Biometric Adaptive Instrument Interaction System for Industrial Vehicle Driving Safety

Abstract

The present invention discloses a biometric adaptive instrument interaction system for the driving safety of industrial vehicles, including a key module for receiving input instructions from the driver and serving as a preliminary interface for system interaction; a face acquisition module for real-time acquisition of the driver's face image and preprocessing of the image; a face recognition module for efficient feature extraction and comparison based on the VoVNet-ETSformer model to perform driver identity verification; a communication module for realizing real-time and secure data transmission among the internal modules of the system and between the system and an external control center or a background data platform; a storage module for storing the collected data; and a control module for integrating, coordinating, and managing each functional module to optimize the overall architecture of the system. The present invention realizes efficient, safe, and intelligent management of industrial vehicle identity recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of biometric adaptive instrument interaction, and particularly to a biometric adaptive instrument interaction system for industrial vehicle driving safety. Background Art

[0002] In the intelligent and safe driving management of industrial vehicles, the application of driver identification and interaction systems is crucial. With the development of industrial automation and intelligent driving technologies, traditional driver authentication methods face many challenges. For example, verification means such as keys, IC cards, or passwords are easily lost, forgotten, or copied, making it difficult to effectively prevent unauthorized operations. These verification methods rely on physical media and have low security. Once lost or stolen, they may lead to serious safety accidents. In addition, traditional authentication methods cannot monitor the driver's state in real time and lack effective means to deal with situations such as fatigue driving and driver identity changes.

[0003] Some current intelligent identity recognition systems have started to introduce biometric recognition technologies, especially face recognition technologies, to improve the security and convenience of authentication. Face recognition technology uses deep learning models to compare identities through facial features, which can effectively prevent illegal use and reduce security risks caused by human operations. However, most existing face recognition systems are applied in static environments or fixed devices, such as access control systems and attendance devices. In the industrial vehicle driving environment, factors such as complex lighting conditions, dynamic posture changes, partial occlusion, and vibration interference will reduce the accuracy of face recognition. In addition, existing face recognition methods still have deficiencies in recognition accuracy and stability when dealing with complex backgrounds, low-light environments, and high-speed dynamic changes, resulting in possible misidentifications or rejections in driver identity verification and affecting the normal use of vehicles.

[0004] To address these problems, some existing technologies have begun to combine deep learning with embedded computing to improve the real-time performance and robustness of face recognition. For example, some systems use convolutional neural networks (CNNs) for feature extraction and combine attention mechanisms to optimize the feature representation of key regions. However, these methods still have limitations. First, the efficiency of feature extraction is low. Traditional CNN structures are difficult to balance computational complexity and recognition accuracy, resulting in inefficient operation of embedded systems. Second, most current face recognition systems are based on static database matching and lack the adaptive ability for real-time driving environments, making it difficult to handle lighting, angle changes, and personalized features of different drivers. In addition, existing recognition methods mainly rely on fixed feature templates for comparison and cannot dynamically optimize feature selection, resulting in insufficient robustness in the face of environmental changes.

[0005] On the other hand, the interaction methods of existing industrial vehicle instrument systems are relatively traditional, mainly relying on physical buttons, touchscreens, or voice commands for operation, lacking real-time monitoring and adaptive adjustment of the driver's identity information. These interaction methods usually do not have a security verification function, and the driver may accidentally touch or perform unauthorized operations during the operation, increasing potential safety hazards. In addition, most existing instrument interaction systems lack intelligent data analysis functions and cannot perform adaptive adjustments in combination with factors such as the driver's identity and environmental conditions, making it difficult to efficiently manage the system in a multi-user environment. For example, in a factory or logistics warehousing environment, multiple drivers share the same industrial vehicle, and the system cannot perform personalized settings according to the habits and permissions of different drivers, resulting in increased management complexity and even potentially affecting operation safety.

[0006] To improve the driving safety of industrial vehicles, in recent years, some studies have attempted to apply optimization algorithms to face recognition to improve the efficiency of feature extraction and identity verification. For example, some methods introduce optimization algorithms such as ant colony optimization algorithm and particle swarm optimization algorithm to optimize the feature matching process to improve the recognition accuracy. However, the search efficiency of traditional optimization algorithms in high-dimensional feature spaces is relatively low, and it is difficult to quickly converge to the optimal solution, affecting real-time performance. In addition, these methods usually rely on fixed feature mapping strategies and lack the ability of adaptive optimization for complex driving environments, resulting in a decrease in recognition accuracy under conditions such as light changes and partial occlusion.

[0007] Most existing face recognition optimization methods are mostly limited to a single feature extraction strategy and cannot fully utilize the multi-scale information expression ability of deep learning networks. For example, some studies use standard convolutional neural networks (CNNs) for feature extraction, but the CNN structure has the problem of limited local feature expression and is difficult to capture global features. Although the self-attention mechanism (Transformer) can model long-range feature dependencies, its computational complexity is relatively high and it is difficult to operate efficiently on embedded systems. In addition, existing feature optimization methods mainly focus on the feature selection level and lack end-to-end optimization of the feature extraction process, resulting in the recognition results being limited by the quality of the initial features and unable to fully adapt to complex environmental changes.

[0008] Generally speaking, the existing industrial vehicle driver identity recognition systems mainly have the following defects: First, the traditional identity verification method has low security and is easily forged or bypassed; second, the existing face recognition methods lack robustness in complex driving environments and are difficult to cope with light changes, pose changes, and dynamic interferences; third, the existing deep learning models have low computational efficiency on embedded systems and are difficult to achieve efficient and real-time identity recognition; finally, the existing instrument interaction systems lack the ability to link with the driver's identity information and are difficult to achieve intelligent management.

[0009] In view of these problems, the present invention proposes a biometric adaptive instrument interaction system for industrial vehicle driving safety, which uses the VoVNet-ETSformer network for efficient feature extraction and combines an improved artificial bee colony optimization algorithm (I-ABC) to optimize the feature selection process, enhancing the recognition accuracy and real-time performance. This system integrates a face acquisition, face recognition, data storage, communication, and control module in the industrial vehicle instrument to achieve efficient verification of the driver's identity and improve the system's adaptive ability by combining intelligent optimization algorithms. Through innovative optimization strategies, the present invention improves the stability of face recognition in complex driving environments, enhances the security and intelligence level of the instrument interaction system, and solves a number of key problems in the prior art.

[0010] Therefore, how to provide a biometric adaptive instrument interaction system for industrial vehicle driving safety is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0011] An object of the present invention is to propose a biometric adaptive instrument interaction system for industrial vehicle driving safety. The present invention makes full use of the VoVNet-ETSformer deep learning model, the improved artificial bee colony optimization algorithm, and the intelligent instrument interaction system, and details the optimization method for efficiently and accurately identifying the driver's identity and performing interactive control in complex driving environments, with the advantages of high identity verification security, high recognition accuracy, strong real-time performance, and strong system intelligent management ability.

[0012] The biometric adaptive instrument interaction system for industrial vehicle driving safety according to an embodiment of the present invention includes:

[0013] A key module for receiving the driver's input instructions and serving as the initial interface for system interaction;

[0014] A face acquisition module for real-time collecting the driver's face image and preprocessing the image;

[0015] A face recognition module for implementing efficient feature extraction and comparison based on the VoVNet-ETSformer model to verify the driver's identity;

[0016] A communication module for realizing real-time and secure data transmission between each module inside the system and with an external control center or a background data platform;

[0017] A storage module for storing the collected data;

[0018] A control module for integrating, coordinating, and managing each functional module to optimize the overall system architecture.

[0019] Optionally, the modules are implemented by the following method:

[0020] S1. Receive the input instructions of the driver through the button module, use the face acquisition module to collect the driver's face image in real time, and preprocess the face image;

[0021] S2. Use the VoVNet-ETSformer model to standardize the features of the preprocessed face image, and use the multi-scale alignment method for geometric transformation;

[0022] S3. Use the single aggregation structure of the VoVNet network to extract multi-scale features from the geometrically transformed image, combine the self-attention mechanism of the ETSformer module to enhance the feature expression ability, and obtain the extracted features;

[0023] S4. Optimize the extracted features based on the improved artificial bee colony optimization algorithm, initialize the bee colony, set the population size, randomly generate candidate feature subsets, employ bees to evaluate the features extracted by VoVNet-ETSformer, adaptively allocate weights, use chaotic mapping for scout bees to enhance the global search ability, eliminate local optima, use the improved LévyFlight search strategy for follower bees to optimize feature selection, and use the adaptive mutation mechanism to adjust the feature subsets, screen the optimal feature subsets, and generate the identity recognition result;

[0024] S5. Transmit the generated identity recognition result to the storage module through the communication module, and compare it with the pre-stored authorization information;

[0025] S6. Generate operation permission information according to the comparison result in the control module, display it in real time through the embedded instrument, encrypt the data transmitted by the communication module, monitor the system status in real time, schedule system resources, and record and store the identity authentication and interaction data.

[0026] Optionally, the specific content of S3 includes:

[0027] S31. Extract multi-scale features from the geometrically transformed face image, use the single aggregation structure of the VoVNet network to establish a cross-layer connection mechanism for different levels of feature information, extract local texture information through the deep convolutional module, and at the same time use the cross-scale aggregation strategy to fuse global features;

[0028] S32. For the extracted multi-scale features, use the dilated convolution technique to expand the features, introduce convolution kernels with different dilation rates to expand the receptive field, and the model learns the key facial features at different scales;

[0029] S33. Combine the self-attention mechanism of the ETSformer module, use multi-head self-attention to calculate the global correlation between facial feature points, calculate feature weights in different feature spaces, emphasize the feature contributions of key regions, and maintain spatial structure information through position encoding technology;

[0030] S34. Perform a non-linear transformation on the features optimized by the ETSformer module, adopt a normalization strategy to enhance feature contrast, and adjust the distribution of feature vectors;

[0031] S35. Perform a feature mapping transformation on the finally extracted features, use the embedding space mapping method to project high-dimensional features into a low-dimensional identity feature space, and output the optimized features.

[0032] Optionally, the specific steps of S4 are as follows:

[0033] S41. Initialize parameters based on the artificial bee colony optimization algorithm, set the feature vector set , where represents the i-th feature vector, set the initial candidate feature subset , represents the j-th candidate feature, define the initial population size N, set the number of iterations T and the fitness calculation function, set the optimization objective function J(S), which reflects the overall fitness of the entire feature subset S, and calculate the initial weight of each feature :

[0034] ;

[0035] Among them, represents the i-th candidate feature, represents the weight of feature , which is not a fixed value and is continuously updated. The value of J(S) will change with changing, represents the similarity between feature and the target feature set D;

[0036] S42. In the employed bee stage, through feature weight calculation, feature contribution degree calculation, and feature distribution balance calculation, screen discriminative features, remove redundant features, and cover various feature information;

[0037] S43. In the scout bee stage, adopt an improved chaotic mapping strategy for global search, combine feature subset fitness calculation and global feature distribution calculation, expand the search range, and avoid local optima;

[0038] S44. In the follower bee stage, use an improved search strategy for local optimization, and combine local feature balance metrics to make the feature subset reasonably distributed in all feature dimensions;

[0039] S45. Calculate the fitness of the newly generated feature subset, screen the feature subset based on the fitness, normalize the selected optimal feature subset, obtain the final feature vector, and generate the identity recognition result.

[0040] Optionally, the S42 specifically includes:

[0041] S421. Calculate the iteratively optimized feature weights of each individual, and introduce feature sparsity regularization: ,

[0042] ;

[0043] where m is the total number of feature subsets, n is the total number of feature vectors, starting from the initial weight of each feature to calculate the iteration, and the finally iteratively optimized weight is denoted as is the matching degree of feature in subset , is the feature sparsity control parameter;

[0044] S422. Calculate the feature contribution degree and construct the feature screening objective function:

[0045] ;

[0046] where, is the distribution density of feature in the candidate subset, , is the adaptive adjustment parameter;

[0047] S423. Calculate the feature distribution balance as the initial measure of the global feature balance:

[0048] ;

[0049] where, represents the probability distribution of feature in subset S, and log is the logarithmic function.

[0050] Optionally, the S43 specifically includes:

[0051] S431. Introduce an improved chaotic mapping search to generate a new feature subset:

[0052] ;

[0053] where, is the global search control parameter, is a perturbation factor, combined with the sine chaotic map, and sin is the sine function;

[0054] S432. Calculate the fitness of the feature subset. For the newly generated feature subset , calculate its fitness:

[0055] ;

[0056] Among them, is the feature selection sensitivity parameter;

[0057] S433. Optimize the global feature distribution. For the new feature subset , calculate the global feature balance metric:

[0058] ;

[0059] Among them, is the feature balance control parameter.

[0060] Optionally, the S44 specifically includes:

[0061] S441. Local optimization search strategy. Introduce improved search to optimize the feature subset :

[0062] ;

[0063] Among them, is the optimized feature subset, is the step size factor, is the search control parameter, t is the current iteration number, Adopt distribution, and the mathematical form is as follows:

[0064] ;

[0065] Among them, is the exponential parameter of the distribution, is the gamma function, x represents the generated random variable, and sin is the sine function;

[0066] S442. Refer to the fitness of the feature subset to screen for a suitable feature subset for balance optimization, and calculate the local feature balance metric:

[0067] ;

[0068] Among them, is the feature balance control parameter.

[0069] Optionally, S45 specifically includes:

[0070] S451. Calculate the final fitness , improve the fitness calculation method, adopt a weighted fusion model, and make the contributions of various measurement indicators more reasonable:

[0071] ;

[0072] Among them, , , , are the weighting coefficients;

[0073] S452. Calculate the fitness of the newly generated feature subset , and define the final screening target:

[0074] ;

[0075] Among them, represents the optimal feature subset finally selected, means to return the feature subset that causes to obtain the maximum value;

[0076] S453. Normalize the selected optimal feature subset to obtain the final feature vector and generate the identity recognition result.

[0077] The present invention combines the VoVNet-ETSformer deep learning model, the improved artificial bee colony optimization algorithm (I-ABC), and the intelligent instrument interaction system, and has made remarkable technological progress in the identity recognition and interaction management of industrial vehicle drivers, which is specifically reflected in the following three beneficial effects:

[0078] (1) The present invention uses the VoVNet-ETSformer deep learning model for efficient feature extraction and identity recognition. Combining multi-scale feature extraction, dilated convolution, and self-attention mechanism, it can still maintain high-precision face recognition ability under complex working conditions such as light changes, driver posture changes, and partial occlusion. At the same time, the multi-scale alignment method is used to perform geometric transformation on facial features, improving the robustness of the model and ensuring that different drivers can quickly and accurately pass the identity verification, reducing the occurrence probability of misrecognition and rejection. Compared with the traditional identity verification methods based on IC cards or passwords, the biometric recognition system of the present invention is more secure and difficult to be forged or bypassed, effectively preventing unauthorized personnel from driving the vehicle and improving the safety of vehicle management.

[0079] (2) In the process of feature optimization, the present invention introduces an improved artificial bee colony optimization algorithm (I-ABC), including feature screening in the employed bee stage, global search in the scout bee stage, and local optimization in the follower bee stage, ensuring rapid convergence to the optimal feature subset in the high-dimensional feature space. The global search strategy of chaotic mapping is used to avoid the optimization falling into local optimum, and the Lévy Flight search strategy is combined to further optimize the feature subset, improving the real-time performance of identity recognition. At the same time, an adaptive mutation mechanism is adopted to adjust the feature subset, enhancing the stability of feature selection, enabling the system to operate efficiently in different driving environments. Compared with the traditional CNN model, the computing efficiency is improved, and the consumption of computing resources is reduced, enabling the present system to operate efficiently in embedded industrial instruments and meet the real-time requirements of vehicle intelligent interaction.

[0080] (3) The present invention integrates intelligent data storage, encrypted communication, and control modules in the industrial vehicle instrument interaction system, dynamically adjusts the driver's permissions based on the authentication results, and encrypts the communication data, improving the security of the system. In addition, the system can adaptively adjust the interaction interface according to the driver's identity, operation habits, and environmental status, providing personalized instrument display solutions for different drivers and enhancing the user experience. At the same time, the system can monitor the driver's operation status in real time, prompt or restrict abnormal driving behaviors (such as fatigue driving, illegal operations), effectively reducing the safety risks of industrial vehicles and improving the intelligent management level. Compared with the existing industrial vehicle instrument systems, the present invention realizes the deep integration of identity recognition and instrument interaction, significantly enhancing the intelligent and automated management capabilities of industrial vehicles.

[0081] In summary, the present invention not only improves the security and accuracy of industrial vehicle driver authentication, but also optimizes feature extraction and computing efficiency, while enhancing the intelligent management capabilities of the instrument interaction system, providing an innovative solution for the safe driving and efficient management of industrial vehicles. Brief Description of the Drawings

[0082] The drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0083] Figure 1 is a flowchart of the biometric adaptive instrument interaction method for industrial vehicle driving safety proposed by the present invention;

[0084] Figure 2 is a schematic diagram of the biometric adaptive instrument interaction system module for industrial vehicle driving safety proposed by the present invention. Detailed Embodiments

[0085] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0086] Reference Figure 1 and Figure 2 , a biometric adaptive instrument interaction system for industrial vehicle driving safety, comprising:

[0087] A key module, used to receive input instructions from the driver and serve as the initial interface for system interaction;

[0088] A face acquisition module, used to acquire the driver's face image in real time and preprocess the image;

[0089] A face recognition module, which realizes efficient feature extraction and comparison based on the VoVNet-ETSformer model for driver identity verification;

[0090] A communication module, used to realize real-time and secure data transmission between various modules within the system and with an external control center or a background data platform;

[0091] A storage module, used to store the acquired data;

[0092] A control module, used to integrate, coordinate and manage each functional module and optimize the overall architecture of the system.

[0093] In this embodiment, the modules are implemented through the following methods:

[0094] S1. Receive the driver's input instructions through the key module, use the face acquisition module to acquire the driver's face image in real time, and preprocess the face image;

[0095] S2. Standardize the features of the preprocessed face image using the VoVNet-ETSformer model, and perform geometric transformation using the multi-scale alignment method;

[0096] S3. Use the single aggregation structure of the VoVNet network to perform multi-scale feature extraction on the geometrically transformed image, and combine the self-attention mechanism of the ETSformer module to enhance the feature expression ability to obtain the extracted features;

[0097] S4. Initialize the bee swarm based on the features optimized by the improved artificial bee colony optimization algorithm, set the population size, randomly generate candidate feature subsets, have employed bees evaluate the features extracted by VoVNet-ETSformer, adaptively allocate weights, have scout bees enhance the global search ability using chaotic mapping to eliminate local optima, have follower bees optimize feature selection using an improved Lévy Flight search strategy, adjust the feature subsets using an adaptive mutation mechanism, screen the optimal feature subsets, and generate identity recognition results;

[0098] S5. Transmit the generated identity recognition results to the storage module through the communication module and compare them with the pre-stored authorized information;

[0099] S6. Generate operation permission information in the control module according to the comparison results, display it in real time through the embedded instrument, encrypt the data transmitted by the communication module, monitor the system status in real time, schedule system resources, and record and store identity authentication and interaction data.

[0100] In this embodiment, S3 specifically includes:

[0101] S31. Perform multi-scale feature extraction on the geometrically transformed face image. Using the single aggregation structure of the VoVNet network, establish a cross-layer connection mechanism for feature information at different levels, extract local texture information through the deep convolutional module, and simultaneously fuse global features using a cross-scale aggregation strategy;

[0102] S32. For the extracted multi-scale features, use dilated convolution technology to expand the features, introduce convolution kernels with different dilation rates to expand the receptive field, and have the model learn key facial features at different scales;

[0103] S33. Combine the self-attention mechanism of the ETSformer module, use multi-head self-attention to calculate the global correlation between facial feature points, calculate feature weights in different feature spaces, emphasize the feature contributions of key regions, and maintain spatial structure information through position encoding technology;

[0104] S34. Perform non-linear transformation on the features optimized by the ETSformer module, use a normalization strategy to enhance the feature contrast, and adjust the distribution of the feature vectors;

[0105] S35. Perform feature mapping transformation on the finally extracted features, use the embedding space mapping method to project high-dimensional features into a low-dimensional identity feature space, and output the optimized features.

[0106] In this embodiment, S4 specifically includes:

[0107] S41. Initialize the parameters based on the artificial bee colony optimization algorithm and set the feature vector set , where represents the i-th eigenvector, and an initial candidate feature subset is set , represents the j-th candidate feature. Define the initial population size N, set the number of iterations T and the fitness calculation function, set the optimization objective function J(S) to reflect the overall fitness of the entire feature subset S, and calculate the initial weight of each feature :

[0108] ;

[0109] Among them, represents the i-th candidate feature, represents the feature The weight of is not a fixed value and is continuously updated. The value of J(S) will change with changing, represents the feature The similarity between and the target feature set D;

[0110] S42. In the employed bee stage, discriminative features are screened, redundant features are removed, and various feature information is covered through feature weight calculation, feature contribution degree calculation, and feature distribution balance calculation;

[0111] S43. In the scout bee stage, an improved chaotic mapping strategy is adopted for global search, combined with feature subset fitness calculation and global feature distribution calculation, to expand the search range and avoid local optima;

[0112] S44. In the follower bee stage, an improved search strategy is used for local optimization, combined with local feature balance measurement to make the feature subset reasonably distributed in all feature dimensions;

[0113] S45. Calculate the fitness of the newly generated feature subset, screen the feature subset based on the fitness, normalize the selected optimal feature subset, obtain the final eigenvector, and generate the identity recognition result.

[0114] In this embodiment, the S42 specifically includes:

[0115] S421. Calculate the iteratively optimized feature weight of each individual , and introduce feature sparsity regularization:

[0116] ;

[0117] Among them, m is the total number of feature subsets, n is the total number of feature vectors, Starting from the initial weight of each feature, calculate the iteration. The finally iteratively optimized weight is denoted as is the feature In the subset The matching degree in is the feature sparsity control parameter;

[0118] S422. Calculate the feature contribution and construct the feature screening objective function:

[0119] ;

[0120] in, Features The distribution density in the candidate subset, , To adjust parameters adaptively;

[0121] S423. Calculate the feature distribution balance as the initial measure of global feature balance:

[0122] ;

[0123] in, Representation characteristics The probability distribution in the subset S, log is the logarithmic function.

[0124] In this implementation manner, the S43 specifically includes:

[0125] S431, introduce improved chaotic mapping search to generate new feature subsets:

[0126] ;

[0127] in, is the global search control parameter, is the disturbance factor, combined with the sine chaotic mapping, sin is the sine function;

[0128] S432, calculate the fitness of the feature subset, for the newly generated feature subset , calculate its fitness:

[0129] ;

[0130] in, Select sensitivity parameters for features;

[0131] S433, optimize the global feature distribution, for the new feature subset , calculate the global feature balance metric:

[0132] ;

[0133] in, is the characteristic equalization control parameter.

[0134] In this embodiment, S44 specifically includes:

[0135] .S441. Local optimization search strategy, introducing improved search to optimize the feature subset :

[0136] ;

[0137] Among them, is the optimized feature subset, is the step size factor, is the search control parameter, t is the current iteration number, Adopt distribution, and the mathematical form is as follows:

[0138] ;

[0139] Among them, is the exponential parameter of the distribution, is the gamma function, x represents the generated random variable, and sin is the sine function;

[0140] S442. Fitness of the reference feature subset Screen a suitable feature subset for balance optimization and calculate the local feature balance metric:

[0141] ;

[0142] Among them, controls the change rate of feature balance.

[0143] In this embodiment, S45 specifically includes:

[0144] S451. Calculate the final fitness , improve the fitness calculation method, and adopt a weighted fusion model to make the contribution of each metric more reasonable:

[0145] ;

[0146] Among them, , , , are the weighting coefficients;

[0147] S452. Calculate the fitness of the newly generated feature subset and define the final screening target:

[0148] ;

[0149] Among them, represents the optimal feature subset finally selected, indicating the return of the feature subset that achieves the maximum value;

[0150] S453. Normalize the selected optimal feature subset to obtain the final feature vector and generate the identity recognition result.

[0151] S453. Normalize the selected optimal feature subset to obtain the final feature vector and generate the identity recognition result.

[0152] Example 1:

[0153] To verify the feasibility of the present invention in implementation, the present invention is applied to the intelligent logistics vehicle management system of a large industrial park. The park contains multiple workshops and warehousing areas, and there are hundreds of industrial vehicles (such as forklifts, tractors, automatic guided vehicles AGV) performing handling, distribution, and loading / unloading tasks in different areas on a daily basis. Due to the large number of industrial vehicles in the park and the chaotic management of driver identities, there are security risks such as unauthorized driving and misoperation. In addition, since drivers often need to change vehicles, the traditional identity verification methods based on IC cards or passwords are not only inefficient but also prone to card loss or password leakage, affecting normal operations.

[0154] The biometric adaptive instrument interaction system for industrial vehicle driving safety of the present invention is tested and applied in this scenario. The system first integrates a key module, a face acquisition module, a face recognition module, a storage module, a communication module, and a control module on the instrument panel of the industrial vehicle. After the driver enters the vehicle, the key module triggers the face acquisition module to collect face images, and performs feature extraction and comparison through the VoVNet-ETSformer network, and combines the improved artificial bee colony optimization algorithm (I-ABC) for feature optimization. The entire recognition process takes less than 1.2 seconds, and the driver can quickly complete identity verification without using a key or an IC card. After the identity verification is passed, the system will automatically load the operation permissions of the driver and adjust the instrument interface according to his operation habits, such as adjusting the brightness of the instrument panel and displaying the layout of the most commonly used control buttons.

[0155] The following are the performance test data of the system under different environments and working conditions:

[0156] Table 1: Performance test results of the biometric adaptive instrument interaction system for industrial vehicle driving safety

[0157] ;

[0158] During the specific testing process, the system conducts face recognition tests under different conditions, including daytime, nighttime, strong light, low light, and complex environments (such as changing lighting in a warehouse and a driver wearing a safety helmet). The experimental results show that the face recognition system of the present invention maintains a high recognition accuracy under various lighting conditions, with an average accuracy rate of 99.2%, which is approximately 3.5% higher than that of traditional CNN recognition models. In addition, in dynamic environments (such as a driver walking towards a vehicle from a distance and approaching the camera from different angles), the system can still complete identity verification within 1.5 seconds, and the rejection rate is lower than 0.8%. Under extreme lighting changes (such as direct sunlight and backlight environments), the recognition accuracy still remains above 97.5%, demonstrating strong environmental adaptability.

[0159] To further verify the security of the intelligent instrument interaction system of the present invention, the system simulates various unauthorized driving scenarios, such as: a non-registered driver attempting to start the vehicle, abnormal driver identity information (such as wearing a mask and covering part of the face), etc. The results show that the system successfully intercepts 98.7% of illegal driving attempts, provides safety prompts on the dashboard, and sends abnormal driving alerts to the background management system through the communication module. In addition, the encryption communication mechanism adopted by the present invention effectively prevents the identity data from being stolen or tampered with, improving data security compared with the traditional IC card-based identity authentication method.

[0160] During the peak scheduling period of industrial vehicles (such as 9:00 - 11:00 am and 14:00 - 16:00 pm every day), the traditional IC card or password verification method usually results in a long queue waiting time, with an average verification time of about 8.5 seconds for each vehicle. However, with face recognition, this system only requires 1.2 seconds, significantly improving the vehicle usage efficiency. After 3 months of test operation, the average identity verification time of industrial vehicles in this park has been shortened by 85.9%, the driver identity management efficiency has been increased by 67.3%, the vehicle misuse situation has been reduced by 92.5%, and the overall security and intelligence level of the industrial vehicle management system have been significantly improved.

[0161] The present invention also tests the system operation performance during the driver identity verification process. It runs under extreme industrial environments such as low temperature (-10°C), high temperature (45°C), high humidity (90%RH), and strong vibration, and the system can normally recognize without any recognition failure or misjudgment. In terms of computing resource occupancy, when the optimized VoVNet-ETSformer network of the present invention runs on an embedded device (NVIDIA Jetson AGX Xavier), the average CPU occupancy rate is 38.6%, which is 17.2% lower than that of the traditional CNN recognition solution, and the energy efficiency ratio is increased by 22.8%, indicating the high efficiency of this system in the embedded environment.

[0162] As described above, it is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A biometric adaptive instrument interaction system for the driving safety of industrial vehicles, characterized in that, Including: A button module, which is used to receive the input instructions of the driver and serves as the initial interface for system interaction; A face acquisition module, which is used to acquire the driver's face image in real time and preprocess the image; A face recognition module, which realizes efficient feature extraction and comparison based on the VoVNet-ETSformer model and an improved artificial bee colony optimization algorithm for driver identity authentication, including: Performing multi-scale feature extraction on the geometrically transformed face image, using the single aggregation structure of the VoVNet network to establish a cross-layer connection mechanism for different levels of feature information, extracting local texture information through a deep convolutional module, and at the same time fusing global features using a cross-scale aggregation strategy; For the extracted multi-scale features, using dilated convolution technology for feature expansion, introducing convolution kernels with different dilation rates to expand the receptive field, and the model learns key face features at different scales; Combining the self-attention mechanism of the ETSformer module, using multi-head self-attention to calculate the global correlation between face feature points, calculating feature weights in different feature spaces, emphasizing the feature contribution of key regions, and maintaining spatial structure information through position encoding technology; Performing a non-linear transformation on the features optimized by the ETSformer module, adopting a normalization strategy to enhance feature contrast and adjust the distribution of feature vectors; Performing a feature mapping transformation on the finally extracted features, using an embedding space mapping method to project high-dimensional features into a low-dimensional identity feature space and output optimized features; Initialize parameters based on the artificial bee colony optimization algorithm and set the feature vector set , where represents the i-th feature vector, and set the initial candidate feature subset , where m is the total number of feature subsets and n is the total number of feature vectors. represents the j-th candidate feature. Define the initial population size N, set the number of iterations T and the fitness calculation function, set the optimization objective function J(S) to reflect the overall fitness of the entire feature subset S, and calculate the initial weight of each feature : ; Among them, represents the i-th candidate feature, represents the feature weight, which is not a fixed value and is constantly updated. The value of J(S) will change with changes, represents the similarity between the feature and the target feature set D; In the employed bee stage, discriminative features are screened by calculating feature weights, feature contribution degrees, and feature distribution balance, removing redundant features to cover a variety of feature information; In the scout bee stage, an improved chaotic mapping strategy is adopted for global search, combined with feature subset fitness calculation and global feature distribution calculation to expand the search range and avoid local optima; The follower bee stage uses an improved search strategy for local optimization, and combines the local feature equilibrium metric to make the feature subset reasonably distributed in all feature dimensions; Calculating the fitness of the newly generated feature subset, screening the feature subset based on the fitness, normalizing the selected optimal feature subset to obtain the final feature vector, and generating an identity recognition result; A communication module, which is used to realize real-time and secure data transmission between each module inside the system and with an external control center or a background data platform; A storage module, which is used to store the collected data; A control module, which is used to integrate, coordinate, and manage each functional module to optimize the overall system architecture.

2. The biometric adaptive instrument interaction system for industrial vehicle driving safety according to claim 1, wherein The modules are implemented through the following methods: S1. Receive the driver's input instructions through the button module, use the face acquisition module to acquire the driver's face image in real time, and preprocess the face image; S2. Perform feature standardization on the preprocessed face image based on the VoVNet-ETSformer model, and perform geometric transformation using a multi-scale alignment method; S3. Use the single aggregation structure of the VoVNet network to perform multi-scale feature extraction on the geometrically transformed image, and combine the self-attention mechanism of the ETSformer module to enhance the feature expression ability to obtain the extracted features, including: S31. Perform multi-scale feature extraction on the geometrically transformed face image. Utilize the single aggregation structure of the VoVNet network to establish a cross-layer connection mechanism for different levels of feature information. Extract local texture information through deep convolutional modules, and at the same time fuse global features using a cross-scale aggregation strategy. S32. For the extracted multi-scale features, use dilated convolution technology for feature expansion. Introduce convolution kernels with different dilation rates to expand the receptive field, and the model learns key facial features at different scales. S33. Combine the self-attention mechanism of the ETSformer module. Use multi-head self-attention to calculate the global correlation between facial feature points, calculate feature weights in different feature spaces, emphasize the feature contributions of key regions, and maintain spatial structure information through position encoding technology. S34. Perform a non-linear transformation on the features optimized by the ETSformer module, and adopt a normalization strategy to enhance feature contrast and adjust the distribution of feature vectors. S35. Perform a feature mapping transformation on the finally extracted features. Use the embedding space mapping method to project high-dimensional features into a low-dimensional identity feature space and output the optimized features. S4. Based on the features optimized by the improved artificial bee colony optimization algorithm, initialize the bee colony, set the population size, randomly generate candidate feature subsets, the employed bees evaluate the features extracted by VoVNet-ETSformer, adaptively allocate weights, the scout bees use chaotic mapping to enhance the global search ability, eliminate local optima, and the follower bees use the improved search strategy to optimize feature selection, the adaptive mutation mechanism adjusts the feature subsets, filters out the optimal feature subset, and generates the identity recognition result, including: S41. Initialize parameters based on the artificial bee colony optimization algorithm and set the feature vector set , where represents the i-th feature vector, and set the initial candidate feature subset , where m is the total number of feature subsets and n is the total number of feature vectors. represents the j-th candidate feature. Define the initial population size N, set the number of iterations T and the fitness calculation function, set the optimization objective function J(S) to reflect the overall fitness of the entire feature subset S, and calculate the initial weight of each feature : ; Among them, represents the i-th candidate feature, represents the weight of feature , which is not a fixed value and is continuously updated. The value of J(S) will change with changing, represents the similarity between feature and the target feature set D; S42. In the employed bee stage, screen discriminative features, remove redundant features, and cover a variety of feature information through feature weight calculation, feature contribution degree calculation, and feature distribution balance calculation. S43. In the scout bee stage, adopt an improved chaotic mapping strategy for global search. Combine feature subset fitness calculation and global feature distribution calculation to expand the search range and avoid local optima. S44. During the follower bee stage, local optimization is carried out using an improved search strategy, and the local feature equilibrium metric is combined to make the feature subset reasonably distributed in all feature dimensions; S45. Calculate the fitness of the newly generated feature subset, screen the feature subset based on fitness, normalize the selected optimal feature subset to obtain the final feature vector, and generate an identity recognition result. S5. Transmit the generated identity recognition result to the storage module through the communication module and compare it with the pre-stored authorized information. S6. Generate operation permission information in the control module according to the comparison result, display it in real time through the embedded instrument, encrypt the data transmitted by the communication module, monitor the system status in real time, schedule system resources, and record and store identity authentication and interaction data.

3. The biometric adaptive instrument interaction system for industrial vehicle driving safety according to claim 2, wherein The specific content of S42 includes: S421. Calculate the feature weights after iterative optimization for each individual , and introduce feature sparsity regularization: ; where m is the total number of feature subsets and n is the total number of feature vectors, starting from the initial weight of each feature calculate iterations, and the finally iteratively optimized weight is denoted as for feature in the subset matching degree, is the feature sparsity control parameter; S422. Calculate the feature contribution degree and construct a feature screening objective function: ; Among them, is the feature distribution density in the candidate subset, , is the adaptive adjustment parameter; S423. Calculate the feature distribution balance as the initial measure of global feature balance: ； Among them, represents the probability distribution of the feature in the subset S, and log is the logarithmic function.

4. The biometric adaptive instrument interaction system for industrial vehicle driving safety according to claim 3, characterized in that, The specific content of S43 includes: S431. Introduce improved chaotic mapping search to generate a new feature subset: ; Among them, is the global search control parameter, is the perturbation factor. Combining with the sine chaotic map, sin is the sine function; S432. Calculate the fitness of the feature subset. For the newly generated feature subset , calculate its fitness: ; Among them, is the feature selection sensitivity parameter; S433. Optimize the global feature distribution. For the new feature subset , calculate the global feature balance metric: ; Among them, is the feature balance control parameter.

Images