Human body posture recognition method, device, equipment and medium

Abstract

The embodiment of the invention provides a human body posture recognition method and device, equipment and a medium, and the method comprises the steps: obtaining a human body posture image collected by a camera, carrying out the feature extraction of the human body posture image through a feature extraction model, obtaining an image feature, and carrying out the feature extraction of the human body posture image through a feature dimension reduction model; and performing feature dimension reduction on the image features to obtain dimension-reduced features, inputting the dimension-reduced features into a classifier model to obtain a classification result output by the classifier model, and determining a human body posture recognition result of the human body posture image according to the classification result. According to the embodiment of the invention, the optimization of human body posture recognition through the feature extraction model, the feature dimension reduction model and the classifier model is realized, the precision and efficiency of human body posture recognition are improved, and the application scene of the posture recognition technology is expanded.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method, device, equipment and medium for human body posture recognition. Background Art

[0002] With the development of artificial intelligence and computer vision technology, human posture recognition has become a hot area of ​​extensive research and application. Human posture recognition technology has important application value in many fields such as intelligent monitoring, virtual reality, medical rehabilitation, etc. However, the existing posture recognition technology has problems such as low recognition accuracy and low efficiency, which affects the application of posture recognition technology. Summary of the invention

[0003] In view of the above problems, a method, device, equipment and medium for human body posture recognition is proposed to overcome the above problems or at least partially solve the above problems, including:

[0004] A method for human body posture recognition, the method comprising:

[0005] Obtain human body posture images collected by the camera;

[0006] Using a feature extraction model to extract features from the human body posture image to obtain image features, and using a feature dimension reduction model to reduce the dimensions of the image features to obtain reduced dimension features;

[0007] The dimension reduction feature is input into a classifier model to obtain a classification result output by the classifier model, and a human body posture recognition result of the human body posture image is determined based on the classification result.

[0008] A device for human body posture recognition, the device comprising:

[0009] A human body posture image acquisition module is used to acquire human body posture images collected by a camera;

[0010] A feature processing module is used to use a feature extraction model to perform feature extraction on the human posture image to obtain image features, and use a feature dimension reduction model to perform feature dimension reduction on the image features to obtain reduced dimension features;

[0011] The human posture recognition module is used to input the dimension reduction feature into a classifier model to obtain a classification result output by the classifier model, and determine a human posture recognition result of the human posture image according to the classification result.

[0012] An electronic device comprises a processor, a memory and a computer program stored in the memory and capable of running on the processor, wherein the computer program implements the method described above when executed by the processor.

[0013] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described above is implemented.

[0014] The embodiments of the present invention have the following advantages:

[0015] In an embodiment of the present invention, a human body posture image captured by a camera is acquired, a feature extraction model is used to perform feature extraction on the human body posture image to obtain image features, and a feature dimension reduction model is used to perform feature dimension reduction on the image features to obtain reduced dimension features, the reduced dimension features are input into a classifier model to obtain a classification result output by the classifier model, and a human body posture recognition result of the human body posture image is determined based on the classification result, thereby achieving optimization of human body posture recognition through feature extraction models, feature dimension reduction models and classifier models, improving the accuracy and efficiency of human body posture recognition, and expanding the application scenarios of posture recognition technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solution of the present invention, the accompanying drawings required for use in the description of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0017] Figure 1 is a flow chart of a method for human body posture recognition provided by some embodiments of the present invention;

[0018] Figure 2a is a flowchart of a first neural network training provided by some embodiments of the present invention;

[0019] Figure 2b is a flowchart of an autoencoding neural network algorithm training provided by some embodiments of the present invention;

[0020] Figure 3a is a flow chart of another method for human body posture recognition provided by some embodiments of the present invention;

[0021] Figure 3b is a schematic diagram of a model training process provided by some embodiments of the present invention;

[0022] Figure 3c It is a schematic diagram of a model usage process provided by some embodiments of the present invention. DETAILED DESCRIPTION

[0023] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] In a related technology, an image with an image of a target object is acquired; key points of the target object in the image are detected according to a preset setting, wherein the key points are multiple structural nodes set according to the structure of the target object; key point coordinates are generated according to the positions of the key points in the image; the key point coordinates are identified according to the preset coordinate conditions corresponding to different postures of the target object, and the posture recognition results corresponding to the target object are obtained. The positional relationship between multiple key points can be mutually verified, and the posture of the target object can be accurately identified when there is interference from other images, which solves the technical problem that the posture recognition is interfered by other images in the image during the posture recognition process, causing misidentification.

[0025] In another related technology, by acquiring images of multiple people at multiple perspectives at the same time, the bounding box of each person in the multiple images is identified, and each image is matched according to the bounding box, and then verified according to the matching results. After the bounding box passes the verification, the posture information of each person is identified and determined. This process obtains multiple images of multiple people at multiple perspectives, and then identifies the bounding box of each person in the image, and determines the posture information of each person through the bounding box, thereby improving the accuracy of posture recognition under multiple perspectives.

[0026] In another related technology, the input original image is classified and converted into a semi-supervised learning data set; the model is built using a deep learning framework; the receptive field of the model is increased; the activation function is optimized; a pulse graph convolution algorithm with a learning rate reward mechanism is added; and the model is trained. The present invention increases the receptive field of the model and extracts more feature values ​​from the image by changing the expansion coefficient of the convolution model. Furthermore, a pulse graph convolution layer with a learning rate reward mechanism is added to the model; wherein the pulse graph convolution layer simulates the transmission mode between neurons and transmits information in the form of pulse signals, thereby speeding up the processing speed of the image; at the same time, the accuracy of the model is improved through the learning rate reward mechanism.

[0027] However, the above related technologies have the following problems:

[0028] 1. In terms of data expansion, it usually relies solely on manual labeling. The data collection and labeling process is time-consuming and labor-intensive, and insufficient data volume can easily lead to poor model generalization ability.

[0029] 2. In terms of feature extraction, traditional neural network structures often encounter problems such as gradient vanishing, gradient exploding, or falling into local optimal solutions, which affects the training stability and performance of the model.

[0030] 3. In terms of feature dimensionality reduction, existing methods often find it difficult to maintain the precise capture and reconstruction of the intrinsic correlation and structural characteristics of the input data while reducing the data dimension, which affects the information flow optimization of the network.

[0031] 4. In terms of classifiers, when dealing with nonlinear separable problems, traditional methods have difficulty in balancing classification performance and computing resource consumption, the model complexity is high, and the classification accuracy and efficiency are low.

[0032] First, the data collection and annotation process is complex and time-consuming, and insufficient data will affect the generalization ability of the model. Secondly, during the feature extraction process, traditional neural networks are prone to problems such as gradient vanishing and gradient explosion, resulting in unstable model training and poor performance. Furthermore, existing feature dimensionality reduction methods are difficult to effectively reduce dimensions while maintaining the intrinsic correlation and structural characteristics of the data, affecting the subsequent classification effect. Finally, when dealing with nonlinear separable problems, it is difficult for the classifier model to control the complexity of the model and computing resource consumption while ensuring classification accuracy.

[0033] In an embodiment of the present invention, by improving data quality, optimizing feature extraction and dimensionality reduction processes, and improving classification accuracy and efficiency, more accurate and efficient human posture recognition is achieved, and the generalization ability and application effect of the model are enhanced. By using a generative adversarial network based on adaptive morphological transformation for data expansion, the quantity and quality of training samples are significantly improved; a neural network algorithm based on adaptive biological influence optimization is used for feature extraction to avoid the problems of gradient vanishing, gradient explosion and local optimal solution; an autoencoder neural network based on feature adaptive refinement is used for feature dimensionality reduction to accurately capture and reconstruct the intrinsic correlation and structural features in the input data, and optimize the information flow of the network; a kernel extreme learning machine based on a variational sparse matrix is ​​used for classification to automatically adjust the sparsity level, reduce the complexity of the model and the consumption of computing resources, thereby improving classification accuracy and efficiency. Specifically, it includes the following aspects:

[0034] 1. A generative adversarial network based on adaptive morphological transformation is used for data augmentation. The generator adjusts the image details through the adaptive morphological transformation layer, and the discriminator is used to distinguish between real images and generated images, thereby improving the quality and diversity of the data.

[0035] 2. Adopt a neural network algorithm based on adaptive biological impact optimization for feature extraction, simulating the ability of organisms to adjust strategies according to resource changes in the environment, avoiding the problems of gradient vanishing, gradient explosion, and falling into local optimal solutions.

[0036] 3. An autoencoder neural network based on feature adaptive refinement is used to perform feature dimensionality reduction. The encoder converts high-dimensional features into low-dimensional representations, the decoder reconstructs them back to high-dimensional space, and the dynamic weight adjustment layer adjusts the weights to optimize the reconstruction error.

[0037] 4. A kernel extreme learning machine based on a variational sparse matrix is ​​used for classification. The construction method of the kernel matrix is ​​optimized by variational sparsity, and the sparsity level is automatically adjusted by combining quantum state information to reduce the complexity of the model and the consumption of computing resources.

[0038] The embodiments of the present invention have the following technical effects:

[0039] 1. Data augmentation improves the quantity and quality of training samples, and enhances the generalization ability and recognition accuracy of the model.

[0040] 2. The problem of gradient vanishing and gradient exploding is avoided in the feature extraction process, which improves the stability and performance of the model.

[0041] 3. The feature dimensionality reduction model can more finely capture and reconstruct the intrinsic correlation and structural characteristics of the input data, optimize the information flow of the network, and reduce the feature dimension.

[0042] 4. The classifier model automatically adjusts the sparsity level, maintaining the classification performance while greatly reducing the complexity of the model and improving the classification accuracy and efficiency.

[0043] The present invention is exemplarily described below with reference to the accompanying drawings:

[0044] Reference Figure 1 , shows a flowchart of a method for human posture recognition provided by some embodiments of the present invention, which may specifically include the following steps:

[0045] Step 101, obtaining a human body posture image captured by a camera.

[0046] In some embodiments of the present invention, after acquiring the human body posture image captured by the camera, the method further includes: performing standardization processing on the human body posture image.

[0047] In practical applications, data is collected through high-resolution cameras installed at different viewing angles to capture multi-dimensional human posture images. The collected image data is cropped to obtain image data in a unified format (and standardized) and stored in jpg format.

[0048] Step 102: Using a feature extraction model, extract features from the human body posture image to obtain image features, and using a feature dimension reduction model, perform feature dimension reduction on the image features to obtain reduced dimension features.

[0049] Step 103: input the dimension reduction feature into a classifier model to obtain a classification result output by the classifier model, and determine a human posture recognition result of the human posture image according to the classification result.

[0050] In some examples, the classification results include any of the following: standing, sitting, bending over, lying down.

[0051] In practical applications, the trained model is used to process new samples for human posture recognition. Specifically, the collected human posture images are input into the trained feature extraction and feature dimension reduction model for feature processing. Furthermore, the processed features are input into the classifier model for classifier training to obtain classification results, such as standing, sitting, bending over, and lying down.

[0052] In an embodiment of the present invention, a human body posture image captured by a camera is acquired, a feature extraction model is used to perform feature extraction on the human body posture image to obtain image features, and a feature dimension reduction model is used to perform feature dimension reduction on the image features to obtain reduced dimension features, the reduced dimension features are input into a classifier model to obtain a classification result output by the classifier model, and a human body posture recognition result of the human body posture image is determined based on the classification result, thereby achieving optimization of human body posture recognition through feature extraction models, feature dimension reduction models and classifier models, improving the accuracy and efficiency of human body posture recognition, and expanding the application scenarios of posture recognition technology.

[0053] In some embodiments of the present invention, the following steps are also included:

[0054] Obtain a training data set, and use a data expansion model to expand the training data set; wherein the training data set includes multiple training data and a labeling result of each training data;

[0055] According to the expanded training data set, a preset feature extraction model is trained to obtain a training data set after feature extraction, and according to the training data set after feature extraction, a preset feature dimension reduction model is trained to obtain a training data set after feature dimension reduction;

[0056] The preset classifier model is trained based on the training data set after feature dimension reduction.

[0057] For the data augmentation model:

[0058] In some embodiments of the present invention, the data augmentation model is provided with a generative adversarial network, the generative adversarial network is provided with a generator, the generator is provided with an adaptive morphological transformation layer, and the data augmentation model is augmented through the adaptive morphological transformation layer.

[0059] In some embodiments of the present invention, the generative adversarial network is further provided with a discriminator, and the generative adversarial network is trained in the following manner: initializing the parameters of the generator and the discriminator; generating an image through the generator according to the current parameters and the input noise data; distinguishing the generated image from the real image through the discriminator to determine the image loss; and updating the parameters of the generator and the discriminator according to the image loss.

[0060] In some embodiments of the present invention, before the data expansion model is used to expand the training data set, the method further includes: standardizing the training data in the training data set.

[0061] In practical applications, data for human posture recognition is collected by high-resolution camera equipment, which is installed at different viewing angles to capture multi-dimensional human posture images. The collected image data is cropped to obtain image data in a unified format and stored in jpg format.

[0062] Furthermore, the collected image data is annotated by manual annotation. In one embodiment, the annotated categories include: standing, sitting, bending, and lying, a total of 4 categories.

[0063] In an embodiment of the present invention, the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization ability of the model and affect the accuracy of the model. The present invention uses a generative adversarial network based on adaptive morphological transformation to generate samples, thereby achieving data expansion.

[0064] The generative adversarial network based on adaptive morphological transformation includes two parts: a generator and a discriminator. The generator is responsible for generating image data that is realistic enough to deceive the discriminator, while the discriminator attempts to distinguish between real images and generated images. The present invention adopts an adaptive morphological transformation layer in the generator, which can automatically adjust the morphological operation according to the characteristics of the input image to optimize the quality and diversity of the generated image.

[0065] Specifically, the training process of the generative adversarial network algorithm based on adaptive morphological transformation is as follows:

[0066] S201, initialize the parameters of the generator and the discriminator. In one embodiment, the generator G c and the discriminator D c The parameter initialization method is expressed as:

[0067]

[0068]

[0069] In the formula, Represents the parameters of the generator; ~ indicates that it obeys a specific distribution; represents the parameters of the discriminator; represents a normal distribution with mean 0 and variance 0.02, used for the initial distribution of parameters.

[0070] S202, start multiple rounds of training cycles, the generator generates images according to the current parameters and input noise data, specifically, the generator G c According to the input noise z c and current parameters Generate image It is expressed as:

[0071]

[0072] In the formula, z c represents the random noise of the input; G c () represents a generator function; Represents the parameters of the generator; Represents the generated image features.

[0073] Furthermore, an adaptive morphological transformation is used on the generated image to adjust the tiny structures in the image to better mimic the details of the real image. Specifically, the operation of the adaptive morphological transformation layer in the generator is defined as:

[0074]

[0075] Where, X c Represents the characteristics of the input image; M c () represents the adaptive morphological transformation function; B c (X c ) and B′ c (X c ) represent the image features X c Adaptively selected dilation and erosion structuring elements; ⊕ and Represent the morphological dilation and erosion operations respectively.

[0076] In one embodiment, according to the feature X of the image c The adaptively selected dilation and erosion structural elements are adaptively adjusted according to the input features, and the adjustment method is expressed as:

[0077]

[0078]

[0079] In the formula, and Represents the convolution kernel used for dilation and erosion; and Represents the corresponding bias term; * represents the convolution operation; Sig() represents the Sigmoid activation function.

[0080] Furthermore, the discriminator tries to distinguish between real images and generated images, and the generator tries to generate images that are increasingly difficult to distinguish in order to deceive the discriminator. The adversarial process is constrained by adversarial loss, which is calculated as follows:

[0081]

[0082] Where, L adv () represents the adversarial loss function; D c (X c ) represents the discriminator's response to the real image X c The judgment result of c (G c (z c )) indicates that the discriminator generates images The judgment result of Indicates the expected operation; p data represents the distribution of real data; ~ represents compliance with a specific distribution; p z () represents the distribution of noise; p data () represents the distribution of data.

[0083] Furthermore, back propagation is used to update the parameters of the generator and discriminator, and the update method is expressed as:

[0084]

[0085]

[0086] Where ← represents the parameter update operation; η c represents the learning rate; c is a weight parameter that controls the impact of morphological consistency loss; and Respectively express and The gradient of L morph represents the morphological consistency loss; ∈ jh is a small constant; represents the second-order momentum estimate for the generator parameters. Preferably, η c Set to 0.01, λ c Set to 0.3, ∈ jh Set to 0.001.

[0087] In one embodiment, the second-order momentum estimate The update method is expressed as:

[0088]

[0089] In the formula, β c represents the momentum decay rate. Preferably, β c Set to 0.95.

[0090] Furthermore, in order to ensure that the generated image is similar to the real image in morphological features, the morphological consistency loss is used to optimize the key parts of the generated image. The calculation method of the morphological consistency loss is defined as:

[0091]

[0092] Where, L morph represents morphological consistency loss; ∥∥1 represents L1 norm; Λ c Represents the weight matrix, which assigns higher weights to the key posture areas in the human image.

[0093] S203, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed.

[0094] In one 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.

[0095] For feature extraction models:

[0096] In some embodiments of the present invention, the feature extraction model is provided with a first neural network, and the feature extraction model performs feature extraction through the first neural network.

[0097] In some examples, the first neural network is trained in the following manner: initializing parameters of the first neural network; evaluating the performance of current parameters in a specific training environment; updating current parameters according to results of resource competition; adjusting sensitivity parameters and learning rate parameters according to results of environmental perception evaluation; and dynamically adjusting hyperparameters that control resource allocation sensitivity between global search and local search.

[0098] The expanded data is input into the feature extraction model to train the feature extraction model. The present invention adopts a 3-layer fully connected neural network for feature extraction.

[0099] In the related art, some solutions use neural networks for feature extraction. In some neural network structures, they may encounter problems such as gradient vanishing, gradient exploding, or falling into local optimal solutions, which will affect the stability of training and the performance of the model. The present invention adopts a neural network algorithm based on adaptive biological impact optimization as a feature extraction model, which is inspired by the adaptive behavior of organisms in nature, especially the ability of organisms to adjust strategies according to resource changes in the environment. The parameters of the neural network are optimized by simulating the survival competition and resource adaptation process of organisms, thereby avoiding the problems of gradient vanishing, gradient exploding, and falling into local optimal solutions, and improving the feature extraction capability.

[0100] like Figure 2a , the training process of the neural network algorithm based on adaptive biological influence optimization (i.e., the first neural network) is as follows:

[0101] S301, initialize the parameters of the neural network, define the weights and biases of the neural network, in one embodiment, the initial weights and initial bias The initialization method is expressed as:

[0102]

[0103]

[0104] In the formula, ~ means obeying a specific distribution, W NN and b NN are the weights and biases of the neural network, σ sn is the initialization standard deviation, and are the initial weights and initial biases of the neural network, respectively. Preferably, σ sn Set to 0.01, is a normal distribution.

[0105] S302: At the beginning of each iteration, evaluate the performance of the current network parameters under a specific training environment. In one embodiment, calculate the loss L under the current parameter configuration. NN (W NN ,b NN ), and estimate the environmental fitness as the basis for subsequent adjustments. The calculation method is expressed as:

[0106] E NN =exp(-β nn ·(W nN ,b NN ))

[0107] In the formula, E NN is the environmental fitness, β NN is the sensitivity parameter, L nn() is the loss function of the neural network.

[0108] And, in this embodiment, the loss L nn (W nn ,b nn ) is calculated using the cross entropy loss method and is expressed as:

[0109]

[0110] In the formula, C represents the number of categories, y c is the one-hot encoding of the true label, is the probability predicted by the model.

[0111] Furthermore, the probability predicted by the model is calculated by the softmax function and is expressed as:

[0112]

[0113] In the formula, z c is the logistic regression probability output by the network.

[0114] In the resource competition behavior in nature, each parameter is regarded as an individual, competing for resources in the data set. The resource allocation is determined by the contribution of the parameter. The resource allocation of each parameter is calculated as follows:

[0115]

[0116] In the formula, R ij The resources allocated for the parameters of the jth neuron in the i-th neural network layer, is the weight parameter of the jth neuron in the i-th neural network layer, is the weight parameter of the lth neuron in the kth neural network layer, k and l are both indices, γ nn is a hyperparameter that controls the sensitivity of resource allocation, is the loss function L nN About parameters The partial derivative of is the loss function L NN About parameters The partial derivative of .

[0117] S303: Update weight W according to the result of resource competition NN and bias b NN , the update method is expressed as:

[0118]

[0119]

[0120] In the formula, R iThe resources allocated for the parameters of the i-th neural network layer, is the bias of the neurons in the i-th layer, is the weight parameter update of the lth neuron in the kth neural network layer, α NN is the learning rate, is the update amount of the bias parameters of the neurons in the i-th layer.

[0121] In one embodiment, the present invention uses the momentum term m ij To smooth the learning process of weight parameters, it can be expressed as:

[0122]

[0123]

[0124] In the formula, is the momentum term of the jth neuron in the ith neural network layer at the t-1th iteration, is the momentum term of the jth neuron in the i-th neural network layer at the t-th iteration, is the weight parameter of the jth neuron in the i-th neural network layer at the t-1th iteration, is the weight parameter of the jth neuron in the i-th neural network layer at the t-th iteration, β xm is the momentum decay parameter. Preferably, β xm Set to 0.95.

[0125] S304: Adjust the sensitivity parameter and the learning rate parameter according to the result of the environmental perception evaluation. The adjustment method is expressed as follows:

[0126]

[0127]

[0128] In the formula, ΔE Nn is the change in environmental fitness, that is, the difference between the environmental fitness of this iteration and the previous iteration; η NN and κ NN is the adjustment factor, is the updated learning rate parameter, is the updated sensitivity parameter. Preferably, η NN and κ NN Both are set to 0.1.

[0129] S305, dynamically adjusting the hyperparameters controlling the resource allocation sensitivity between the global search and the local search to avoid falling into the local optimal solution, the adjustment method is expressed as:

[0130]

[0131] In the formula, || is the absolute value symbol, is the hyperparameter of the adjusted control resource allocation sensitivity, and ρ is a balanced hyperparameter. Preferably, ρ NN Set to 0.1.

[0132] S306, repeat the above steps until the preset stop iteration condition is met, which means 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.

[0133] For feature dimensionality reduction models:

[0134] In some embodiments of the present invention, the feature dimensionality reduction model is provided with an autoencoder neural network, and the feature dimensionality reduction model performs feature dimensionality reduction through the autoencoder neural network.

[0135] The autoencoding neural network is provided with an encoder and a decoder, wherein the encoder is used to convert the features extracted from the high-dimensional input into a low-dimensional hidden representation, and the decoder is used to reconstruct the hidden representation back to the original high-dimensional space.

[0136] In some examples, the autoencoding neural network is trained in the following manner: initializing the parameters of the encoder; converting the input feature-extracted data into a low-dimensional representation through the encoder; adjusting the weights in the hidden layer of the encoder through a dynamic weight adjustment layer; and updating the parameters of the encoder and the decoder using the gradient descent method.

[0137] The extracted features are input into a feature dimension reduction model to train the feature dimension reduction model. The present invention uses an autoencoding neural network algorithm based on feature adaptive refinement to perform feature dimension reduction.

[0138] The autoencoding neural network algorithm includes two parts: an encoder and a decoder.

[0139] The encoder converts the extracted features of the high-dimensional input into a low-dimensional hidden representation, and the decoder reconstructs this hidden representation back to the original high-dimensional space to form a reconstructed output. The present invention adopts a feature coupling mechanism so that the autoencoder can not only reduce the dimension of the data, but also more finely capture and reconstruct the inherent correlation and structural features in the input data, thereby optimizing the information flow of the network.

[0140] like Figure 2b , the training process of the autoencoder neural network algorithm based on feature adaptive refinement is as follows:

[0141] S401, initializing the network parameters of the autoencoder, including the weight and bias of the autoencoder. In one embodiment, the initialization method is expressed as:

[0142]

[0143]

[0144] Where W r is the weight parameter of the autoencoder, b r is the bias parameter of the autoencoder, is the initial weight of the autoencoder; is the initial bias of the autoencoder; It means the mean is 0 and the variance is Normal distribution of is the number of neurons in the input layer of the autoencoder.

[0145] S402, input feature extracted data X r Propagate through the sequence of encoder layers and transform into a low-dimensional representation H r , expressed as:

[0146]

[0147]

[0148] Where Re() is the ReLU activation function; H r is the low-dimensional representation of the encoder output; and are the weight and bias of the encoder respectively; and are the weights and biases of the decoder; Reconstructed output.

[0149] S403. In the hidden layer of the autoencoder, the weight is adjusted by the dynamic weight adjustment layer to achieve the purpose of adaptive feature refinement. The calculation method of the weight update amount of the autoencoder is expressed as:

[0150]

[0151] In the formula, ΔW r is the weight update amount; tanh() is the hyperbolic tangent function; λ r is the learning rate; r is the adjustment factor; is the loss function L r About weight W r The gradient of L r is the reconstruction error. Preferably, λ r Set to 0.01, γr Set to 3.14.

[0152] Furthermore, the gradient Solve the loss function L by chain rule r The gradient with respect to the weight is calculated and expressed as:

[0153]

[0154] In the formula, is the partial derivative of the loss function with respect to the reconstructed output.

[0155] In one embodiment, the calculation method of the mean square error of the loss function is expressed as:

[0156]

[0157] In the formula, ∥∥ is the L2 norm, α r is the regularization coefficient, is the L1 norm of the encoder weights. Preferably, α r Set to 0.3.

[0158] S404: Update the parameters of the encoder and decoder using the gradient descent method. The updating method is expressed as:

[0159]

[0160]

[0161] Where η r is the learning rate; is the weight parameter of the autoencoder at the t+1th iteration; is the weight parameter of the autoencoder at the tth iteration; is the bias parameter of the autoencoder at the t+1th iteration; is the bias parameter of the autoencoder at the tth iteration; t represents the number of iterations; and are the gradients of the loss function with respect to weights and biases, respectively.

[0162] In one embodiment, the learning rate η r It is set by adaptive adjustment to dynamically adjust according to different stages of network training, expressed as:

[0163]

[0164] In the formula, is the learning rate for the t+1th iteration, is the learning rate of the tth iteration, δ r is the sensitivity parameter of the learning rate adjustment. Preferably, δr Set to 0.5.

[0165] S405, repeat the above steps until the preset stop iteration condition is met, which means 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.

[0166] For the classifier model:

[0167] In some embodiments of the present invention, the classifier model is provided with a kernel extreme learning machine, and the classifier model is classified by the kernel extreme learning machine, and the kernel extreme learning machine is trained in the following manner: initializing the parameters of the kernel extreme learning machine; constructing the kernel matrix in the kernel extreme learning machine according to the input data; optimizing the hidden layer neuron state in the kernel extreme learning machine to obtain the weight of the output layer.

[0168] The reduced-dimensional data is input into the classifier to train the classifier model. The present invention adopts a variational sparse matrix-based kernel extreme learning machine as a classification algorithm. The kernel extreme learning machine is based on a single-layer feedforward neural network, uses kernel techniques to handle nonlinear separable problems, and uses extreme learning to perform rapid training. The present invention uses a variational sparsity optimization kernel matrix construction method to enable the kernel matrix to automatically adjust its sparsity level while maintaining classification performance, greatly reducing the complexity of the model and computing resource consumption, and improving classification accuracy.

[0169] Specifically, the training process of the kernel extreme learning machine algorithm based on variational sparse matrix is ​​as follows:

[0170] S501: Initialize the parameters of the kernel extreme learning machine, including the weight and bias of the kernel extreme learning machine. In one embodiment, the weight W u and bias b u The initialization method is expressed as:

[0171]

[0172] b u =0

[0173] Where W u is the weight matrix of the kernel extreme learning machine; b u is the bias vector, initialized to zero; n u Indicates the number of input layer nodes; m u Indicates the number of nodes in the output layer; randn(n u ,m u ) Generate n u ×m u A standard normally distributed random number.

[0174] S502: construct a kernel matrix according to the input data. The kernel matrix K u It is constructed using Gaussian kernel-based variational sparse combined with quantum state information, expressed as:

[0175]

[0176] In the formula, K u is the kernel matrix; u is the bandwidth parameter of the kernel function; X u is the input data matrix; is the transpose of the input data matrix; Q u is the quantum state matrix; K u (i,j) represents the element in the i-th row and j-th column of the kernel matrix; d is the feature dimension of the input data; X u,i,k and X u,j,k Represents the input data matrix X u The kth feature of the i-th and j-th samples in .

[0177] Furthermore, we define the data change metric V u is the trace of the covariance matrix of the input data, indicating the breadth of the data distribution, expressed as:

[0178] V u =trace(cov(X u ))

[0179] In the formula, cov(X u ) represents the input data X u The covariance matrix of ; trace() represents the trace of the matrix; V u Represents a measure of data change.

[0180] Furthermore, according to the data change metric V u Dynamically adjust kernel bandwidth u , expressed as:

[0181]

[0182] In the formula, According to V u The square of the calculated standard deviation; ∈ jy is a small constant that prevents division by zero. Preferably, ∈ jy Set to 0.001.

[0183] In one embodiment, the quantum state matrix Q u It is used to enhance the data representation capability of the kernel matrix. The calculation method is expressed as:

[0184] Q u =Su CNOT(H u (S u ))

[0185] In the formula, S u is the initial setting of the quantum state; CNOT() is the controlled NOT gate operation in quantum computing, which is used to generate entangled states; H u () is the Hadamard gate, which is used to create a superposition state on a qubit; H u (S u ) indicates that in S u The result after using the Hadamard gate is shown above. Preferably, the initial setting of the quantum state S u Select the uniform superposition state.

[0186] S503, optimize the hidden layer neuron state, use the rapid learning ability of the extreme learning machine, quickly train the classifier, and output the weight β u The calculation of is achieved by minimizing the regularized risk function, expressed as:

[0187]

[0188] In the formula, β u is the weight of the output layer; u is the regularization parameter; I is the identity matrix; Y u is the target output matrix; represents the inverse of the kernel matrix under the regularization term. Preferably, λ u Set to 2.

[0189] Furthermore, the calculation process of output weights is solved by Cholesky decomposition optimization, which is expressed as:

[0190]

[0191] L u =cholesky(Ω u )

[0192]

[0193] In the formula, Ω u is the regularized kernel matrix; L u is the lower triangular matrix obtained by Cholesky decomposition; Indicates L u The inverse matrix of the transpose of ; Indicates L u The inverse matrix of .

[0194] S504, after the weight calculation of the output layer is completed, it means that the training of the kernel extreme learning machine model based on the variational sparse matrix is ​​completed.

[0195] like Figure 3a , Figure 3b , Figure 3c The specific process includes: S1, data collection and annotation; S2, data expansion; S3, feature extraction model training; S4, feature dimension reduction model training; S5, classifier model training; S6, human posture recognition. In the model training process, data expansion, feature extraction model training, feature dimension reduction model training, and classifier model training are performed. Then, in the process of using the model, the trained model is used to process new samples (i.e., human posture images) to obtain classification results.

[0196] It should be noted that, for the sake of simplicity, the method embodiments are described as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because according to the embodiments of the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0197] Some embodiments of the present invention also provide a device for human body posture recognition, which may specifically include:

[0198] The human body posture image acquisition module is used to obtain the human body posture image collected by the camera.

[0199] The feature processing module is used to use a feature extraction model to perform feature extraction on the human posture image to obtain image features, and use a feature dimension reduction model to perform feature dimension reduction on the image features to obtain reduced dimension features.

[0200] The human posture recognition module is used to input the dimension reduction feature into a classifier model to obtain a classification result output by the classifier model, and determine a human posture recognition result of the human posture image according to the classification result.

[0201] Optionally, after acquiring the human body posture image collected by the camera, the method further includes:

[0202] The human body posture image is standardized.

[0203] Optionally, the classification result includes any one of the following: standing, sitting, bending over, lying down.

[0204] Optionally, before acquiring the human body posture image collected by the camera, the method further includes:

[0205] Obtain a training data set, and use a data expansion model to expand the training data set; wherein the training data set includes multiple training data and a labeling result of each training data;

[0206] According to the expanded training data set, a preset feature extraction model is trained to obtain a training data set after feature extraction, and according to the training data set after feature extraction, a preset feature dimension reduction model is trained to obtain a training data set after feature dimension reduction;

[0207] The preset classifier model is trained based on the training data set after feature dimension reduction.

[0208] Optionally, before the data expansion model is used to expand the training data set, the method further includes:

[0209] The training data in the training data set is standardized.

[0210] Optionally, the data augmentation model is provided with a generative adversarial network, the generative adversarial network is provided with a generator, the generator is provided with an adaptive morphological transformation layer, and the data augmentation model is augmented through the adaptive morphological transformation layer.

[0211] Optionally, the generative adversarial network is further provided with a discriminator, and the generative adversarial network is trained in the following manner:

[0212] Initialize the parameters of the generator and the discriminator;

[0213] Generate an image according to the current parameters and input noise data through the generator;

[0214] Using the discriminator, distinguishing the generated image from the real image to determine the image loss;

[0215] According to the image loss, parameters of the generator and the discriminator are updated.

[0216] Optionally, the feature extraction model is provided with a first neural network, and the feature extraction model performs feature extraction through the first neural network, and the first neural network is trained in the following manner:

[0217] Initializing parameters of the first neural network;

[0218] Evaluate the performance of current parameters in a specific training environment;

[0219] Update current parameters according to the results of resource competition;

[0220] Adjust the sensitivity parameter and learning rate parameter according to the results of the environmental perception evaluation;

[0221] The hyperparameters that control the sensitivity of resource allocation are dynamically adjusted between global and local search.

[0222] Optionally, the feature dimension reduction model is provided with an autoencoding neural network, and the feature dimension reduction model performs feature dimension reduction through the autoencoding neural network;

[0223] The autoencoding neural network is provided with an encoder and a decoder, wherein the encoder is used to convert the features extracted from the high-dimensional input into a low-dimensional hidden representation, and the decoder is used to reconstruct the hidden representation back to the original high-dimensional space.

[0224] Optionally, the autoencoding neural network is trained in the following manner:

[0225] Initializing parameters of the encoder;

[0226] The input feature-extracted data is converted into a low-dimensional representation through an encoder;

[0227] In the hidden layer of the encoder, weights are adjusted by a dynamic weight adjustment layer;

[0228] The parameters of the encoder and the decoder are updated using the gradient descent method.

[0229] Optionally, the classifier model is provided with a kernel extreme learning machine, and the classifier model is classified by the kernel extreme learning machine, and the kernel extreme learning machine is trained in the following manner:

[0230] Initializing parameters of the kernel extreme learning machine;

[0231] Constructing a kernel matrix in the kernel extreme learning machine according to input data;

[0232] The hidden layer neuron states in the kernel extreme learning machine are optimized to obtain the weights of the output layer.

[0233] Some embodiments of the present invention further provide an electronic device, comprising a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the above method is implemented when the computer program is executed by the processor.

[0234] Some embodiments of the present invention further provide a computer-readable storage medium, on which a computer program is stored, and the computer program implements the above method when executed by a processor.

[0235] Some embodiments of the present invention further provide a computer program product, including a computer program, which implements the above method when executed by a processor.

[0236] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0237] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0238] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0239] Those skilled in the art will appreciate that the embodiments of the present invention may be provided as methods, devices, or computer program products. Therefore, the embodiments of the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the embodiments of the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0240] The embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0241] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0242] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0243] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present invention.

[0244] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the above elements.

[0245] The above is a detailed introduction to the method, device, equipment and medium for human posture recognition provided. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for human body posture recognition, characterized in that: The method comprises: Obtain human body posture images collected by the camera; Using a feature extraction model to extract features from the human body posture image to obtain image features, and using a feature dimension reduction model to reduce the dimensions of the image features to obtain reduced dimension features; The dimension reduction feature is input into a classifier model to obtain a classification result output by the classifier model, and a human body posture recognition result of the human body posture image is determined based on the classification result.

2. The method according to claim 1, characterized in that After obtaining the human body posture image collected by the camera, the method further includes: The human body posture image is standardized.

3. The method according to claim 1, characterized in that The classification result includes any of the following: standing, sitting, bending over, and lying down.

4. The method according to any one of claims 1 to 3, characterized in that: Before obtaining the human body posture image collected by the camera, the method further includes: Obtain a training data set, and use a data expansion model to expand the training data set; wherein the training data set includes multiple training data and a labeling result of each training data; According to the expanded training data set, a preset feature extraction model is trained to obtain a training data set after feature extraction, and according to the training data set after feature extraction, a preset feature dimension reduction model is trained to obtain a training data set after feature dimension reduction; The preset classifier model is trained based on the training data set after feature dimension reduction.

5. The method according to claim 4, characterized in that Before the data expansion model is adopted to expand the training data set, the method further includes: The training data in the training data set is standardized.

6. The method according to claim 4, characterized in that The data expansion model is provided with a generative adversarial network, the generative adversarial network is provided with a generator, the generator is provided with an adaptive morphological transformation layer, and the data expansion model is expanded through the adaptive morphological transformation layer.

7. The method according to claim 6, characterized in that The generative adversarial network is also provided with a discriminator, and the generative adversarial network is trained in the following manner: Initialize the parameters of the generator and the discriminator; Generate an image according to the current parameters and input noise data through the generator; Using the discriminator, distinguishing the generated image from the real image to determine the image loss; According to the image loss, parameters of the generator and the discriminator are updated.

8. The method according to claim 4, characterized in that The feature extraction model is provided with a first neural network, and the feature extraction model performs feature extraction through the first neural network, and the first neural network is trained in the following manner: Initializing parameters of the first neural network; Evaluate the performance of current parameters in a specific training environment; Update current parameters according to the results of resource competition; Adjust the sensitivity parameter and learning rate parameter according to the results of the environmental perception evaluation; The hyperparameters that control the sensitivity of resource allocation are dynamically adjusted between global and local search.

9. The method according to claim 4, characterized in that The feature dimension reduction model is provided with an autoencoder neural network, and the feature dimension reduction model performs feature dimension reduction through the autoencoder neural network; The autoencoding neural network is provided with an encoder and a decoder, wherein the encoder is used to convert the features extracted from the high-dimensional input into a low-dimensional hidden representation, and the decoder is used to reconstruct the hidden representation back to the original high-dimensional space.

10. The method according to claim 9, characterized in that The autoencoder neural network is trained in the following way: Initializing parameters of the encoder; The input feature-extracted data is converted into a low-dimensional representation through an encoder; In the hidden layer of the encoder, weights are adjusted by a dynamic weight adjustment layer; The parameters of the encoder and the decoder are updated using the gradient descent method.

11. The method according to claim 4, characterized in that The classifier model is provided with a kernel extreme learning machine, and the classifier model is classified by the kernel extreme learning machine, and the kernel extreme learning machine is trained in the following manner: Initializing parameters of the kernel extreme learning machine; Constructing a kernel matrix in the kernel extreme learning machine according to input data; The hidden layer neuron states in the kernel extreme learning machine are optimized to obtain the weights of the output layer.

12. A device for human body posture recognition, characterized in that: The device comprises: A human body posture image acquisition module is used to acquire human body posture images collected by a camera; A feature processing module is used to use a feature extraction model to perform feature extraction on the human posture image to obtain image features, and use a feature dimension reduction model to perform feature dimension reduction on the image features to obtain reduced dimension features; The human posture recognition module is used to input the dimension reduction feature into a classifier model to obtain a classification result output by the classifier model, and determine a human posture recognition result of the human posture image according to the classification result.

13. An electronic device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program implements the method according to any one of claims 1 to 11 when executed by the processor.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 11 is implemented.

Images