after the change, P i =(p xi , p yi ) is the coordinate point of the sampling point pixel; w represents the weight that needs to be trained during the transformation process; p n is the offset of each sampling point in the convolution kernel relative to the center position, and the offset is {(-1, -1), (-1, 0), ..., (0, 1), (1, 1)}; Δp xi Indicates the adaptive offset in the x direction, Δp yi Indicates the adaptive offset in the y direction; p x0 represents the original x coordinate; p y0 Represents the original y coordinate.

[0087] Specifically, when extracting features, convolutional neural networks generally use regular square convolution kernels such as 1x1, 3x3, and 5x5, which will result in the same receptive field scale of the activation units of the same layer of CNN during the convolution process. Traditional methods will limit the position of the receptive field. These two convolution methods add a separate direction vector to each convolution kernel, so that the convolution kernel can adaptively change to any shape, achieving the following results: Figure 4 The designed convolution kernels for adaptive detection of target sizes (ASC convolution kernel and ACC convolution kernel) can more efficiently extract and abstract the semantic information and visual features of objects of different sizes in the image, achieving the following results: Figure 5 These features contain various structural and contextual information of the operators and safety work equipment, providing key clues for subsequent target recognition and positioning.

[0088] The construction site safety detection model also includes:

[0089] In each convolution process of the construction site safety detection model, an adaptive amplitude parameter is added;

[0090] The specific formula for adding adaptive amplitude parameters to the adaptive shrinking deformation convolution layer and adaptive enlargement deformation convolution layer of the construction site safety detection model is as follows:

[0091] X(P i )=∑w(p xi )*x(p x0 +p n +Δp xi )*Δk (Formula 6)

[0092] Y(P i )=∑w(p yi )*y(p y0 +p n +Δp yi )*Δk (Formula 7)

[0093] Where Δk is the amplitude parameter that needs to be learned;

[0094] Specifically, in construction scenes, there may be background blur, interference from external objects, etc. In order to solve the problem of interference from irrelevant areas, this paper adds a weight coefficient as an amplitude parameter on the basis of adding sampling point offset to distinguish whether the introduced area is the area of ​​interest. The specific approach is to add adaptive amplitude parameters to each convolution part of feature maps T1, T2, and T3, so as to enhance the expressiveness and adaptability of features when cross-level or cross-scale feature fusion is involved, reduce the interference of background errors, and produce more accurate, specific and diverse feature representations.

[0095] During the convolution process, the addition of the amplitude parameter will add an extra channel to the single convolution layer on the same input feature map. This mainly includes the above-mentioned spatial resolution offset and the newly added weight channel. Unlike the weight parameter w of the offset, k is a judgment parameter for determining whether it is an area of ​​interest. When the area of ​​the sampling point is not of interest, k is 0; in other cases, the initial value of k is set to 0.5 for subsequent training.

[0096] Step S3 includes:

[0097] Adaptively enlarge the feature map F output by the deformed convolutional layer in the convolutional neural network P5 , perform maximum pooling operations of different scales to obtain multi-scale feature maps;

[0098] Combine the multi-scale feature map with the feature map F output by the fourth convolutional layer in the convolutional neural network P4 Perform splicing to obtain feature map F1;

[0099] The feature map F1 is outputted by the adaptive shrinking and deformation convolution layer in the convolutional neural network after one upsampling and C2f layer. P3 Fusion is performed to obtain feature map T1;

[0100] After the feature map T1 passes through a convolution layer once and is fused with the feature map F1, it passes through a C2f layer once to obtain the feature map T2;

[0101] After the feature map T2 is fused with the feature map T3 through a convolution layer, a C2f layer is used to obtain the feature map T3;

[0102] The feature maps T1, T2, and T3 are feature maps of different scales output by the construction site safety detection model;

[0103] When the feature maps are fused through the fully connected layer, the amplitude parameter and the scaling ratio r are used to adjust the fusion process. The calculation process is as follows:

[0104] s=∑r*w*f(z i ) (Formula 8)

[0105] Where r is the scaling ratio, w is the weight of each operation, and f(z i ) is the current feature layer; s represents the output result.

[0106] Specifically, the C2f layer is a special convolutional layer that is used to fuse feature maps of different scales together to improve the accuracy of object detection. Feature fusion and detection network are key steps in the object detection process. At this stage, features of different levels or scales are integrated and fused to produce a more comprehensive, rich and diverse feature representation.

[0107] like Figure 8 As shown, a real-time high-precision monitoring system for safety equipment in a complex construction environment, the system includes: a high-resolution camera, an embedded device, a touch display, and a wireless transmission module;

[0108] The high-resolution camera is connected to the embedded device through a wireless transmission module; the touch display is connected to the embedded development board;

[0109] High-resolution cameras are used to obtain on-site operation pictures of the construction site to be inspected;

[0110] The embedded device is used to monitor the on-site operation pictures of the construction site to be inspected in real time through the deployed construction site safety detection model and generate detection results;

[0111] The wireless transmission module is used to transmit image data in real time;

[0112] The touch display is used to visualize the test results.

[0113] Specifically, complete the construction site safety detection model (such as Figure 2 After the design is completed, it can be deployed to embedded devices for on-site inspection at the construction site. The device uses high-performance NVIDIA chips as the core computing platform to analyze and process data. First, the embedded development board will be equipped with a suitable image acquisition device, combined with a high-resolution camera and a wireless transmission module to achieve efficient data acquisition and processing. In addition, during the real-time detection process, the main chip of the development board must have strong computing power and low power consumption, so that it is suitable for use in resource-limited environments. The high-resolution camera provides high-definition images and a wide-angle field of view to ensure a wide monitoring coverage and clear data. With the wireless transmission module, real-time data upload and processing can be achieved. When the development board processes huge data, heat will be generated, causing temperature rise. Therefore, in order to ensure computing efficiency, this project adds a heat fan to the chip in the hardware design to help the development board dissipate heat and maintain a suitable temperature. Finally, a multi-function interface is incorporated into the development board, which is convenient for equipping with a display screen and a control panel, so that users can view monitoring data, test results and alarm information in real time, and can adjust monitoring parameters and view historical data through a touch screen or keyboard, so as to achieve the purpose of friendly human-computer interaction.

[0114] Specifically, after building the target detection model (construction site safety detection model), it is necessary to select a suitable data set for its training. The operation data of multiple construction scenes on the construction site were collected, and the collected image data were preprocessed by data cleaning, denoising, format conversion, etc. Through these steps, the quality and consistency of the data were ensured, so as to build a data set covering multiple construction scenes. This data set is not only rich and diverse, but also highly representative, and can reflect various situations in actual construction. After the data set is built, the data set is trained multiple times using the deep learning model designed and built previously. In order to obtain better results, a variety of optimization techniques are used in the training process, such as data enhancement, transfer learning, and hyperparameter tuning, to improve the generalization ability and accuracy of the model. After each training, the performance of the model is evaluated, and the model parameters are adjusted and optimized according to the evaluation results. After multiple iterations of training and optimization, an optimal model is obtained. This optimal model can efficiently and accurately identify and analyze various operation conditions on the construction site, and accurately detect in real time whether the construction workers on site are wearing safety equipment such as safety helmets and reflective vests.

[0115] The present application also discloses a computer-readable storage medium storing a plurality of instructions suitable for loading by a processor to execute the above-mentioned method for real-time and high-precision monitoring of safety equipment in a complex construction environment.

[0116] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure.

[0117] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A real-time high-precision monitoring method for safety equipment in a complex construction environment, characterized in that: The method comprises the following steps: S1: Obtain construction site pictures and perform preprocessing; S2: Based on the amplitude parameter, the adaptive downscaling deformation convolution layer and the adaptive upscaling deformation convolution layer, a construction site safety detection model is constructed; the construction site safety detection model includes: a convolutional neural network, a feature fusion and detection network; S3: Train the construction site safety detection model through the pre-processed construction site images; S4: Obtain on-site operation pictures of the construction site to be detected, and obtain feature maps of different scales through the trained construction site safety detection model; S5: Sample feature maps of different scales through the ARoI pooling method; S6: The feature maps of different scales after ARoI pooling are input into the decoupling head of the construction site safety detection model for identification, so as to realize real-time monitoring of safety working equipment.

2. A method for real-time high-precision monitoring of safety equipment in a complex construction environment as claimed in claim 1, characterized in that: Step S1 includes: the preprocessing includes: mosaic enhancement, mixed enhancement, random dithering, color dithering and size adjustment.

3. A real-time high-precision monitoring method for safety equipment in a complex construction environment as claimed in claim 1, characterized in that: Step S3 includes: The convolutional neural network includes: a first convolutional layer, a second convolutional layer, an adaptive shrinking and deforming convolutional layer, a fourth convolutional layer, and an adaptive enlarging and deforming convolutional layer; The first convolution layer, the second convolution layer, the adaptive downscaling deformation convolution layer, the fourth convolution layer, and the adaptive upscaling deformation convolution layer are connected in sequence.

4. A real-time high-precision monitoring method for safety equipment in a complex construction environment as claimed in claim 3, characterized in that: In the adaptive downscaling deformation convolution layer and the adaptive upscaling deformation convolution layer, the sampling points of the original convolution kernel are augmented by using the offset, as follows: During the training process of the construction site safety detection model, the coordinate points of the sampling point pixels of the adaptive downscaling deformation convolution layer and the adaptive upscaling deformation convolution layer are transformed as follows: X(p i ) = Σw(p xi ) * x(p x0 + p n + Δp xi ) (Equation 4) Y(P i ) = Σw(p yi ) * y(p y0 + p n + Δp yi ) (Equation 5) Where (X, Y) represents the position coordinates of the sampling point pixel after the change, P i =(p xi ,p yi ) is the coordinate point of the sampling point pixel; w represents the weight that needs to be trained during the transformation process; p n is the offset of each sampling point in the convolution kernel relative to the center position, and the offset is {(-1,-1), (-1,0), …, (0,1), (1,1)}; Δp xi Represents the adaptive offset in the x direction, ΔP yi Indicates the adaptive offset in the y direction; p x0 represents the original x coordinate; p y0 Represents the original y coordinate.

5. A method for real-time high-precision monitoring of safety equipment in a complex construction environment as claimed in claim 4, characterized in that: The construction site safety detection model also includes: In each convolution process of the construction site safety detection model, an adaptive amplitude parameter is added; The specific formula for adding adaptive amplitude parameters to the adaptive shrinking deformation convolution layer and adaptive enlargement deformation convolution layer of the construction site safety detection model is as follows: X(P i ) = Σw(p xi ) * x(p x0 + p n + Δp xi ) * Δk (Equation 6) Y(P i ) = Σw(p yi ) * y(p y0 + p n + Δp yi ) * Δk (Equation 7) Where Δk is the amplitude parameter that needs to be learned.

6. A real-time high-precision monitoring method for safety equipment in a complex construction environment as claimed in claim 1, characterized in that: Step S3 includes: Adaptively enlarge the feature map F output by the deformed convolutional layer in the convolutional neural network P5 , perform maximum pooling operations of different scales to obtain multi-scale feature maps; The multi-scale feature map is combined with the feature map F output by the fourth convolutional layer in the convolutional neural network P4 Perform splicing to obtain feature map F1; The feature map F1 is outputted by the adaptive shrinking and deformation convolution layer in the convolutional neural network after one upsampling and C2f layer. P3 Fusion is performed to obtain feature map T1; After the feature map T1 passes through a convolution layer once and is fused with the feature map F1, it passes through a C2f layer once to obtain the feature map T2; After the feature map T2 is fused with the feature map T3 through a convolution layer, a C2f layer is used to obtain the feature map T3; The feature maps T1, T2, and T3 are feature maps of different scales output by the construction site safety detection model; When the feature maps are fused through the fully connected layer, the amplitude parameter and the scaling ratio r are used to adjust the fusion process. The calculation process is as follows: s = Σr * w * f(z i ) (Equation 8) Where r is the scaling ratio, w is the weight of each operation, and f(z i ) is the current feature layer; s represents the output result.

7. A real-time high-precision monitoring system for safety equipment in a complex construction environment, used to implement a real-time high-precision monitoring method for safety equipment in a complex construction environment as described in claims 1-6, characterized in that: The system includes: high-resolution camera, embedded device, touch display, wireless transmission module; The high-resolution camera is connected to the embedded device through a wireless transmission module; the touch display is connected to the embedded development board; High-resolution cameras are used to obtain on-site operation pictures of the construction site to be inspected; The embedded device is used to monitor the on-site operation pictures of the construction site to be inspected in real time through the deployed construction site safety detection model and generate detection results; The wireless transmission module is used to transmit image data in real time; The touch display is used to visualize the test results.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed by a computer, the method according to any one of claims 1 to 6 is executed.

Images