Building constructor safety helmet monitoring method based on real-time monitoring

Abstract

The invention discloses a building construction personnel safety helmet monitoring method based on real-time monitoring, and the method comprises the steps: collecting image data, and marking an image of a construction personnel wearing a safety helmet and an image of a construction personnel not wearing the safety helmet in the image data; expanding the image data through a generative adversarial network model, and reinforcing and expanding the generated image data by using recursive reinforcement learning; inputting the enhanced and expanded image data into a three-layer full-connection neural network model for feature extraction model training; inputting the image data after feature extraction into a self-encoding neural network model for feature dimension reduction model training; inputting the image data after dimension reduction into a classifier model, and carrying out model training by adopting a mechanical learning machine classification algorithm; and inputting newly acquired image data into the trained classifier model to output an identification result. According to the invention, the problems of insufficient data volume and weak model generalization ability of the existing construction site safety helmet wearing monitoring method are solved.

Description

Technical Field

[0001] The invention relates to the technical field of building construction, and in particular to a method for monitoring safety helmets of building construction workers based on real-time monitoring. Background Art

[0002] In the field of construction, the safety of construction workers has always been a major concern due to the complex and changing environment. Among them, wearing a hard hat is one of the most basic and important safety measures, because it is directly related to the life safety of construction workers in the event of an accident. However, there are many challenges in ensuring that every worker on site strictly abides by the regulations on wearing a hard hat in actual operation, especially in large construction sites where the resources and scope of manual supervision are very limited. With the advancement of technology, automatic safety monitoring using real-time monitoring systems has become a feasible solution. Existing technologies mainly rely on video surveillance systems to capture image data and use computer vision technology to identify the wearing of hard hats in images. Although this method improves the efficiency of monitoring to a certain extent, it usually faces problems such as insufficient data volume, weak model generalization ability, and insufficient ability to handle complex scenarios. In addition, existing monitoring systems often cannot fully meet the strict requirements of high-risk environments in terms of stability and accuracy. Summary of the invention

[0003] In order to overcome the defects of the existing technology, a construction worker safety helmet monitoring method based on real-time monitoring is provided to solve the problems of insufficient data volume and weak model generalization ability in the existing construction site safety helmet wearing monitoring method.

[0004] To achieve the above purpose, a construction worker helmet monitoring method based on real-time monitoring is provided, comprising the following steps:

[0005] Collecting image data of a real-time monitoring camera, and marking images of construction workers wearing helmets and images of construction workers not wearing helmets in the image data;

[0006] The image data is augmented by a generative adversarial network model, and the generated image data is augmented by recursive reinforcement learning;

[0007] The enhanced and expanded image data is input into a three-layer fully connected neural network model for feature extraction model training;

[0008] The image data after feature extraction is input into the autoencoder neural network model to perform feature dimension reduction model training;

[0009] The image data after dimensionality reduction is input into the classifier model and the machine learning machine classification algorithm is used to train the model;

[0010] The newly acquired image data is input into the trained classifier model to output the recognition result.

[0011] Furthermore, the step of expanding the image data by generating an adversarial network model and strengthening the expanded generated image data by recursive reinforcement learning includes:

[0012] Initializing weight parameters of the generative adversarial network model;

[0013] A recursive reinforcement learning module is used to enable the generative adversarial network model generator to gradually learn to improve the generated image during multiple iterations, and in each iteration, the generator adjusts parameters based on the result of the previous generation and the feedback of the generative adversarial network model discriminator;

[0014] During the training process, dynamically adjusting the learning rate and the weight of the loss function of the generator and the discriminator according to the characteristics of the current training data;

[0015] At the end of each training cycle, an online evaluation of the generated data is automatically performed using the preset image quality evaluation indicators;

[0016] Repeat the above steps until the preset stop iteration condition is met.

[0017] Furthermore, the step of inputting the enhanced and expanded image data into a three-layer fully connected neural network model for feature extraction model training includes:

[0018] Initializing the parameters and basic topology of the neural network;

[0019] Before each iteration, the neural network evaluates the requirements of the current task on the network structure through a built-in topology sensor, and the topology sensor outputs a topology adjustment indicator based on the current loss and network structure;

[0020] Calculate the prediction error and topology adaptability penalty of the loss function under the current network configuration;

[0021] Determining whether to perform parameter optimization or topology adjustment according to the topology adjustment index and the value of the loss function;

[0022] Execute parameter or topology adjustments;

[0023] Optimize topology and parameter tuning through feedback loops;

[0024] Dynamically adjust the reconstruction strategy of the topological structure by monitoring the performance feedback of the neural network;

[0025] Repeat the above steps until the preset stop iteration condition is met.

[0026] Furthermore, the step of inputting the image data after feature extraction into the autoencoder neural network model to perform feature dimension reduction model training includes:

[0027] Initializing the weights and biases in the autoencoder neural network with random values;

[0028] The high-dimensional feature data after feature extraction is input to the encoder, propagated through the activation function and multi-layer network structure to generate low-dimensional feature representation;

[0029] Calculate total losses;

[0030] Updating the parameters of the autoencoder neural network using a gradient descent method according to the total loss function;

[0031] Repeat the above steps until the preset stop iteration condition is met.

[0032] Furthermore, the step of inputting the dimension-reduced image data into the classifier model and using a machine learning machine classification algorithm to perform model training includes:

[0033] Initialize network weights and biases, set the weight matrix and bias vector from the input layer to the hidden layer;

[0034] In the initial training stage, the number of hidden layer nodes is dynamically adjusted according to the complexity of the input data to ensure that the model can effectively capture the intrinsic structure of the data;

[0035] The data is forward propagated through the network to calculate the activation value of each hidden layer node and the final output;

[0036] Calculate the gradient of the loss function with respect to the output layer, sparse the gradient, and only retain the important gradients for back propagation;

[0037] Merge feature gradients with similar properties to simplify the network update process;

[0038] Update network parameters using optimized gradients;

[0039] Evaluate the classification performance of the model;

[0040] Repeat the above steps until the preset stop iteration condition is met.

[0041] The beneficial effect of the present invention is that the construction workers' helmet monitoring method based on real-time monitoring of the present invention utilizes a generative adversarial network based on a recursive reinforcement learning strategy to generate higher quality image data to overcome the problem of insufficient training data, and can improve the authenticity and details of the generated images.

[0042] The construction workers' helmet monitoring method based on real-time monitoring of the present invention solves the problems of gradient vanishing, gradient exploding or falling into local optimum that traditional neural networks may encounter in specific tasks by dynamically adjusting the topological structure of the internal connections of the network, thereby improving the learning efficiency and stability of the model.

[0043] The construction worker safety helmet monitoring method based on real-time monitoring of the present invention uses an autoencoder to perform feature dimensionality reduction. The present invention uses dual loss calculation to optimize the fitness of the encoded features and the target dimensionality reduction space, thereby improving the expression ability of the features.

[0044] The construction worker helmet monitoring method based on real-time monitoring of the present invention can generate more high-quality training samples and improve the model's ability to process complex data through improved data expansion technology and neural network structure optimization.

[0045] The construction worker helmet monitoring method based on real-time monitoring of the present invention uses topology adaptive learning and feature dimension reduction optimization strategies to make the model more stable during the training process and avoid common training problems such as gradient disappearance and explosion.

[0046] The construction worker helmet monitoring method based on real-time monitoring of the present invention is combined with the optimization strategy of the extreme learning machine. The present invention provides a more accurate classification result of construction worker helmet monitoring, and improves the efficiency and reliability of real-time monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0048] Figure 1 The present invention is a flowchart of a method for monitoring construction workers' safety helmets based on real-time monitoring according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0050] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0051] Reference Figure 1 As shown, the present invention provides a construction worker helmet monitoring method based on real-time monitoring, comprising the following steps:

[0052] S1. Collect image data from a real-time monitoring camera, and mark images of construction workers wearing helmets and images of construction workers not wearing helmets in the image data.

[0053] The image data of the present invention is collected from a real-time monitoring camera at a construction site. The real-time monitoring camera is installed according to a preset position and angle, and the collected image data is stored in a JPEG format.

[0054] In this embodiment, the collected data is annotated manually. The annotator classifies and marks the construction workers in each image and whether they are wearing safety helmets. The annotated categories include "wearing safety helmets" and "not wearing safety helmets".

[0055] S2. Expand the image data through the generative adversarial network model, and use recursive reinforcement learning to strengthen the expanded generated image data.

[0056] The acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization of the model and affect the accuracy of the model. The present invention uses a generative adversarial network to generate samples and thus achieve data expansion. Traditional generative adversarial networks rely on adversarial losses between generators and discriminators to train models, which can easily lead to mode collapse or unstable quality of generated images. The present invention uses recursive reinforcement learning to strengthen the similarity between generated images and real images at the feature level to improve the details and quality of generated images.

[0057] Specifically, step S2, expanding the image data by generating an adversarial network model, and using recursive reinforcement learning to strengthen the expanded generated image data, includes:

[0058] S21. Initialize the weight parameters of the generative adversarial network model.

[0059] The weight parameter initialization calculation method of the generative adversarial network is:

[0060]

[0061] In the formula, Initialize the weight of the lth layer, Re b () is the ReLU activation function, k bi is the convolution kernel, is the input feature map of the previous layer, is the bias term, and M is the number of convolution kernels.

[0062] When the generator generates a fake image, the feature error fed back in the discriminator is input into the generator as an embedding vector to improve the quality and authenticity of the generated image, which can be expressed as:

[0063]

[0064] In the formula, G b is the generator function, z is the input random noise vector, y b is the conditional variable, θ g is the generator parameter, is the weight matrix of the generator, ⊕ represents the feature fusion operation, and φ is the weight matrix of the discriminator D b The returned feature extraction function.

[0065] In this embodiment, the specific implementation of the feature fusion operation can be expressed as:

[0066] z⊕φ(D b (G b (z,y b )))=z+α b ·φ(D b (G b (z,y b ))),

[0067] In the formula, α b is a tuning parameter used to balance the input noise vector z and the feature φ(D b (G b (z,y b )))’s contribution ratio.

[0068] In this embodiment, the adjustment parameter α b The dynamic adjustment method can be expressed as:

[0069]

[0070] In the formula, κ b Control the adjustment speed, t is the current training iteration number, τ b is the iteration threshold at which the adjustment parameter begins to change significantly. Preferably, κ b Set to 0.3, τ b Set to 100.

[0071] S22. A recursive reinforcement learning module is used to enable the generative adversarial network model generator to gradually learn and improve the generated images during multiple iterations. In each iteration, the generator adjusts parameters based on the results of the previous generation and the feedback from the generative adversarial network model discriminator.

[0072] The adjustment method can be expressed as:

[0073]

[0074] In the formula, θ g is the parameter of the generator, α b is the learning rate, λ b and γ b is the loss weight, L GAN To generate the adversarial loss, L feat is the feature matching loss, G b and D b are the generator and discriminator respectively, X b is the real image data.

[0075] In one embodiment, the feature matching loss L feat The similarity between the optimized generated image and the real image at the feature level can be expressed as:

[0076]

[0077] In the formula, f i () represents the i-th feature extraction function applied to the image, F is the total number of feature extraction functions, and ∥∥2 is the L2 norm.

[0078] S23. During the training process, the learning rate and loss function weights of the generator and discriminator are dynamically adjusted according to the characteristics of the current training data.

[0079] The adjustment method can be expressed as:

[0080]

[0081]

[0082] In the formula, and are the learning rate and loss weight of the tth training cycle, respectively. and is the initial value of the learning rate and loss function weight of the generator and discriminator, η b and β b is the adjustment factor, perf(D b ,t) represents the performance index of the discriminator in the tth cycle. Preferably, and Set to 0.01 and 0.03 respectively.

[0083] In one embodiment, the performance indicator function perf() is characterized by dynamically adjusting the loss weight, and the calculation method can be expressed as:

[0084]

[0085] in, and They represent the discriminator D at the tth iteration respectively. b The number of true positives, true negatives, false positives, and false negatives.

[0086] S24. At the end of each training cycle, an online evaluation of the generated data is automatically performed, and the evaluation is performed using a preset image quality evaluation index.

[0087] The evaluation method can be expressed as:

[0088]

[0089]

[0090] In the formula, Q b is the quality score of the generated image, ψ b () is the evaluation function, z n is random noise input, y b is the conditional variable, N is the number of images to be evaluated, and ρ b To optimize the step size, For the generator G b Image quality score Q b In one embodiment, the evaluation function ψ b () is set to the same function as the performance indicator function perf().

[0091] S25, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed. b Decide whether to stop iteration. Preferably, the preset threshold for stopping iteration is 0.95.

[0092] After the data expansion model training is completed, the trained data expansion model is used to increase the number of samples. In one embodiment, assuming that the original collected samples are 800, and the data expansion model expands and generates 200 samples, the expanded data set contains 1000 samples.

[0093] S3. Input the enhanced and expanded image data into a three-layer fully connected neural network model for feature extraction model training.

[0094] The expanded data is input into the feature extraction model to train the feature extraction model. The present invention uses a 3-layer fully connected neural network for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, gradient disappearance, gradient explosion or falling into local optimal solutions may be encountered, affecting the stability of training and the performance of the model. The present invention uses a feature extraction model of a neural network optimized based on a topological adaptive learning algorithm, and adopts a topological dynamic adjustment mechanism, allowing the network to adjust not only weights and biases during the learning process, but also the topological structure of its internal connections.

[0095] Specifically, step S3, inputting the enhanced and expanded image data into a three-layer fully connected neural network model for feature extraction model training, includes:

[0096] S31. Initialize the parameters of the neural network (weight w and bias b) and the basic topology, and define the initial weight matrix W (0) and the bias vector B (0) for:

[0097] W (0) =randn(d in ,d out ),

[0098] B (0) =randn(0,d out ),

[0099] Where, d in and d out They represent the input and output dimensions of the network layer respectively, and randn(,) is a randomly selected value in a specific interval.

[0100] S32. Before each iteration, the neural network evaluates the requirements of the current task on the network structure through the built-in topology sensor. The topology sensor is defined as a function T. The topology sensor outputs a topology adjustment index τ based on the current loss L and the network structure S, which can be expressed as:

[0101] τ=T(L,S),

[0102]

[0103] Where Sig() is the Sigmoid activation function.

[0104] Furthermore, the gradient The calculation method can be expressed as:

[0105]

[0106] Where E is the prediction error, which is calculated by the preset Softmax function.

[0107] S33. Calculate the prediction error E and topology adaptability penalty P of the loss function L under the current network configuration.

[0108] The loss function L consists of two parts: prediction error E and topology adaptability penalty P. The prediction error E and topology adaptability penalty P can be calculated as:

[0109] L=E+λ sa P;

[0110]

[0111]

[0112] In the formula, λ sa is the topology weight hyperparameter, Δs j represents the change in the jth structural element (such as connection weight). Preferably, λ sa is set to 0.1.

[0113] Δs j The calculation method can be expressed as:

[0114]

[0115] In the formula, and are the structural states in consecutive iterations, respectively.

[0116] S34. According to the values ​​of the topology adjustment index τ and the loss function L, decide whether to perform parameter optimization or topology adjustment.

[0117] The adjustment method is realized through the decision function D, which can be expressed as:

[0118]

[0119] In the formula, θ cp is the threshold value. Preferably, θ cp Set to 0.5.

[0120] S35. Execute parameter or topology adjustment.

[0121] The parameter adjustment rule can be expressed as:

[0122]

[0123]

[0124] In the formula, w (t+1) and b (t+1)are the adjusted neural network weights and biases, w (t) and b (t+1) is the neural network weight and bias before adjustment, η vh is the learning rate. Preferably, η vh Set to 0.01.

[0125] gradient Calculated by the chain rule, it can be expressed as:

[0126]

[0127] Similarly, Calculated by the chain rule, it can be expressed as:

[0128]

[0129] In the formula, y i is the actual label of the i-th sample, is the predicted label of the i-th sample output by the preset Softmax function.

[0130] S36. Optimize topology and parameter tuning through feedback loops.

[0131] Specifically, adjusting η or λ to optimize the learning process can be expressed as:

[0132] η vh (t+1) =η vh (t) ×β bu ;

[0133] λ sa (t+1) =λ sa (t) +α bu ;

[0134] Where η vh (t+1) and λ sa (t+1) are the adjusted learning rate and topology weight hyperparameters, η vh (t) and λ sa (t) are the learning rate and topology weight hyperparameters before adjustment, β bu and α bu is the adjustment factor. Preferably, β bu and α bu Set to 0.9 and 0.01 respectively.

[0135] S37. By monitoring the performance feedback of the neural network, the reconstruction strategy of the topology structure is dynamically adjusted to achieve more sophisticated network optimization. Specifically, the performance monitoring function R is defined as a measure of the stability and accuracy of the network output, which can be expressed as:

[0136] R=ω1R s +ω2R p ;

[0137] In the formula, R s Represents output stability, obtained by calculating the output variance, R p represents accuracy, which is measured by the inverse loss function value. The weights ω1 and ω2 adjust the weights of stability and accuracy. Preferably, ω1 and ω2 are set to 0.6 and 0.4 respectively.

[0138] R s and R p The calculation method can be expressed as:

[0139]

[0140]

[0141] In the formula, yes The average value, E max is the maximum possible error. Preferably, E max Set to 0.5.

[0142] Adopting adaptive adjustment factor δ bu , dynamically adjust the topology according to the value of R, which can be expressed as:

[0143]

[0144] In the formula, κ up is the upper limit of the adjustment rate, λ vu is the sensitivity of the adjustment. Preferably, κ up and λ vu Set to 0.1 and 5 respectively.

[0145] According to δ, the topology can be expressed as:

[0146] S (t+1) =S (t) +δ bu ΔS;

[0147] Where S is the topology before adjustment, S (t+1) is the adjusted topological structure, and ΔS is the topological adjustment direction derived based on the gradient or other optimization algorithms.

[0148] S38, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0149] S4. Input the image data after feature extraction into the autoencoder neural network model to perform feature dimension reduction model training.

[0150] The data after feature extraction is input into the feature dimensionality reduction model to train the feature dimensionality reduction model. The present invention adopts a micro-adjustable strategy autoencoder neural network as the feature dimensionality reduction model. The micro-adjustable strategy autoencoder neural network includes an encoder and a decoder. The encoder is responsible for compressing the input high-dimensional features into a low-dimensional feature space, and the decoder attempts to reconstruct the original input data from this low-dimensional feature space. In order to improve the training efficiency and reconstruction quality of the model, the present invention adopts a dual loss calculation method, which not only considers the reconstruction error, but also evaluates the fitness of the encoded low-dimensional features and the target dimensionality reduction space, thereby optimizing the quality of feature representation.

[0151] Specifically, step S4, the step of inputting the image data after feature extraction into the autoencoder neural network model to perform feature dimension reduction model training includes:

[0152] S41. Initialize the weights and biases in the autoencoder neural network with random values.

[0153] In the initialization stage, all weights w and biases b in the network are initialized with random values. In one embodiment, Gaussian distribution is used for initialization, which can be expressed as:

[0154]

[0155]

[0156] In the formula, w pij represents the connection weight from the jth neuron to the ith neuron in the autoencoder; b pi represents the bias of the i-th neuron; σ cb is the standard deviation, used to control the dispersion of the initialization. Preferably, σ cb Set to 0.01.

[0157] S42, input the high-dimensional feature data after feature extraction to the encoder, propagate through the activation function and multi-layer network structure, and generate low-dimensional feature representation. Each layer of the encoder outputs z pl Defined as:

[0158] z pl =Sig(∑ j wplj ·z p(l-1) +b pl );

[0159] Further, the low-dimensional features enter the decoder, and the decoder attempts to reconstruct the original high-dimensional data. The output of the decoder reconstructs for:

[0160]

[0161] Where, is the output of the encoder layer l; y pk is the k-th layer input of the decoder; Sig() is the Sigmoid activation function.

[0162] Furthermore, let the last layer output of the encoder be z pm , a differentiable adjustment layer is used to adjust the encoder output z pm for It can be expressed as:

[0163]

[0164] Where ⊙ represents element-wise multiplication; α p is a learnable adjustment factor; δ p is a parameter adjusted according to the gradient information. Preferably, δ p Set to 0.3.

[0165] In this embodiment, the adjustment factor α p The calculation method can be expressed as:

[0166]

[0167] In the formula, Represents the feature after encoding from the kth to the weight of the adjustment factor; is the bias of the adjustment factor; Re() is the ReLU activation function.

[0168] S43. Calculate the total loss L p .

[0169] Total loss L p , including reconstruction loss and dual loss.

[0170] As for the reconstruction loss, in this embodiment, the mean square error is used to measure the difference between the reconstructed data and the original data.

[0171] For the dual loss, it involves evaluating the matching degree between the encoded features and the predetermined low-dimensional target features to optimize the representation ability of the features.

[0172] Specifically, the total loss Lp The calculation method can be expressed as:

[0173]

[0174] Where N is the number of samples; x pi is the i-th original data point; is the i-th reconstructed data point; D() is the dual loss function; z target is the target low-dimensional space; p is the weight factor of the dual loss. Preferably, λ p Set to 0.4.

[0175] Furthermore, the dual loss function D() is used to measure the encoded feature z pm and the target low-dimensional space z target The degree of matching between them can be calculated as:

[0176]

[0177] In the formula, z pmi is the i-th feature dimension after encoding; z targeti is the i-th feature dimension in the target low-dimensional space.

[0178] S44. Update the parameters of the autoencoder neural network using the gradient descent method according to the total loss function.

[0179] In the micro-adjustable layer, the parameters are adjusted according to the gradient information to achieve better information compression and feature expression. Specifically, let the gradient and Used to update parameters, the update method of the autoencoder parameters can be expressed as:

[0180]

[0181]

[0182] In the formula, and are the updated autoencoder weights and bias parameters, and are the autoencoder weights and bias parameters before updating; γ p is the learning rate. Preferably, γ p Set to 0.01.

[0183] Furthermore, the gradient of the weight The calculation method can be expressed as:

[0184]

[0185] In the formula, is the partial derivative of the loss function with respect to the model output, It is the partial derivative of the model output with respect to the weight.

[0186] Similarly, the gradient of the bias The calculation method can be expressed as:

[0187]

[0188] In the formula, is a constant term, preferably set to 1.

[0189] S45. Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed.

[0190] In this embodiment, the preset condition for stopping iteration is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0191] S5. Input the reduced-dimensional image data into the classifier model and use the machine learning machine classification algorithm to perform model training.

[0192] The reduced-dimensional data is input into a classifier model to train the classifier model. The present invention adopts an extreme learning machine classification algorithm based on dynamic hierarchical gradient sparsification and gradient merging optimization. The extreme learning machine classification algorithm includes an input layer, a hidden layer, and an output layer. During the training process, the gradient is hierarchically and sparsified so that it pays more attention to the features meaningful to the classification in each layer of the network, while maintaining the generalization ability of the model and reducing the risk of overfitting.

[0193] Specifically, step S5, inputting the reduced-dimensional image data into the classifier model and using a machine learning machine classification algorithm to perform model training comprises:

[0194] S51, initialize network weights and biases, set the weight matrix W from the input layer to the hidden layer q and the bias vector b q .

[0195] In this embodiment, the initialization method is random initialization.

[0196] S52. In the initial training stage, the number of hidden layer nodes is dynamically adjusted according to the complexity of the input data to ensure that the model can effectively capture the intrinsic structure of the data.

[0197] The adjustment strategy can be expressed as:

[0198] N q =β q ×feas;

[0199] Where N q is the number of hidden layer nodes; β q is the adjustment coefficient set according to the complexity of the data; feas represents the number of features of the input data.

[0200] S53, data is forward propagated through the network to calculate the activation value of each hidden layer node and the final output.

[0201] Specifically, for the input feature vector x q , the output of the hidden layer h q The calculation method can be expressed as:

[0202] h q =Re(W q ·x q +b q );

[0203] Where Re() represents the ReLU activation function; h q represents the output of the hidden layer.

[0204] S54. Calculate the gradient of the loss function with respect to the output layer The gradient is sparsely processed, and only important gradients are retained for back propagation. The calculation method can be expressed as:

[0205]

[0206] In the formula, is the output layer gradient; θ q is the sparsification threshold, which is used to determine which gradient values ​​are retained.

[0207] S55. Merge feature gradients with similar properties to simplify the network update process, which can be expressed as:

[0208] ΔW q =Mer(ΔW q );

[0209] In the formula, ΔW q Represents the gradient change of weight.

[0210] In this embodiment, the calculation method of the gradient merging function Mer() can be expressed as:

[0211]

[0212] In the formula, is the gradient change corresponding to the i-th feature; α i is the weight assigned according to feature similarity, and is equal to 1; k is the total number of similar features to be merged.

[0213] S56. Use the optimized gradient to update the network parameters.

[0214] The network parameters include weights and biases, and the update method can be expressed as:

[0215] W q ←W q -η q ΔW q ;

[0216] b q ←b q -η q Δb q ;

[0217] Where η q represents the learning rate; ΔW q and Δb q are the gradient updates for weights and biases, respectively.

[0218] Furthermore, the calculation method for the update amount of weights and biases can be expressed as:

[0219]

[0220]

[0221] In the formula, γ q is the momentum coefficient, used to smooth the update process; q is the learning rate; R represents the L2 regularization term, which is used to prevent overfitting. Preferably, γ q Set to 0.3, λ q Set to 0.05.

[0222] S57. Evaluate the classification performance of the model.

[0223] The dynamic number of nodes and learning rate hyperparameters are adjusted according to the evaluation results. In one embodiment, the method of adjusting the learning rate and number of nodes according to the performance feedback during the training process can be expressed as:

[0224] η q =η q ×exp(-δ q ×MSE);

[0225] N q =N q +κ q ×(acc-tar);

[0226] In the formula, δ q is the decay factor for adjusting the learning rate; MSE is the mean square error; κ qis the sensitivity factor of node number adjustment; acc is the current model accuracy; ta is the target accuracy. Preferably, δ q Set to 0.95, κ q Set to 0.3.

[0227] S58, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0228] S6. Input the newly acquired image data into the trained classifier model to output the recognition result.

[0229] The trained model is used to process newly collected image data to achieve real-time monitoring of the helmet wearing status of construction workers.

[0230] In one embodiment, the input data is a construction site image captured by a real-time monitoring camera. It is first processed by a trained feature extraction model. Furthermore, the extracted features are further processed by an autoencoder for dimensionality reduction. Finally, the reduced feature vector is fed into a classifier model to output the worker helmet monitoring recognition result. In this embodiment, the recognition results include "wearing a helmet" and "not wearing a helmet".

[0231] The construction worker helmet monitoring method based on real-time monitoring of the present invention utilizes a generative adversarial network based on a recursive reinforcement learning strategy to generate higher quality image data to overcome the problem of insufficient training data, and can improve the authenticity and details of the generated images.

[0232] The construction workers' helmet monitoring method based on real-time monitoring of the present invention solves the problems of gradient vanishing, gradient exploding or falling into local optimum that traditional neural networks may encounter in specific tasks by dynamically adjusting the topological structure of the internal connections of the network, thereby improving the learning efficiency and stability of the model.

[0233] The construction worker safety helmet monitoring method based on real-time monitoring of the present invention uses an autoencoder to perform feature dimensionality reduction. The present invention uses dual loss calculation to optimize the fitness of the encoded features and the target dimensionality reduction space, thereby improving the expression ability of the features.

[0234] The construction workers' helmet monitoring method based on real-time monitoring of the present invention adopts the dynamic hierarchical gradient sparsification and gradient merging optimization of the extreme learning machine, which enhances the model's learning of key features, reduces the risk of overfitting of the model, and improves the generalization ability.

[0235] The construction worker helmet monitoring method based on real-time monitoring of the present invention can generate more high-quality training samples and improve the model's ability to process complex data through improved data expansion technology and neural network structure optimization.

[0236] The construction worker helmet monitoring method based on real-time monitoring of the present invention uses topology adaptive learning and feature dimension reduction optimization strategies to make the model more stable during the training process and avoid common training problems such as gradient disappearance and explosion.

[0237] The construction worker helmet monitoring method based on real-time monitoring of the present invention is combined with the optimization strategy of the extreme learning machine. The present invention provides a more accurate classification result of construction worker helmet monitoring, and improves the efficiency and reliability of real-time monitoring.

[0238] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. For example, the above features are replaced with (but not limited to) technical features with similar functions disclosed in the present application.

Claims

1. A construction worker helmet monitoring method based on real-time monitoring, characterized in that: The following steps are involved: Collecting image data of a real-time monitoring camera, and marking images of construction workers wearing helmets and images of construction workers not wearing helmets in the image data; The image data is augmented by a generative adversarial network model, and the generated image data is augmented by recursive reinforcement learning; The enhanced and expanded image data is input into a three-layer fully connected neural network model for feature extraction model training; The image data after feature extraction is input into the autoencoder neural network model to perform feature dimension reduction model training; The image data after dimensionality reduction is input into the classifier model and the machine learning machine classification algorithm is used to train the model; The newly acquired image data is input into the trained classifier model to output the recognition result.

2. The method for monitoring construction workers' helmets based on real-time monitoring according to claim 1 is characterized in that: The step of expanding the image data by generating an adversarial network model and strengthening the expanded generated image data by recursive reinforcement learning comprises: Initializing weight parameters of the generative adversarial network model; A recursive reinforcement learning module is used to enable the generative adversarial network model generator to gradually learn to improve the generated image during multiple iterations, and in each iteration, the generator adjusts parameters based on the result of the previous generation and the feedback of the generative adversarial network model discriminator; During the training process, dynamically adjusting the learning rate and the weight of the loss function of the generator and the discriminator according to the characteristics of the current training data; At the end of each training cycle, an online evaluation of the generated data is automatically performed using the preset image quality evaluation indicators; Repeat the above steps until the preset stop iteration condition is met.

3. The method for monitoring construction workers' helmets based on real-time monitoring according to claim 1 is characterized in that: The step of inputting the enhanced and expanded image data into a three-layer fully connected neural network model for feature extraction model training comprises: Initializing the parameters and basic topology of the neural network; Before each iteration, the neural network evaluates the requirements of the current task on the network structure through a built-in topology sensor, and the topology sensor outputs a topology adjustment indicator based on the current loss and network structure; Calculate the prediction error and topology adaptability penalty of the loss function under the current network configuration; Determining whether to perform parameter optimization or topology adjustment according to the topology adjustment index and the value of the loss function; Execute parameter or topology adjustments; Optimize topology and parameter tuning through feedback loops; Dynamically adjust the reconstruction strategy of the topological structure by monitoring the performance feedback of the neural network; Repeat the above steps until the preset stop iteration condition is met.

4. The method for monitoring construction workers' helmets based on real-time monitoring according to claim 1 is characterized in that: The step of inputting the image data after feature extraction into the autoencoder neural network model to perform feature dimension reduction model training comprises: Initializing the weights and biases in the autoencoder neural network with random values; The high-dimensional feature data after feature extraction is input to the encoder, propagated through the activation function and multi-layer network structure to generate low-dimensional feature representation; Calculate total losses; Updating the parameters of the autoencoder neural network using a gradient descent method according to the total loss function; Repeat the above steps until the preset stop iteration condition is met.

5. The method for monitoring construction workers' helmets based on real-time monitoring according to claim 1 is characterized in that: The step of inputting the reduced-dimensional image data into the classifier model and using the machine learning machine classification algorithm to perform model training comprises: Initialize network weights and biases, set the weight matrix and bias vector from the input layer to the hidden layer; In the initial training stage, the number of hidden layer nodes is dynamically adjusted according to the complexity of the input data to ensure that the model can effectively capture the intrinsic structure of the data; The data is forward propagated through the network to calculate the activation value of each hidden layer node and the final output; Calculate the gradient of the loss function with respect to the output layer, sparse the gradient, and retain only the important gradients for back propagation; Merge feature gradients with similar properties to simplify the network update process; Update network parameters using optimized gradients; Evaluate the classification performance of the model; Repeat the above steps until the preset stop iteration condition is met.

Images