Identity recognition method based on intelligent safety helmet

Abstract

The invention relates to the technical field of identity recognition, and discloses an identity recognition method based on an intelligent safety helmet, which comprises the following steps of: constructing a feature vector by combining built-in sensor data, a facial image and behavior mode data, and establishing and converting the feature vector into a probability calculation model; a neural network algorithm is used for solving the model, an initial recognition result is based on safety helmet distribution and work arrangement, path results higher than a threshold value are screened out by cyclically calculating an optimal path and a discriminant value of a feature association graph, iteration is conducted till a solution is stable, and finally accurate identity recognition is achieved. According to the invention, the identification accuracy and efficiency can be improved, and the safety of operators and the reliability of identity verification are ensured.

Description

Technical Field

[0001] The invention relates to the technical field of identity recognition, and in particular to an identity recognition method based on a smart helmet. Background Art

[0002] Existing identity recognition technologies mainly rely on biometrics, smart cards, RFID technology, etc. Biometrics recognition technologies such as fingerprint recognition, iris recognition, and facial recognition are widely used because of their uniqueness and difficulty in duplication. However, these technologies may be affected by factors such as light and angle in complex environments, resulting in reduced recognition accuracy. Smart cards and RFID technologies use electronic tags and readers to authenticate identities. Although they are easy to operate, they are prone to copying and theft. In high-risk working environments such as industry and construction, ensuring the accuracy of the identity of operators is crucial to ensuring personnel safety and improving work efficiency. However, existing technologies often find it difficult to provide stable and reliable recognition results in dynamic and changing working environments.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the prior art: the existing identity recognition methods are not adaptable enough in complex industrial environments, and it is difficult to take into account both the accuracy and real-time performance of recognition; in addition, for scenarios where multiple people work at the same time, the existing technology is often difficult to efficiently process and distinguish individual identities, resulting in low recognition efficiency; furthermore, the existing identity recognition system lacks an effective fusion and processing mechanism when faced with multi-source information such as sensor data, facial images, and behavioral patterns, and is unable to fully utilize this information to improve recognition accuracy and robustness. Summary of the invention

[0004] The embodiment of the present invention aims to provide an identity recognition method based on a smart helmet to solve the technical problems raised in the prior art.

[0005] The embodiment of the present invention solves the technical problem by adopting the following technical solution:

[0006] In a first aspect, a method for identifying an identity based on a smart helmet is provided, comprising:

[0007] Based on the built-in sensor data of the helmet, the facial image data of the personnel and the behavior pattern data, a feature vector is constructed, an identity recognition model is established, and then converted into a probability calculation model;

[0008] Based on the personnel activity set corresponding to each helmet, a corresponding feature association graph is constructed according to the discriminant value of the probability calculation model;

[0009] A deep learning algorithm with a neural network algorithm as the core is used to solve the probability calculation model, including: based on the helmet allocation information and the personnel work arrangement information, an initial identity recognition result corresponding to each helmet is obtained as an initial recognition result set, the solution of the restricted main problem of the probability calculation model is solved, the weight parameter is calculated, the optimal path and the discrimination value of the feature association graph of each helmet are cyclically obtained, and the recognition result corresponding to the optimal path with a discrimination value greater than a preset threshold is added to the recognition result set to continue to solve the solution of the restricted main problem;

[0010] Through deep learning algorithms and continuous iterations, all solutions to the restricted master problem are stable solutions;

[0011] Based on the stable solution and the final set of recognition results, a helmet identification solution is obtained.

[0012] Furthermore, the built-in sensors of the helmet include a temperature sensor and an acceleration sensor;

[0013] The helmet built-in sensor data is obtained according to the weighted sum of the sensor measurement values ​​of each helmet, wherein: when the sensor is a temperature sensor, the weight of its measurement value is α;

[0014] When the sensor is an acceleration sensor, the weight of its measurement value is β, where α and β are preset non-negative weight values.

[0015] Furthermore, the facial image data of the person is constructed based on key facial feature points and texture features.

[0016] Furthermore, the step of acquiring the facial key feature points and texture features includes:

[0017] Pre-process the image captured by the built-in camera of the helmet, obtain the facial area in the image, and obtain the key facial feature points and texture features through the feature extraction algorithm;

[0018] Taking the feature range of the standard facial image as the basis of the statistical range, taking the preset feature distance as the frequency band, counting the feature matching probability in each frequency band within the statistical range, and obtaining the facial feature distribution;

[0019] According to the feature changes of the facial images of the same person at different times, the facial feature change range is obtained, and based on the facial feature change range and facial feature distribution, the facial key feature points and texture features used for identification are obtained.

[0020] Furthermore, the calculation formula for obtaining the facial key feature points and texture features for identification based on the facial feature variation range and facial feature distribution is as follows:

[0021]

[0022] Among them, p i is the initial eigenvalue of facial feature point i, p i′ is the feature value of facial feature point i′ after the change, in pixels, p i ∈[P MIN ,P MAX ], P MIN is the minimum eigenvalue of facial feature point i, P MAX is the maximum eigenvalue of facial feature point i, ρ(p i ) is the facial feature point i with feature value p i The probability, ρ(p i′ ) is the facial feature point i′ with feature value p i′ The probability of ii′ is the relative distance between facial feature points i and i′ of the same person, and F is the comprehensive facial feature value used for identification.

[0023] Furthermore, the behavior pattern data is obtained by summing the personnel activity trajectory and action frequency data corresponding to each helmet, wherein: when the personnel action frequency is higher than a preset frequency, the weight of the behavior pattern data corresponding to the personnel activity is γ, otherwise it is δ, and γ and δ are preset non-negative weight values.

[0024] Furthermore, the feature vector is constructed based on the helmet built-in sensor data, the personnel facial image data and the behavior pattern data to establish the identity recognition model, including:

[0025] The weighted sum of the helmet built-in sensor data, the personnel facial image data and the behavior pattern data is used as the total eigenvalue, and maximizing the correct recognition probability is used as the objective function;

[0026] The first constraint is that each person has a corresponding helmet;

[0027] The second constraint condition is that each helmet can only be assigned to one recognition result or not be assigned to any recognition result;

[0028] The third constraint condition is that the number of people identified at the same time is not greater than the number that the system can handle;

[0029] Whether the identification result is confirmed is used as the decision variable constraint.

[0030] Furthermore, the personnel activity set corresponding to each helmet is constructed based on the discriminant value of the probability calculation model, including:

[0031] For each helmet, arrange the activities in the corresponding personnel activity set in chronological order, with each activity as a node;

[0032] An initial feature matrix is ​​set, whose element values ​​are calculated based on the values ​​of the helmet built-in sensor data, the personnel facial image data and the behavior pattern data at the activity moment;

[0033] Calculate the feature difference between every two adjacent active nodes. This difference takes into account the changes in sensor data, facial features, and behavior patterns. The weight of the edge is determined based on the difference and the preset mapping relationship.

[0034] A special starting node and ending node are set, and each active node is connected according to specific rules and assigned corresponding weights to construct a feature association graph.

[0035] Furthermore, the step of obtaining the optimal path and the discriminant value of the feature association graph of each helmet includes:

[0036] The dynamic programming algorithm is used to process the feature association graph. Starting from the starting node, the optimal cumulative feature value reaching each node is calculated in turn, and the predecessor node information of each node is recorded;

[0037] After traversing all nodes, trace back from the end node to the start node to get the optimal path;

[0038] The discriminant value is obtained by calculating the weight product of all edges on the optimal path, and the discriminant value indicates the credibility of the identity recognition result represented by this path.

[0039] Furthermore, the deep learning algorithm and continuous iteration until the solutions to the restricted master problem are all stable solutions include:

[0040] The discriminant value of the optimal path of the feature association graph of each helmet is input into an adaptively adjusted classifier, and the classifier adjusts itself according to the discriminant value and historical recognition results. If the discriminant value meets the current classification standard of the classifier, the recognition result corresponding to the corresponding optimal path is added to the recognition result set, and the parameters of the classifier are updated. If it does not meet the requirements, the reasons for the non-compliance are analyzed. If it is caused by factors such as data anomalies, the discriminant value and path are recalculated after the data is corrected;

[0041] If it is due to inaccurate model, adjust the classifier structure or parameters and iterate again until all solutions of the restricted main problem are stable solutions.

[0042] The above-mentioned embodiments of the present invention have at least the following beneficial effects: The present invention can significantly improve the accuracy and efficiency of identity recognition by combining the built-in sensor data of the helmet, the facial image data of the personnel and the behavior pattern data, constructing a feature vector and establishing an identity recognition model. The present method utilizes deep learning algorithms and probabilistic calculation models, which can not only process and analyze multi-source data, but also optimize the recognition results through continuous iterations to ensure stable and reliable identity verification of personnel in complex and changing working environments. In addition, the present method can effectively identify and distinguish individual identities by constructing a feature association graph and calculating the optimal path, and can maintain high efficiency and accuracy even in scenarios where multiple people are working at the same time.

[0043] The method of the present invention can effectively solve the problems existing in the prior art, such as low recognition accuracy, poor adaptability and low processing efficiency. Through the built-in sensors and cameras of the smart helmet, combined with the work arrangement information of the personnel, the behavior and facial features of the operators can be monitored and analyzed in real time, thereby realizing the rapid and accurate identification of the identity of the operators. This method can not only improve the safety of operations, but also provide a new means of personnel management and safety monitoring for industries such as industry and construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] One or more embodiments are exemplarily described by the figures in the corresponding drawings, and these exemplifications do not constitute limitations on the embodiments. Elements with the same reference numerals in the drawings represent similar elements, and unless otherwise stated, the figures in the drawings do not constitute proportional limitations.

[0045] Figure 1 A schematic diagram of a flow chart of an identity recognition method based on a smart helmet provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to facilitate the understanding of the present invention, the present invention is described in more detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that when an element is described as "connecting" another element, it can be directly on another element, or there can be one or more centered elements therebetween. The orientation or positional relationship indicated by the terms "upper", "lower", "left", "right", "upper end", "lower end", "top" and "bottom" used in this specification is based on the orientation or positional relationship shown in the accompanying drawings, only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0047] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0048] Combine the following Figure 1 , the identity recognition method 100 based on the smart helmet provided in the embodiment of the present application is described in detail through a specific embodiment.

[0049] Figure 1 : is a flow chart of an identity recognition method based on a smart helmet provided by the present invention. An identity recognition method based on a smart helmet provided by one embodiment of the present invention comprises:

[0050] Step 101, constructing a feature vector based on the helmet built-in sensor data, the personnel facial image data and the behavior pattern data, establishing an identity recognition model, and converting it into a probability calculation model;

[0051] Step 102, based on the personnel activity set corresponding to each helmet, construct a corresponding feature association graph according to the discriminant value of the probability calculation model;

[0052] Step 103, using a deep learning algorithm with a neural network algorithm as the core to solve the probability calculation model, including: based on the helmet allocation information and the personnel work arrangement information, obtaining an initial identity recognition result corresponding to each helmet as an initial recognition result set, solving the solution of the restricted main problem of the probability calculation model, calculating the weight parameter, cyclically obtaining the optimal path and the discrimination value of the feature association graph of each helmet, adding the recognition result corresponding to the optimal path with a discrimination value greater than a preset threshold to the recognition result set to continue solving the solution of the restricted main problem;

[0053] Step 104, using a deep learning algorithm and continuous iteration until all solutions to the restricted master problem are stable solutions;

[0054] Step 105: obtaining a helmet identification solution based on the stable solution and the final identification result set.

[0055] It should be noted that the present invention relates to an identity recognition method based on a smart helmet. The core of the method is to use the sensor data built into the helmet, the facial image data of the personnel and the behavior pattern data to construct a feature vector, establish an identity recognition model, and finally convert the model into a probability calculation model. Here, the feature vector refers to a collection of a series of data points that can represent the identity of an individual, and the identity recognition model refers to an algorithm or mathematical model used to distinguish different individual identities.

[0056] Specifically, the built-in sensors of the helmet include temperature sensors and acceleration sensors, and the data of these sensors are obtained based on the weighted sum of the sensor measurement values ​​of each helmet. Here, the weighted sum means that the measurement value of each sensor is assigned a weight according to its importance or reliability. For example, the weight of the temperature sensor is α, and the weight of the acceleration sensor is β. These weights are pre-set non-negative values. Facial image data is constructed based on facial key feature points and texture features. These feature points and texture features are obtained by pre-processing the images taken by the built-in camera of the helmet and then obtaining them through feature extraction algorithms.

[0057] Preferably, the acquisition of behavior pattern data can be achieved by analyzing the activity trajectory and action frequency data of each person wearing a helmet. When the action frequency of a person is higher than a preset frequency, the weight of the behavior pattern data corresponding to the person's activity is γ, otherwise it is δ, and γ and δ are also preset non-negative weight values.

[0058] Furthermore, in practical applications, these parameters can be adjusted according to specific working environments and safety requirements to achieve the best recognition effect. For example, in high temperature or high vibration environments, the weights of temperature sensors and acceleration sensors may need to be increased to improve recognition accuracy.

[0059] In some embodiments, the built-in sensor of the helmet includes a temperature sensor and an acceleration sensor;

[0060] The helmet built-in sensor data is obtained according to the weighted sum of the sensor measurement values ​​of each helmet, wherein: when the sensor is a temperature sensor, the weight of its measurement value is α;

[0061] When the sensor is an acceleration sensor, the weight of its measurement value is β, where α and β are preset non-negative weight values.

[0062] It should be noted that the specific composition of the built-in sensors of the smart helmet is described, including temperature sensors and acceleration sensors, and how to obtain the built-in sensor data of the helmet based on the weighted sum of the measured values ​​of these sensors is explained. Here, the built-in sensor refers to a device directly installed inside the helmet to collect the wearer's physiological and activity data.

[0063] Specifically, the temperature sensor is used to measure the wearer's ambient temperature or body surface temperature, while the acceleration sensor is used to detect the wearer's head movement and possible impact. The measurements of these sensors are weighted according to preset weights α and β, where α represents the weight of the temperature sensor and β represents the weight of the acceleration sensor. These weight values ​​can be set according to the accuracy and importance of the sensor and the specific requirements of the working environment. For example, when working in a high temperature environment, it may be necessary to increase the weight α of the temperature sensor to more accurately monitor the thermal stress that the wearer may face.

[0064] Preferably, in order to more accurately reflect the importance of different sensor data, the values ​​of α and β can be adjusted according to the actual working environment and historical data analysis results. For example, if it is found that the data of the acceleration sensor is more critical for identifying the wearer's identity under specific working conditions, the value of β can be appropriately increased.

[0065] Furthermore, it is also possible to consider introducing other types of sensors, such as humidity sensors or heart rate sensors, to further enrich the dimensions of the built-in sensor data of the helmet and improve the accuracy and reliability of identity recognition. These additional sensor data can also be integrated into the identity recognition model through a similar weighting mechanism.

[0066] In some embodiments, the facial image data of the person is constructed based on key facial feature points and texture features.

[0067] It should be noted that the method of constructing facial image data is further explained, and the importance of key facial feature points and texture features is particularly pointed out. Here, key facial feature points refer to some specific locations on the face, such as the corners of the eyes, the tip of the nose, the corners of the mouth, etc. These points are very important in facial recognition because they have a high degree of differentiation between different individuals. Texture features involve the texture information of the facial skin, such as wrinkles, spots, etc. These features also help to distinguish different individuals.

[0068] Specifically, the steps of acquiring facial key feature points and texture features include preprocessing the image captured by the built-in camera of the helmet, acquiring the facial area in the image, and acquiring facial key feature points and texture features through a feature extraction algorithm. Preprocessing includes steps such as denoising, contrast enhancement, and standardized lighting conditions to improve the accuracy of feature extraction. The feature extraction algorithm can be an existing computer vision algorithm, such as Haar features, LBP (local binary pattern) or a deep learning model, which can identify and extract key feature points and texture features from the image.

[0069] Preferably, in order to improve the accuracy of facial recognition, a multi-scale and multi-angle feature extraction method can be used to adapt to different lighting and posture changes. For example, feature points can be extracted at different scales to capture different levels of features from local to global.

[0070] Furthermore, 3D facial reconstruction techniques can be combined to improve the robustness of feature extraction, especially in complex lighting or partial occlusion. Deep learning frameworks, such as convolutional neural networks (CNNs), can also be considered to automatically learn and extract facial features, which can generally provide more accurate feature representations, especially when trained on large-scale datasets. The combined application of these techniques can significantly improve the effectiveness of facial image data in identity recognition.

[0071] In some embodiments, the step of acquiring facial key feature points and texture features includes:

[0072] Pre-process the image captured by the built-in camera of the helmet, obtain the facial area in the image, and obtain the key facial feature points and texture features through the feature extraction algorithm;

[0073] Taking the feature range of the standard facial image as the basis of the statistical range, taking the preset feature distance as the frequency band, counting the feature matching probability in each frequency band within the statistical range, and obtaining the facial feature distribution;

[0074] According to the feature changes of the facial images of the same person at different times, the facial feature change range is obtained, and based on the facial feature change range and facial feature distribution, the facial key feature points and texture features used for identification are obtained.

[0075] It should be noted that the steps for obtaining facial key feature points and texture features are described in detail, including preprocessing the image taken by the built-in camera of the helmet, obtaining the facial area in the image, and obtaining facial key feature points and texture features through a feature extraction algorithm. Here, preprocessing refers to a series of processing steps on the image in order to better extract facial features; feature extraction algorithm refers to a calculation method used to identify and extract key information in the image.

[0076] Specifically, the preprocessing steps may include grayscale, histogram equalization, edge detection, etc. of the image to enhance the recognizability of facial features. The acquisition of facial regions can be achieved through a face detection algorithm, such as a Haar cascade classifier or a deep learning face detection model. The feature extraction algorithm can be a method based on traditional computer vision, such as SIFT (Scale Invariant Feature Transform), SURF (Speeded Up Robust Features), or a method based on deep learning, such as using a pre-trained CNN (Convolutional Neural Network) model to extract facial features.

[0077] Preferably, in order to improve the accuracy and robustness of feature extraction, a multi-stage feature extraction process can be adopted. For example, in the preprocessing stage, spatial filtering and frequency filtering techniques can be combined to further optimize image quality. In the feature extraction stage, local feature extraction and global feature extraction methods can be combined to obtain a more comprehensive facial feature description.

[0078] Furthermore, liveness detection technologies such as blink detection or facial muscle movement analysis can be introduced to ensure that the facial images collected are from real living people rather than photos or videos. The combined use of these technologies can significantly improve the security and reliability of facial recognition systems.

[0079] In some embodiments, the calculation formula for obtaining the facial key feature points and texture features for identification based on the facial feature variation range and facial feature distribution is as follows:

[0080]

[0081] Among them, p i is the initial eigenvalue of facial feature point i, p i′ is the feature value of facial feature point i′ after the change, in pixels, p i ∈[P MIN ,P MAX ], P MIN is the minimum eigenvalue of facial feature point i, P MAX is the maximum eigenvalue of facial feature point i, ρ(p i ) is the facial feature point i with feature value p i The probability, ρ(p i′ ) is the facial feature point i′ with feature value p i′ The probability of ii′ is the relative distance between facial feature points i and i′ of the same person, and F is the comprehensive facial feature value used for identification.

[0082] It should be noted that a calculation formula is provided for determining the key facial feature points and texture features used for identification based on the facial feature variation range and facial feature distribution. Here, the facial feature variation range refers to the variation interval of the facial features of the same person at different time points, and the facial feature distribution refers to the frequency or probability distribution of each feature value within the standard facial image feature range.

[0083] Specifically, the calculation formula involves the initial eigenvalues ​​and changed eigenvalues ​​of facial feature points, as well as the probability distribution and relative distance between them. In this formula, x i represents the initial eigenvalue of facial feature point i, x i ′ Represents facial feature point i′ After the change in eigenvalue, p(x i ) and p(x i ′ ) are facial feature points with eigenvalues ​​x i and x i ′ The probability of d(x i ,x i ′ ) are facial feature points i and i ′ This formula calculates the changes of all feature points and their probability distribution to obtain a comprehensive facial feature value for recognition.

[0084] Preferably, this calculation process can be further optimized by machine learning methods. For example, a machine learning algorithm can be used to determine the importance of each facial feature point and their weight in the recognition process. In addition, more contextual information, such as ambient lighting, facial expression changes, etc., can be introduced to adjust the calculation of facial feature values.

[0085] Furthermore, in practical applications, we can also consider using deep learning techniques, such as autoencoders or generative adversarial networks (GANs), to automatically learn the changes and distribution of facial features, thereby improving the accuracy and robustness of the recognition system. The application of these technologies can make facial recognition systems more adaptable to different environments and conditions, and improve their effectiveness in practical applications.

[0086] In some embodiments, the behavior pattern data is obtained by summing the personnel activity trajectory and action frequency data corresponding to each safety helmet, wherein: when the personnel action frequency is higher than a preset frequency, the weight of the behavior pattern data corresponding to the personnel activity is γ, otherwise it is δ, and γ and δ are preset non-negative weight values.

[0087] It should be noted that the method for obtaining behavior pattern data is described, which is obtained based on the sum of the activity trajectory and action frequency data of the person corresponding to each helmet. Here, behavior pattern data refers to data for identifying the behavior pattern of an individual by analyzing the activity trajectory and action frequency of the individual, and the preset frequency refers to a threshold set in advance to distinguish normal and abnormal behavior patterns.

[0088] Specifically, the weight setting of the behavior pattern data can be determined according to the frequency of the personnel's actions. When the personnel's action frequency is higher than this preset frequency, the corresponding behavior pattern data weight is γ, otherwise it is δ. Here, γ and δ are pre-set non-negative weight values, which can be adjusted according to the actual helmet usage scenario and safety requirements. For example, on a construction site, if the worker's movement speed suddenly increases, it may indicate an emergency. At this time, a higher γ value can be set to reflect the importance of this emergency behavior.

[0089] Preferably, the analysis of the behavior pattern data can be further refined, including but not limited to the movement speed, dwell time, operation frequency, etc. For example, an algorithm can be set to monitor whether a worker completes a specific task within a specified time, or enters a restricted area during non-working hours.

[0090] Further, machine learning techniques can be introduced to automatically identify and learn normal behavior patterns and issue warnings when abnormal behavior is detected. Such a system can not only improve safety, but can also be used to optimize workflows and improve efficiency. Alternatives may include using different sensors to collect behavioral data, or adjusting weight parameters to adapt to different work environments and personnel behaviors.

[0091] In some embodiments, the step of constructing a feature vector based on the helmet built-in sensor data, the person's facial image data, and the behavior pattern data to establish an identity recognition model includes:

[0092] The weighted sum of the helmet built-in sensor data, the personnel facial image data and the behavior pattern data is used as the total eigenvalue, and maximizing the correct recognition probability is used as the objective function;

[0093] The first constraint is that each person has a corresponding helmet;

[0094] The second constraint condition is that each helmet can only be assigned to one recognition result or not be assigned to any recognition result;

[0095] The third constraint condition is that the number of people identified at the same time is not greater than the number that the system can handle;

[0096] Whether the identification result is confirmed is used as the decision variable constraint.

[0097] It should be noted that the construction process of the identity recognition model based on the smart helmet includes how to use the built-in sensor data of the helmet, the facial image data of the person and the behavior pattern data to build the identity recognition model. Here, the feature vector refers to the integration of data from these different sources into a mathematical representation that can be used for recognition, and the objective function refers to the specific performance indicator optimized when building the model, that is, maximizing the probability of correct recognition.

[0098] Specifically, the establishment of the identity recognition model involves weighting the data of the built-in sensor of the helmet, the facial image data of the person, and the behavior pattern data to form a total eigenvalue. In this process, different weight parameters need to be set to reflect the importance of each type of data in the recognition process. For example, the weights of sensor data, facial image data, and behavior pattern data can be determined based on historical data and experimental results. These weights can be automatically adjusted by machine learning algorithms to maximize the accuracy of the recognition model. The objective function can be recognition accuracy, recognition speed, or a balance between the two.

[0099] Preferably, the construction of the identity recognition model can be further refined into the following steps: first, preprocess the collected data, including normalization and noise removal; second, select a suitable feature extraction method, such as principal component analysis (PCA) or autoencoder, to reduce the data dimension and extract key features; then, use these features to train a classifier, such as a support vector machine (SVM) or a neural network, to establish an identity recognition model; finally, evaluate the performance of the model through methods such as cross-validation, and adjust the model parameters to optimize the objective function.

[0100] Furthermore, one can consider introducing online learning mechanisms to enable the model to adapt to new data and changing environmental conditions. Alternatives may include using different feature fusion techniques or adopting different machine learning frameworks to build the model.

[0101] In some embodiments, the personnel activity set corresponding to each helmet is constructed based on the discriminant value of the probability calculation model, including:

[0102] For each helmet, arrange the activities in the corresponding personnel activity set in chronological order, with each activity as a node. Set an initial feature matrix whose element values ​​are calculated based on the values ​​of the helmet's built-in sensor data, personnel facial image data, and behavior pattern data at the time of the activity. Calculate the feature difference between every two adjacent activity nodes. The difference comprehensively considers the changes in sensor data, facial features, and behavior patterns, and determines the weight of the edge based on the difference and the preset mapping relationship. At the same time, set a special start node and end node, which are connected to each activity node according to specific rules and assigned corresponding weights, thereby constructing a feature association graph.

[0103] It should be noted that how to construct a feature association graph based on the set of personnel activities corresponding to each helmet and the discriminant value of the probability calculation model is described. Here, the feature association graph is a data structure used to represent the association relationship and feature difference between different activity nodes, and the activity node refers to the data point related to a specific activity, which is arranged in chronological order, with each activity as a node.

[0104] Specifically, in the process of constructing the feature association graph, it is first necessary to arrange the activities in the personnel activity set in chronological order and set a node for each activity. Then, based on the built-in sensor data of the helmet, the personnel facial image data and the behavior pattern data, the initial feature matrix is ​​calculated at the time of the activity. Then, the feature difference between every two adjacent activity nodes is calculated. This difference comprehensively considers the changes in sensor data, facial features and behavior patterns. The weight of the edge is determined according to the difference and the preset mapping relationship. At the same time, a special starting node and an end node are set, which are connected to each activity node according to specific rules and assigned corresponding weights.

[0105] Preferably, the construction of the feature association graph can be further refined into the following steps: first, define a feature vector of an activity node, which contains information such as sensor readings, facial features, and behavior patterns; second, calculate the Euclidean distance or cosine similarity between adjacent activity nodes to quantify the feature difference; then, determine the existence of the edge and the size of the weight based on the difference and the preset threshold. The start node and the end node can represent the start and end states of identity recognition, and the connection weights between them and the activity node can be set according to the importance of the activity and the confidence of the recognition.

[0106] Furthermore, we can consider introducing algorithms from graph theory, such as the shortest path algorithm or the minimum spanning tree algorithm, to optimize the process of building the feature association graph. Alternatives may include using different graph structures, such as weighted undirected graphs or directed graphs, or adopting different feature difference calculation methods, such as anomaly detection algorithms based on machine learning.

[0107] In some embodiments, the step of obtaining the optimal path and the discriminant value of the feature association graph of each helmet includes:

[0108] The feature association graph is processed using a dynamic programming algorithm. Starting from the starting node, the optimal cumulative feature value to each node is calculated in turn, and the predecessor node information of each node is recorded. After traversing all nodes, the optimal path is obtained by traversing from the end node to the starting node. The discriminant value is obtained by calculating the weight product of all edges on the optimal path. The discriminant value reflects the credibility of the identity recognition result represented by this path.

[0109] It should be noted that the process of how to obtain the optimal path and discriminant value of the feature association graph of each helmet is described. Here, the optimal path refers to the path from the starting node to the end node in the feature association graph, and the cumulative feature value on this path is the largest; and the discriminant value refers to the weight product of all edges on this optimal path, which is used to reflect the credibility of the identity recognition result represented by the path.

[0110] Specifically, the process of obtaining the optimal path and its discriminant value involves processing the feature association graph using a dynamic programming algorithm. Starting from the starting node, the optimal cumulative feature value to each node is calculated in turn, and the predecessor node information of each node is recorded. This process involves traversing the path and updating the optimal value until all nodes are traversed. After the traversal is completed, the optimal path is obtained by backtracking from the end node to the starting node. The discriminant value is obtained by calculating the weight product of all edges on the optimal path. The larger this value is, the more reliable the identity recognition result represented by the path is.

[0111] Preferably, the process of obtaining the optimal path can be further refined into the following steps: first, initialize the optimal cumulative eigenvalue of the starting node to 0, and the values ​​of all other nodes to negative infinity or a sufficiently small number to indicate that they are initially unreachable; second, for each node, update the optimal cumulative eigenvalue of the current node according to the optimal cumulative eigenvalue of its predecessor node and the edge weight of the current node; then, select the node with the largest optimal cumulative eigenvalue as the current node, and update the optimal cumulative eigenvalue of its successor node. This process continues until all nodes have been visited. Finally, backtrack from the end node to find the path with the largest cumulative eigenvalue, that is, the optimal path. The calculation of the discriminant value can be completed by simply multiplying the weight of each edge on the optimal path.

[0112] Further alternatives include using other graph search algorithms, such as A* search or Bellamy's algorithm, or introducing heuristic information to optimize the path search process.

[0113] In some embodiments, the method of using a deep learning algorithm and continuously iterating until all solutions to the restricted master problem are stable solutions includes:

[0114] The discriminant value of the optimal path of the feature association graph of each helmet is input into an adaptively adjusted classifier, and the classifier adjusts itself according to the discriminant value and historical recognition results. If the discriminant value meets the current classification standard of the classifier, the recognition result corresponding to the corresponding optimal path is added to the recognition result set, and the parameters of the classifier are updated. If it does not meet the requirements, the reasons for the non-compliance are analyzed. If it is caused by factors such as data anomalies, the discriminant value and path are recalculated after the data is corrected;

[0115] If it is due to inaccurate model, adjust the classifier structure or parameters and iterate again until all solutions of the restricted main problem are stable solutions.

[0116] It should be noted that the process described is to use deep learning algorithms and continuous iterations until the solutions to the restricted main problem are all stable solutions. Here, the restricted main problem refers to the main problem that needs to be solved in the identity recognition process, but is subject to some constraints, such as data integrity, model accuracy, etc. A stable solution means that after multiple iterations, the recognition result no longer changes significantly, indicating that the model has converged to a reliable recognition state.

[0117] Specifically, the deep learning algorithm and the iterative process involve inputting the discriminant value of the optimal path of the feature association graph of each helmet into an adaptively adjusted classifier. The classifier adjusts itself based on the discriminant value and historical recognition results. This process includes updating the classifier parameters, as well as analyzing and correcting the data and model.

[0118] More specifically, the settings of specific parameters may include learning rate, number of iterations, threshold, etc. These parameters can be adjusted according to the actual recognition effect and model performance. Conceptually, this process can include multiple steps such as data preprocessing, feature extraction, model training, and result verification.

[0119] Preferably, the iterative process can be further refined into the following steps: first, preprocess the collected data to ensure the quality and consistency of the data; second, use the deep learning model to extract and learn the features of the data to obtain the initial recognition results; then, adjust the parameters of the classifier based on the initial results and historical data to improve the accuracy of recognition; then, verify the new recognition results, and stop the iteration if the results are stable; if the results are unstable, analyze the reasons and make corresponding data corrections or model adjustments.

[0120] Furthermore, it is possible to consider introducing more data sources and features to enhance the generalization ability of the model. Alternatives may include using different deep learning architectures, such as convolutional neural networks (CNN) or recurrent neural networks (RNN), or adopting ensemble learning methods, such as random forests or gradient boosting machines (GBM), to improve the stability and accuracy of the model.

[0121] The above-mentioned embodiments of the present invention have the following beneficial effects: The identity recognition method based on the smart helmet described in the present invention can provide a highly accurate and reliable personnel identity authentication solution. By integrating the temperature and acceleration sensor data, facial image data and behavior pattern data built into the helmet, the method can construct a comprehensive feature vector and then establish an identity recognition model. This method can effectively associate the identity of an individual with his or her activities, solve the probability calculation model through a deep learning algorithm, and realize rapid and accurate identification of the identity of a person. In addition, the method can improve the stability and accuracy of the recognition process by continuously iterating and optimizing until a stable identity recognition result is obtained.

[0122] This method can ensure efficient identification in a changing work environment, even when personnel activities are frequent and complex. By processing the feature association graph with a dynamic programming algorithm, the optimal path and its discriminant value can be found, thus reflecting the credibility of the identification result. This method can not only reduce the possibility of misidentification, but also improve the safety management level of the work site, ensuring that only authorized personnel can enter a specific work area, thereby enhancing the safety of high-risk environments such as industrial and construction sites.

[0123] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or terminal devices such as a computer, a server, a mobile phone, and a tablet.

[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Under the concept of the present invention, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes in different aspects of the present invention as above, which are not provided in detail for the sake of simplicity. Although the present invention is described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An identity recognition method based on a smart helmet, characterized in that: The steps include: Based on the built-in sensor data of the helmet, the facial image data of the personnel and the behavior pattern data, a feature vector is constructed, an identity recognition model is established, and then converted into a probability calculation model; Based on the personnel activity set corresponding to each helmet, a corresponding feature association graph is constructed according to the discriminant value of the probability calculation model; A deep learning algorithm with a neural network algorithm as the core is used to solve the probability calculation model, including: based on the helmet allocation information and the personnel work arrangement information, an initial identity recognition result corresponding to each helmet is obtained as an initial recognition result set, the solution of the restricted main problem of the probability calculation model is solved, the weight parameter is calculated, the optimal path and the discrimination value of the feature association graph of each helmet are cyclically obtained, and the recognition result corresponding to the optimal path with a discrimination value greater than a preset threshold is added to the recognition result set to continue to solve the solution of the restricted main problem; Through deep learning algorithms and continuous iterations, all solutions to the restricted master problem are stable solutions; Based on the stable solution and the final set of recognition results, a helmet identification solution is obtained.

2. The identity recognition method based on the smart helmet according to claim 1 is characterized in that: The built-in sensors of the helmet include a temperature sensor and an acceleration sensor; The helmet built-in sensor data is obtained according to the weighted sum of the sensor measurement values ​​of each helmet, wherein: when the sensor is a temperature sensor, the weight of its measurement value is α; When the sensor is an acceleration sensor, the weight of its measurement value is β, where α and β are preset non-negative weight values.

3. The identity recognition method based on the smart helmet according to claim 1 is characterized in that: The facial image data of the person is constructed based on key facial feature points and texture features.

4. The identity recognition method based on the smart helmet according to claim 3 is characterized in that: The steps of acquiring the facial key feature points and texture features include: Pre-process the image captured by the built-in camera of the helmet, obtain the facial area in the image, and obtain the key facial feature points and texture features through the feature extraction algorithm; Taking the feature range of the standard facial image as the basis of the statistical range, taking the preset feature distance as the frequency band, counting the feature matching probability in each frequency band within the statistical range, and obtaining the facial feature distribution; According to the feature changes of the facial images of the same person at different times, the facial feature change range is obtained, and based on the facial feature change range and facial feature distribution, the facial key feature points and texture features used for identification are obtained.

5. The identity recognition method based on the smart helmet according to claim 4 is characterized in that: The calculation formula for obtaining the facial key feature points and texture features for identification based on the facial feature variation range and facial feature distribution is as follows: Among them, p i is the initial eigenvalue of facial feature point i, p i′ is the feature value of facial feature point i′ after the change, in pixels, p i ∈[P MIN ,P MAX ], P MIN is the minimum eigenvalue of facial feature point i, P MAX is the maximum eigenvalue of facial feature point i, ρ(p i ) is the facial feature point i with feature value p i The probability, ρ(p i′ ) is the facial feature point i′ with feature value p i′ The probability of ii′ is the relative distance between facial feature points i and i′ of the same person, and F is the comprehensive facial feature value used for identification.

6. The identity recognition method based on the smart helmet according to claim 1 is characterized in that: The behavior pattern data is obtained by summing the personnel activity trajectory and action frequency data corresponding to each helmet, wherein: when the personnel action frequency is higher than the preset frequency, the weight of the behavior pattern data corresponding to the personnel activity is γ, otherwise it is δ, and γ and δ are preset non-negative weight values.

7. The identity recognition method based on the smart helmet according to claim 1 is characterized in that: The method of constructing a feature vector based on the helmet built-in sensor data, the personnel facial image data and the behavior pattern data, and establishing an identity recognition model includes: The weighted sum of the helmet built-in sensor data, the personnel facial image data and the behavior pattern data is used as the total eigenvalue, and maximizing the correct recognition probability is used as the objective function; The first constraint is that each person has a corresponding helmet; The second constraint condition is that each helmet can only be assigned to one recognition result or not be assigned to any recognition result; The third constraint condition is that the number of people identified at the same time is not greater than the number that the system can handle; Whether the identification result is confirmed is used as the decision variable constraint.

8. The identity recognition method based on the smart helmet according to claim 7 is characterized in that: The personnel activity set corresponding to each helmet is constructed based on the discriminant value of the probability calculation model, including: For each helmet, arrange the activities in the corresponding personnel activity set in chronological order, with each activity as a node; An initial feature matrix is ​​set, whose element values ​​are calculated based on the values ​​of the helmet built-in sensor data, the personnel facial image data and the behavior pattern data at the activity moment; Calculate the feature difference between every two adjacent active nodes. This difference takes into account the changes in sensor data, facial features, and behavior patterns. The weight of the edge is determined based on the difference and the preset mapping relationship. A special starting node and ending node are set, and each active node is connected according to specific rules and assigned corresponding weights to construct a feature association graph.

9. The identity recognition method based on the smart helmet according to claim 8, characterized in that: The method of obtaining the optimal path and the discriminant value of the feature association graph of each helmet includes: The dynamic programming algorithm is used to process the feature association graph. Starting from the starting node, the optimal cumulative feature value reaching each node is calculated in turn, and the predecessor node information of each node is recorded; After traversing all nodes, trace back from the end node to the start node to get the optimal path; The discriminant value is obtained by calculating the weight product of all edges on the optimal path, and the discriminant value indicates the credibility of the identity recognition result represented by this path.

10. The identity recognition method based on the smart helmet according to claim 1 or 9, characterized in that: The deep learning algorithm and continuous iteration are used until the solutions to the restricted master problem are all stable solutions, including: The discriminant value of the optimal path of the feature association graph of each helmet is input into an adaptively adjusted classifier, and the classifier adjusts itself according to the discriminant value and historical recognition results; If the discriminant value meets the current classification standard of the classifier, the recognition result corresponding to the corresponding optimal path is added to the recognition result set, and the parameters of the classifier are updated; If it does not meet the requirements, analyze the reasons for the non-compliance. If it is caused by data anomalies or other factors, correct the data and recalculate the discriminant value and path; If it is due to inaccurate model, adjust the classifier structure or parameters and iterate again until all solutions of the restricted main problem are stable solutions.

Images