A driver identity dual verification apparatus and method for industrial vehicles

Abstract

The application provides a driver identity double verification device and method for an industrial vehicle, which combines a lightweight face recognition algorithm and a wireless radio frequency identification technology, optimizes a GhostFaceNets network structure, improves an attention mechanism, and introduces a multi-sub-center improved loss function, so as to improve the recognition accuracy and speed of a small model on an embedded device, the system adopts FP16 / FP32 mixed precision and 8-bit quantization deployment, significantly reduces the computing resources and power consumption, designs a Bluetooth registration and binding process, realizes local pairing and verification of a driver IC card UID and face features and identity information, and thus ensures the safety of the system while enhancing the flexibility and industrial adaptability of device deployment.

Description

Technical Field

[0001] This invention relates to the fields of facial recognition and lightweight models, specifically to a dual verification device and method for the identity of industrial vehicle drivers that combines a lightweight facial recognition algorithm with radio frequency identification technology. Background Technology

[0002] Currently, in the industrial vehicle sector, especially in forklifts and other material handling equipment, identity verification is a crucial aspect of enhancing security. Traditional identity verification methods rely on cards or passwords, but these methods have inherent security vulnerabilities, such as lost or stolen cards. Therefore, incorporating more secure and convenient biometric technologies, particularly facial recognition, has become an important development direction for intelligent forklift identity verification.

[0003] Currently, there are some forklift authentication systems on the market based on radio frequency identification (RFID) technology, but the following problems still exist: the only way to verify the driver's identity is still limited to cards, lacking multiple security protections; the deployment of existing facial recognition systems in embedded environments is relatively complex and consumes a lot of computing resources; and there is a lack of an authentication scheme that is both efficient and low-power and does not rely on complex hardware.

[0004] Current forklift authentication methods involving RFID technology include: Patent application number 202411134411.X, entitled "An Invention Patent for a Non-Contact Identification Method Based on Infrared RFID PAD Terminals," which relies solely on RFID tags for authentication. While this simplifies the process, it also results in low system security. If the RFID tag is copied or stolen, attackers can easily impersonate legitimate users, posing a serious security risk. Its limitations demonstrate that a single RFID authentication method is significantly insufficient for scenarios with high security requirements.

[0005] On the other hand, forklift authentication methods involving facial recognition algorithms include: Patent application number 202420865728.X, entitled "An Invention Patent for a Bluetooth-based Facial Recognition Forklift Control System." Although it achieves communication with the instrument cluster via a Bluetooth module, it requires the operator to carry a specific mobile terminal for Bluetooth pairing each time the forklift is started. This design significantly increases the system's complexity and dependency in practical applications, especially in work environments where forklifts are frequently used. It is not only time-consuming but also prone to misoperation.

[0006] In summary, current forklift authentication technologies have the following shortcomings: (1) they rely on a single authentication method and lack multiple security protection measures; (2) facial recognition systems consume significant computing resources in embedded environments, making it difficult to meet low power consumption requirements. Therefore, developing a forklift authentication method that combines facial recognition and RFID technology can simultaneously improve system security and efficiency while reducing power consumption, and has significant market demand and application prospects. This method can address the shortcomings of current technologies, and has broad application value, especially in high-security and high-reliability application scenarios. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, the present invention aims to provide a dual verification device and method for driver identity in industrial vehicles. This method combines lightweight facial recognition and radio frequency identification technologies, which can solve the defects in existing technologies and provide a safe, efficient, and embedded solution.

[0008] A dual verification device for driver identity in industrial vehicles includes a core processing module, an IC card sensing module, a camera module, a display module, and a relay module.

[0009] The IC card sensing module is responsible for reading the driver's identity card and obtaining its unique identifier; the camera module captures images and transmits the facial image as input data to the main control chip in the core processing module, where it is processed using a deep learning algorithm to determine identity; the display module displays the images captured by the camera and the system's verification status in real time, providing feedback to the driver; the relay module controls the start and stop of the industrial vehicle via control signals; the core processing module processes the data from the IC card sensing module and the camera module, runs a lightweight facial recognition algorithm, and achieves identity verification by combining the authentication information from both; after identity verification, the main control chip sends a signal to the relay module to control the start or stop of the industrial vehicle.

[0010] A method for dual authentication of driver identity for industrial vehicles, based on a dual authentication device for driver identity for industrial vehicles, includes the following steps:

[0011] Step 1: Construct a face recognition model. The specific steps are as follows:

[0012] Step 1.1: Employ a two-stage face detection and feature point extraction method, specifically: n frames of images captured from the camera, denoted as F... i Let i = 1, 2, ..., n, and let the existing face detection model based on a lightweight SSD structure be passed in to recognize the input image F. i From the potential face regions in the image, m candidate rectangular boxes for faces are obtained, denoted as B. ijLet j = 1, 2, ..., m, and set the face candidate bounding box B. ij The input is fed into the existing MTCNN face fine detection module to confirm the final bounding box B of the face location. i Simultaneously, four key points are extracted from the face: the left eye, the right eye, the tip of the nose, and the corner of the mouth. Based on the positions of these four key points, the detected face is aligned.

[0013] Step 1.2: Optimize the lightweight face recognition neural network GhostFaceNets, specifically as follows:

[0014] Step 1.2.1: Change the fixed value r in the traditional phantom module and dynamically adjust r according to different stages. Assign 0.8 to the shallow layer near the input, 0.7 to the middle layer, and 0.6 to the deep layer near the output. The phantom module is a module proposed in GhostNet. r is an internal hyperparameter in the design of the neural network, as shown in Equation (1), which represents the proportion of output features generated by cheap operations.

[0015]

[0016] Where, n out n represents the total number of output features. ghost n represents the number of features generated by the cheap operation. primary This represents the number of features generated by direct convolution;

[0017] Step 1.2.2: Simplify the attention mechanism, specifically by removing the deep convolution-fully connected attention mechanism in GhostFaceNets to pursue inference speed on embedded devices. At the same time, replace the fully connected layer in the SE module with a 1×1 convolution according to Equation (2), and then expand it after reducing the number of channels.

[0018] SE(X)=σ(Conv1×1 expand (δ(Conv1×1 reduce (GAP(X))))) (2)

[0019] Where X represents the input feature map, σ represents Sigmoid normalization, δ represents the PReLU activation function, GAP represents the global average pooling layer, and Conv1×1 reduce This indicates that the number of channels is reduced using a 1×1 convolution, Conv1×1 expand This represents a 1×1 convolution that recovers the number of channels;

[0020] Step 1.2.3: Using the improved loss function, according to formula (3), the similarity is calculated by selecting multiple center points for each face identity. When calculating the loss, the center point closest to the sample is selected. The final loss value is calculated according to formula (4), and the selected center point is substituted into the formula.

[0021]

[0022] Where N represents the batch size, m represents the angular interval, s represents the scaling factor, K represents the subcenter of the current category, k represents the index of all subcenters, and y i This indicates the corresponding category of the input sample. This represents the input sample and its corresponding category y. i The angle θ between the weight vectors j This represents the angle between the input sample and the weight vector of the j-th class;

[0023] Step 1.3: Train the optimized face recognition model from Step 1.2 on a publicly available face recognition dataset. This includes the following steps:

[0024] Step 1.3.1: Obtain face images from the public face recognition dataset, perform alignment processing on the input face images, then perform normalization processing, and then uniformly crop the face image size to 112×112;

[0025] Step 1.3.2: Train the model using a loss function with angular intervals. In this case, multiple sub-centers are set for each category. During training, the center with the smallest angle to the input sample features is selected and the loss is calculated by substituting it into the formula, as shown in equations (3) and (4).

[0026] Step 1.3.3: Reduce computational resource consumption without compromising model performance by using a mixed-precision calculation method of FP16 and FP32, and use L2 normalization as a feature output constraint;

[0027] Step 2: Quantize and deploy the face recognition model. The specific steps are as follows:

[0028] Step 2.1: The steps for quantizing the face recognition model are as follows: First, save the trained face recognition model as a standard Keras model format (.h5). Then, use a model conversion tool to convert it to a deep learning model exchange format (.onnx) that supports multiple hardware platforms. Utilize the ESP-DL toolchain to quantize the ONNX model, converting the original 32-bit floating-point weights and activation values ​​to 8-bit fixed-point integer values. This significantly reduces model storage requirements and computational overhead. The quantization process employs a symmetric quantization mechanism, converting the floating-point weights x... float Convert to integer weight x int The specific conversion method is shown in equation (5);

[0029]

[0030] Where Δ represents the scaling factor and z represents the zero-point offset;

[0031] Step 2.2: The specific steps for model deployment are as follows: The model format is converted to C language format, generating two files: an .hpp file and a .cpp file. The .hpp file contains the declaration of parameters and structure of each layer of the model, and the .cpp file stores the specific weights and computation graph data of the model, which is convenient for loading in the embedded environment. The above model code file is merged with the target embedded project code. The program is written into the core processing module through compilation and burning tools. The inference engine running on the chip can perform real-time face recognition inference on the input face image.

[0032] Step 3: The driver's identity dual verification process is as follows:

[0033] Step 3.1: Bluetooth communication completes identity registration, as follows: The IC card sensing module obtains the driver's unique IC card identifier (UID), and then the camera module captures the driver's facial image and performs feature extraction. The main control chip receives the driver's identity information sent by the management terminal, including name and ID card number, through low-power Bluetooth feature value communication, and binds the above information and stores it in non-volatile memory NVS, establishing a mapping relationship between UID-facial feature-ID card number-name;

[0034] Step 3.2: IC card identity verification, the specific steps are as follows: The driver swipes the card to verify their identity, the IC card sensing module reads the card UID and transmits the UID data to the core processing module main control chip, the chip searches the list of registered UIDs in the local NVS storage; if the UID matches successfully, the system enters the face recognition process; if the UID is not registered, the identity registration process is started; if the UID verification fails, a prompt message is displayed, requiring the card to be swiped again;

[0035] Step 3.3: The specific steps of face recognition verification are as follows: After the IC card verification is successful, the camera module starts face image acquisition. The acquired face image is input into the quantized face recognition model. The core processing module performs inference operations, extracts face features and matches them with registered face features. The matching process uses cosine similarity calculation to compare the current input features with the reference features bound to the UID. If the similarity exceeds the set threshold, face recognition is successful.

[0036] Step 3.4: Control execution. Specifically, when both IC card verification and facial recognition are successful, the core processing module sends a control command to the relay module to drive the relay to engage and complete the industrial vehicle start-up operation.

[0037] By employing the above-described technology, the beneficial effects of the present invention compared to the prior art are as follows:

[0038] This invention integrates lightweight face recognition algorithms and radio frequency identification (RFID) technology, proposing a dual forklift authentication method suitable for embedded devices. It adopts an improved network based on GhostFaceNets to achieve low-power, high-precision face recognition, and introduces an optimized loss function to improve the discrimination ability of the small model. The model is compressed and deployed in the main control chip through post-training quantization technology to achieve local real-time inference. At the same time, the driver's identity information is registered and bound through the Bluetooth interface. This invention effectively solves the problems of insufficient security, high resource consumption, and complex deployment in existing solutions, significantly improving the security, reliability, and application flexibility of forklift authentication in industrial scenarios, and adapting to the needs of high personnel mobility and high authentication frequency in industrial scenarios.

[0039] This invention integrates lightweight face recognition algorithms with radio frequency identification (RFID) technology. By optimizing the GhostFaceNets network structure, improving the attention mechanism, and introducing multiple subcenters to improve the loss function, it enhances the recognition accuracy and speed of small models on embedded devices. The system adopts FP16 / FP32 mixed precision and 8-bit quantization deployment, significantly reducing computing resources and power consumption. A Bluetooth registration and binding process is designed to achieve local pairing and verification of the driver's IC card UID with facial features and identity information, thereby enhancing the flexibility and industrial adaptability of device deployment while ensuring system security. Attached Figure Description

[0040] Figure 1 A physical diagram of the verification device.

[0041] Figure 2 Images captured by the camera;

[0042] Figure 3 Face images from a face recognition dataset;

[0043] Figure 4 for Figure 3 Face image after face alignment processing;

[0044] Figure 5 for Figure 4 A face image after normalization;

[0045] Figure 6 for Figure 5 A human face image after cropping;

[0046] Figure 7 The facial image captured by the camera. Detailed Implementation

[0047] The following detailed description of the dual verification device and method for the identity of industrial vehicle drivers of the present invention, with reference to specific embodiments, should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the present invention.

[0048] The present invention relates to a dual verification device and method for industrial vehicle drivers. First, it integrates a core processing module, an IC card sensing module, a camera module, a display module, and a relay module. It acquires the driver's facial image and performs face detection and alignment using a lightweight SSD and MTCNN. An optimized GhostFaceNets model is constructed, and an optimized loss function is used to improve embedded recognition performance. Post-training quantization technology is then employed to compress the model into an 8-bit format and deploy it to the chip. Driver identity information is registered via Bluetooth, binding the IC card UID, facial features, name, and ID number. Finally, the system achieves forklift control through dual verification of the IC card and face, balancing safety and deployment efficiency.

[0049] A dual verification device for driver identity in industrial vehicles, the physical construction diagram of which is shown below. Figure 1 As shown, it consists of a core processing module, an IC card sensing module, a camera module, a display module, and a relay module. The IC card sensing module is responsible for reading the driver's identification card and obtaining the card's unique identifier; the camera module captures images, such as... Figure 2 As shown, the system transmits facial images as input data to the main control chip, where they are processed using a deep learning algorithm to determine identity. The display module displays the images captured by the camera and the system's verification status in real time, providing feedback to the driver. The relay module controls the start and stop of the industrial vehicle via control signals. The core processing module processes data from the IC card sensing module and the camera module, runs a lightweight facial recognition algorithm, and achieves identity verification by combining the authentication information from both. After identity verification, the main control chip sends a signal to the relay module to control the start or stop of the industrial vehicle.

[0050] A dual authentication method for drivers of industrial vehicles includes the following steps:

[0051] Step 1: Construct a face recognition model. The specific steps are as follows:

[0052] Step 1.1: Employ a two-stage face detection and feature point extraction method, specifically: n frames of images captured from the camera, denoted as F... i Let i = 1, 2, ..., n, and let i be the face detection model optimized based on a lightweight SSD structure to recognize the input image F. i From the potential face regions in the image, m candidate rectangular boxes for faces are obtained, denoted as B. ij Let j = 1, 2, ..., m, and set the face candidate bounding box B.ij The data is input into the MTCNN's fine-grained face detection module to confirm the final bounding box B of the face location. i Simultaneously, four key points are extracted from the face: the left eye, the right eye, the tip of the nose, and the corner of the mouth. Based on the positions of these four key points, the detected face is aligned.

[0053] Step 1.2: Optimize the lightweight face recognition neural network GhostFaceNets, specifically as follows:

[0054] Step 1.2.1: Change the fixed value r in the traditional phantom module and dynamically adjust r according to different stages. Assign 0.8 to the shallow layer near the input, 0.7 to the middle layer, and 0.6 to the deep layer near the output. The phantom module is a module proposed in GhostNet. r is an internal hyperparameter in the design of the neural network, as shown in Equation (1). It determines how much of the output features are generated by cheap operations.

[0055]

[0056] Where, n out n represents the total number of output features. ghost n represents the number of features generated by the cheap operation. primary This represents the number of features generated by direct convolution;

[0057] Step 1.2.2: Simplify the attention mechanism, specifically by removing the deep convolution-fully connected attention mechanism in GhostFaceNets to pursue inference speed on embedded devices. At the same time, replace the fully connected layer in the SE module with a 1×1 convolution according to Equation (2), and then expand it after reducing the number of channels.

[0058] SE(X)=σ(Conv1×1 expand (δ(Conv1×1 reduce (GAP(X))))) (2)

[0059] Where X represents the input feature map, σ represents Sigmoid normalization, δ represents the PReLU activation function, GAP represents the global average pooling layer, and Conv1×1 reduce This indicates that the number of channels is reduced using a 1×1 convolution, Conv1×1 expand This represents a 1×1 convolution that recovers the number of channels;

[0060] Step 1.2.3: Using the improved loss function, according to formula (3), the similarity is calculated by selecting multiple center points for each category. When calculating the loss, the center point closest to the sample is selected; the final loss value is calculated according to formula (4), and the selected center point is substituted into it;

[0061]

[0062]

[0063] Where N represents the batch size, m represents the angular interval, s represents the scaling factor, K represents the subcenter of the current category, k represents the index of all subcenters, and y i This indicates the corresponding category of the input sample. This represents the input sample and its corresponding category y. i The angle θ between the weight vectors j This represents the angle between the input sample and the weight vector of the j-th class;

[0064] Step 1.3: Train the optimized face recognition model from Step 1.2 on a publicly available face recognition dataset. This includes the following steps:

[0065] Step 1.3.1: As Figure 3 The image shown is used to obtain face images from a publicly available face recognition dataset. Figure 4 The input face image is aligned as shown, and then, as... Figure 5 The image is then normalized as shown, and then the face image is processed as follows. Figure 6 The dimensions shown are uniformly cut to 112×112;

[0066] Step 1.3.2: Train the model using a loss function with angular intervals. In this case, multiple sub-centers are set for each category. During training, the center with the smallest angle to the input sample features is selected and the loss is calculated by substituting it into the formula, as shown in equations (3) and (4).

[0067] Step 1.3.3: Reduce computational resource consumption without compromising model performance by using a joint FP16 and FP32 computation method, and use L2 normalization as a feature output constraint;

[0068] Step 2: Quantize and deploy the face recognition model. The specific steps are as follows:

[0069] Step 2.1: The steps for quantizing the face recognition model are as follows: First, save the trained face recognition model as a standard Keras model format (.h5). Then, use a model conversion tool to convert it to a deep learning model exchange format (.onnx) that supports multiple hardware platforms. Utilize the ESP-DL toolchain to quantize the ONNX model, converting the original 32-bit floating-point weights and activation values ​​to 8-bit fixed-point integer values. This significantly reduces model storage requirements and computational overhead. The quantization process employs a symmetric quantization mechanism, converting the floating-point weights x... float Convert to integer weight x int The specific conversion method is shown in equation (5);

[0070]

[0071] Where Δ represents the scaling factor and z represents the zero-point offset;

[0072] Step 2.2: The specific steps for model deployment are as follows: The model format is converted to C language format, generating two files: an .hpp file and a .cpp file. The .hpp file contains the declaration of parameters and structure of each layer of the model, and the .cpp file stores the specific weights and computation graph data of the model, which is convenient for loading in the embedded environment. The above model code file is merged with the target embedded project code. The program is written into the core processing module through compilation and burning tools. The inference engine running on the chip can perform real-time face recognition inference on the input face image.

[0073] Step 3: The driver's identity dual verification process is as follows:

[0074] Step 3.1: Bluetooth communication completes identity registration, as follows: The IC card sensing module obtains the driver's unique IC card identifier (UID), and then the camera module captures the driver's facial image and performs feature extraction. The main control chip receives the driver's identity information sent by the management terminal, including name and ID card number, through low-power Bluetooth feature value communication, and binds the above information and stores it in non-volatile memory NVS, establishing a mapping relationship between UID-facial feature-ID card number-name;

[0075] Step 3.2: IC card identity verification, the specific steps are as follows: The driver swipes the card to verify their identity, the IC card sensing module reads the card UID and transmits the UID data to the core processing module main control chip, the chip searches the list of registered UIDs in the local NVS storage; if the UID matches successfully, the system enters the face recognition process; if the UID is not registered, the identity registration process is started; if the UID verification fails, a prompt message is displayed, requiring the card to be swiped again;

[0076] Step 3.3: The specific steps of face recognition verification are as follows: After the IC card verification is successful, the camera module starts face image acquisition, such as... Figure 7 As shown, the acquired face image is input into the quantized face recognition model, and the core processing module performs inference operations to extract face features and match them with registered face features. The matching process uses cosine similarity calculation to compare the current input features with the reference features bound to the UID. If the similarity exceeds the set threshold, the face recognition is successful.

[0077] Step 3.4: Control execution. Specifically, when both IC card verification and facial recognition are successful, the core processing module sends a control command to the relay module to drive the relay to engage and complete the industrial vehicle start-up operation.

[0078] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms described in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for dual authentication of driver identity for industrial vehicles, based on a dual authentication device for driver identity for industrial vehicles, wherein the device includes: The core processing module, IC card sensing module, camera module, display module, and relay module are all included. The IC card sensing module is responsible for reading the driver's identification card and obtaining the card's unique identifier; The camera module captures images and transmits the facial images as input data to the main control chip in the core processing module, where they are processed using deep learning algorithms to determine identity. The display module is used to display images captured by the camera and the system's verification status in real time, providing feedback to the driver's operation. The relay module controls the start and stop of industrial vehicles via control signals; The core processing module is responsible for processing data from the IC card sensing module and the camera module, running a lightweight face recognition algorithm, and achieving identity verification by combining the authentication information from both. After identity verification, the main control chip sends a signal to the relay module to control the start or stop of the industrial vehicle. The method is characterized by comprising the following steps: Step 1: Construct a face recognition model. The specific steps are as follows: Step 1.1: Employ a two-stage face detection and feature point extraction method, specifically: n frames of images captured from the camera, denoted as F... i Let i = 1, 2, ..., n, and let the existing face detection model based on a lightweight SSD structure be passed in to recognize the input image F. i From the potential face regions in the image, m candidate rectangular boxes for faces are obtained, denoted as B. ij Let j=1,2,…,m, and select the face candidate bounding box B. ij The fine-grained face detection module input into the MTCNN model confirms the final bounding box B of the face location. i Simultaneously, four key points are extracted from the face: the left eye, the right eye, the tip of the nose, and the corner of the mouth. Based on the positions of these four key points, the detected face is aligned. Step 1.2: Optimize the lightweight face recognition neural network GhostFaceNets, specifically as follows: Step 1.2.1: Change the fixed value r in the phantom module of GhostNet and dynamically adjust r according to different stages. r is an internal hyperparameter during neural network design, as shown in Equation (1), which represents the proportion of output features generated through cheap operations. (1)； Where, n out n represents the total number of output features. ghost n represents the number of features generated by the cheap operation. primary This represents the number of features generated by direct convolution; Step 1.2.2: Simplify the attention mechanism, specifically by deleting the depthwise convolution-fully connected attention mechanism in GhostFaceNets. At the same time, replace the fully connected layer in the SE module with a 1×1 convolution according to Equation (2), and then expand it after reducing the number of channels. (2)； Where X represents the input feature map. This indicates Sigmoid normalization. This represents the PReLU activation function, and GAP represents the global average pooling layer. This means using a 1×1 convolution to reduce the number of channels. This represents a 1×1 convolution that recovers the number of channels; Step 1.2.3: Using an improved loss function, according to equation (3), the similarity is calculated by selecting multiple center points for each face identity. When calculating the loss, the center point closest to the sample is selected. The final loss value is calculated according to equation (4), and the selected center point is substituted into the equation. (3)； (4)； Where N represents the batch size, m represents the angular interval, s represents the scaling factor, K represents the subcenter of the current category, k represents the index of all subcenters, and y i This indicates the corresponding category of the input sample. This represents the input sample and its corresponding category y. i The angle between the weight vectors, This represents the angle between the input sample and the weight vector of the j-th class; Step 1.3: Train the optimized lightweight face recognition model GhostFaceNets from Step 1.2 on a public face recognition dataset. This includes the following steps: Step 1.3.1: Obtain face images from the public face recognition dataset, perform alignment processing on the input face images, then perform normalization processing, and then uniformly crop the face image size to 112×112; Step 1.3.2: Train the model using a loss function with angular intervals. In this case, multiple sub-centers are set for each category. During training, the center with the smallest angle to the input sample features is selected and the loss is calculated by substituting it into the formula, as shown in equations (3) and (4). Step 1.3.3: Reduce computational resource consumption without compromising model performance by using a mixed-precision calculation method of FP16 and FP32, and use L2 normalization as a feature output constraint; Step 2: Quantize and deploy the face recognition model. The specific steps are as follows; Step 2.1: The steps for quantizing the face recognition model are as follows: First, save the face recognition model completed in training step 1.3 as a standard Keras model format (.h5). Then, use a model conversion tool to convert it to a deep learning model exchange format (.onnx) that supports multiple hardware platforms. Utilize the ESP-DL toolchain to quantize the ONNX model, converting the original 32-bit floating-point weights and activation values ​​into 8-bit fixed-point integer values. This significantly reduces model storage requirements and computational overhead. The quantization process employs a symmetric quantization mechanism, converting the floating-point weights x... float Convert to integer weight x int The specific conversion method is shown in equation (5); (5)； in, This represents the scaling factor, and z represents the zero-point offset; Step 2.2: The specific steps for model deployment are as follows: The model format is converted to C language format, generating two files: an .hpp file and a .cpp file. The .hpp file contains the declaration of parameters and structure of each layer of the model, and the .cpp file stores the specific weights and computation graph data of the model, which is convenient for loading in the embedded environment. The above model code file is merged with the target embedded project code. The program is written into the core processing module through compilation and burning tools. The inference engine running on the chip can perform real-time face recognition inference on the input face image. Step 3: The driver's identity dual verification process is as follows, and the specific steps are as follows; Step 3.1: Bluetooth communication completes identity registration, as follows: The IC card sensing module obtains the driver's unique IC card identifier UID, and then the camera module captures the driver's face image and performs feature extraction. The main control chip receives the driver's identity information sent by the management terminal through low-power Bluetooth feature value communication, including name and ID card number, and binds the above information and stores it in non-volatile memory NVS, establishing a mapping relationship between UID-face feature-ID card number-name; Step 3.2: IC card identity verification, the specific steps are as follows: The driver swipes the card to verify their identity, the IC card sensing module reads the card UID and transmits the UID data to the core processing module main control chip, the chip searches the list of registered UIDs in the local NVS storage; if the UID matches successfully, the system enters the face recognition process; if the UID is not registered, the identity registration process is started; if the UID verification fails, a prompt message is displayed, requiring the card to be swiped again; Step 3.3: The specific steps of face recognition verification are as follows: After the IC card verification is successful, the camera module starts face image acquisition. The acquired face image is input into the quantized face recognition model. The core processing module performs inference operations, extracts face features and matches them with registered face features. The matching process uses cosine similarity calculation to compare the current input features with the reference features bound to the UID. If the similarity exceeds the set threshold, face recognition is successful. Step 3.4: Control execution. Specifically, when both IC card verification and facial recognition are successful, the core processing module sends a control command to the relay module to drive the relay to engage and complete the industrial vehicle start-up operation.

Images