Security control method and device based on environment perception and electronic equipment

Abstract

The application discloses a security control method and device based on environment perception and electronic equipment, and relates to the technical field of Internet of Things. The method comprises the following steps: collecting environment characteristic parameters, determining at least two target identification models and target biological feature data types to be activated from a model configuration library; collecting target biological feature data, obtaining an available resource pool composed of idle computing resources of each edge computing node; calculating the inference complexity according to the model parameter quantity and data quantity of each model, and allocating target hardware resources for each model in the available resource pool; performing inference in a parallel mode to obtain an identity recognition result and a confidence score; determining the weight coefficient of each identity recognition result by using the environment characteristic parameters, obtaining a comprehensive confidence score through weighted fusion, determining a final identity recognition result, and controlling the access control device to perform a passing control action. The application can improve the adaptability and accuracy of the security system in a complex environment and reduce the multi-modal comprehensive inference delay.

Description

Technical Field

[0001] This application relates to the field of Internet of Things (IoT) technology, and in particular to a security control method and electronic device based on environmental perception. Background Technology

[0002] With the continuous advancement of smart park construction, security systems, as the infrastructure for park security management, are receiving increasing attention for their level of intelligence. A typical park security system usually includes various types of security equipment, such as video surveillance equipment, access control equipment, and fire alarm equipment. Among them, access control equipment, as a key node controlling personnel access, requires biometric identification technology to verify the identity of those passing through. Biometric identification technology utilizes inherent physiological or behavioral characteristics of the human body for identity verification, offering high security and convenience, and has become the mainstream verification method for modern access control systems. During peak commuting hours in the park, access control equipment needs to process a large number of identity verification requests from people passing through in a short period of time, thus placing high demands on both recognition speed and accuracy.

[0003] Currently, the identity recognition functions of access control devices in park security systems mostly rely on recognition algorithms deployed on cloud servers. Specifically, the data acquisition device deployed at the front end of the access control device uploads the collected video data and biometric data to the cloud server in real time via the network. The cloud server then calls the recognition algorithm to complete feature extraction and identity comparison, and returns the recognition results to the access control device to execute access control. However, this cloud-based centralized processing architecture has the following problems: First, both data uploading and result return rely on network transmission. When network conditions are poor or the data transmission volume is large, the transmission latency is high, making it difficult to meet the real-time requirement of second-level access control. Second, biometric data, including user facial features, needs to be transmitted over public networks, which carries the risk of interception or leakage during transmission, making it difficult to effectively protect user privacy. Third, when the network connection is interrupted, the access control device cannot upload data to the cloud server, the recognition function completely fails, and the overall system stability and availability are poor.

[0004] Furthermore, existing access control devices typically employ a single biometric identification modality for authentication, such as facial recognition or voiceprint recognition alone. The performance of a single identification modality is highly susceptible to environmental factors. In real-world park applications, lighting conditions, ambient noise, and other environmental factors are often dynamically changing. Existing access control devices based on a single identification modality struggle to maintain stable recognition accuracy under various environmental conditions, leading to reduced overall access efficiency and a poor user experience. Summary of the Invention

[0005] This application provides a security control method and electronic device based on environmental perception, which can solve the technical problems of insufficient recognition accuracy and large processing delay in existing security devices. The technical solution is as follows: In a first aspect, embodiments of this application provide a security control method based on environmental perception, the method comprising: In response to the detection of a target object entering the recognition area, environmental feature parameters are collected through an associated multi-sensor array; Based on the environmental feature parameters, at least two target recognition models to be activated are determined from the preset model configuration library, as well as the target biometric data types corresponding to each target recognition model; Based on the determined target biometric data type, collect the target biometric data corresponding to the target object; Obtain an available resource pool, which consists of the idle computing resources of each edge computing node in the security system; For each target recognition model, the reasoning complexity required for the target recognition model to complete the reasoning is calculated based on the number of its model parameters and the amount of corresponding target biometric data collected. Based on the reasoning complexity of each target recognition model, target hardware resources are allocated to each target recognition model from the available resource pool; The biometric data of each target are sent to the target hardware resources allocated to them, and the inference of each target recognition model is executed in parallel to obtain the identity recognition result and its confidence score output by each target recognition model. When there are multiple identity recognition results, the weight coefficient of each identity recognition result in the final decision is determined by using the environmental feature parameters and the confidence score corresponding to each identity recognition result. The confidence scores corresponding to each of the identity recognition results are weighted and fused using the weight coefficients to obtain a comprehensive confidence score, and the final identity recognition result is determined based on the comprehensive confidence score. Based on the final identity recognition result, the access control device associated with the edge computing node is controlled to perform corresponding access control actions.

[0006] Secondly, embodiments of this application provide a security control device based on environmental perception, the device comprising: The acquisition module is used to collect environmental feature parameters through an associated multi-sensor array in response to the detection of a target object entering the recognition area; The determination module is used to determine, based on the environmental feature parameters, at least two target recognition models to be activated from a preset model configuration library, and the target biometric data types corresponding to each target recognition model; The acquisition module is also used to collect target biometric data corresponding to the target object according to the determined target biometric data type; The acquisition module is also used to acquire an available resource pool, which consists of the idle computing resources of each edge computing node in the security system. The calculation module is used to calculate the reasoning complexity required for each target recognition model to complete the reasoning, based on the number of its model parameters and the amount of corresponding target biometric data collected. The allocation module is used to allocate target hardware resources to each target recognition model from the available resource pool according to the inference complexity of each target recognition model. The inference module is used to send the biometric data of each target to the target hardware resources allocated to them, and execute the inference of each target recognition model in parallel to obtain the identity recognition result and confidence score output by each target recognition model. The calculation module is also used to determine the weight coefficient of each identity recognition result in the final decision by using the environmental feature parameters and the confidence score corresponding to each identity recognition result when there are multiple identity recognition results. The weighting module is used to perform weighted fusion of the confidence scores corresponding to each of the identity recognition results using the weight coefficients to obtain a comprehensive confidence score, and to determine the final identity recognition result based on the comprehensive confidence score. The control module is used to control the access control device associated with the edge computing node to perform corresponding access control actions based on the final identity recognition result.

[0007] Thirdly, embodiments of this application provide a computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the above-described method steps.

[0008] Fourthly, embodiments of this application provide an electronic device that may include: a processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and to execute the above-described method steps.

[0009] The beneficial effects of the technical solutions provided in some embodiments of this application include at least the following: By dynamically determining at least two target recognition models to be activated and their corresponding target biometric data types from a preset model configuration library based on environmental characteristic parameters, the system can adaptively select the optimal combination of recognition models according to actual environmental conditions. This effectively avoids the technical problem of a significant decrease in recognition accuracy of a single biometric recognition modality under adverse environmental conditions such as insufficient light and noise interference, and can improve the recognition adaptability of security systems in complex and ever-changing environments.

[0010] By constructing an available resource pool composed of idle computing resources from various edge computing nodes in the security system, and allocating target hardware resources to each model in the available resource pool according to the inference complexity of each target recognition model, multimodal recognition tasks can achieve computing power collaboration and parallel inference among multiple edge computing nodes. Compared with the serial execution method of a single node, the end-to-end latency of multimodal integrated inference can be reduced, which helps to improve the real-time response performance of the access control system.

[0011] By dynamically adjusting the weight coefficients of each identity recognition result using environmental feature parameters, and combining the weighted fusion of the confidence scores corresponding to each identity recognition result to generate a comprehensive confidence score, the final decision-making mechanism can automatically amplify the decision weight of the recognition modality with strong environmental adaptability and suppress the interference of the recognition modality with weak environmental adaptability under complex environmental conditions. This helps to improve the accuracy and robustness of the final identity recognition result and can reduce the false recognition rate and the rejection rate. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the architecture of the security system provided in the embodiments of this application; Figure 2 This is a schematic flowchart of a security control method based on environmental perception provided in an embodiment of this application; Figure 3 This is a flowchart illustrating the calculation of reasoning complexity provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a security control device based on environmental perception provided in this application; Figure 5 This is a schematic diagram of a computer storage medium provided in this application; Figure 6This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0015] It should be noted that the security control method based on environmental perception provided in this application is generally executed by a diagnostic terminal, and correspondingly, the security control device based on environmental perception is generally installed in the target vehicle 102.

[0016] Figure 1 This is a schematic diagram of the architecture of a security system provided in an embodiment of this application. The security system includes: a server 1, security devices distributed in various locations in the park, each security device including an edge computing node and at least one security device, the edge computing node being used to control one or more security devices, each security device communicating based on a distributed soft bus, and each edge computing node having the same operating system installed. Figure 1 In the middle, the edge computing nodes are 21, 22 and 23 respectively, and the security device is an access control device. Edge computing node 21 controls security device 211 and security device 212.

[0017] Server 1 is the centralized management device for the security system, deployed in the central computer room of the park or connected to the cloud. Server 1 establishes communication connections with various security devices within the park via wired or wireless networks, using either Transmission Control Protocol (TCP) or User Datagram Protocol (UDP). Server 1's main functions include storing the authorized identity database, managing model configuration files for each identification model in the model configuration library, distributing model update data to edge computing nodes in each security device, and receiving access event records uploaded by each edge computing node. Server 1 can be a rack-mount server, tower server, or a cloud-based virtual server. Its processor configuration should meet the requirements of multi-processor concurrent data handling, with a memory capacity of at least 32 gigabytes. Solid-state drives or disk arrays should be used for storage to ensure fast data read / write response times. Server 1 also undertakes the registration management and status monitoring of each security device. When an edge computing node goes offline or experiences resource anomalies, Server 1 can issue a maintenance alarm to the system administrator.

[0018] Security equipment refers to field terminal devices deployed throughout the park within a security system. Multiple security devices are installed at key locations requiring identification and access control, such as park entrances / exits, building entrances, floor corridors, and computer room entrances. Each security device, as an independent field unit, comprises an edge computing node and at least one security device. The edge computing node and security device interact via an internal bus or short-range communication interface. The internal bus can be a serial peripheral interface bus or an integrated circuit bus, while the short-range communication interface can be a universal serial bus interface or a Bluetooth Low Energy communication interface. Multiple security devices can be integrated within a single security device; for example, a single security device may include a multi-sensor array and an access control device. The multi-sensor array collects environmental characteristic parameters and target biometric data, while the access control device executes access control actions. Communication between security devices within the park is based on a distributed soft bus. This distributed soft bus logically forms a peer-to-peer device cluster, allowing any security device's edge computing node to initiate resource query requests to other security devices' edge computing nodes to obtain idle computing resources, achieving cross-device computing power collaboration.

[0019] Edge computing nodes are the core processing units of security equipment, used to execute the environment-adaptive multimodal biometric recognition method provided in this application embodiment. The hardware form of the edge computing node can be an embedded ARM processor motherboard, an industrial control motherboard based on the x86 architecture, or an edge artificial intelligence computing module integrating a neural network processor. An edge computing node includes at least a central processing unit (CPU), a graphics processing unit (GPU), memory, and storage units. The CPU runs the operating system and schedules various task processes, the GPU accelerates the inference calculation of the target recognition model, the memory capacity is no less than 4 gigabytes, and the storage unit uses an embedded multimedia card or a solid-state drive to store model configuration libraries, authorized identity databases, and various preset threshold parameters. All edge computing nodes are uniformly installed with the same operating system, which can be an embedded Linux system or a real-time operating system. The identical operating system ensures consistency in resource management interfaces, communication protocol stacks, and task scheduling mechanisms across edge computing nodes, enabling the distributed soft bus to transparently schedule idle computing resources of each edge computing node, eliminating the need for developers to perform adaptive development for different edge computing nodes. The edge computing node also has a network communication module for exchanging data with server 1 and other edge computing nodes. The network communication module supports at least one of Ethernet communication and Wi-Fi communication.

[0020] Security devices are end-point execution or sensing devices in security equipment that directly interact with target objects and are controlled by edge computing nodes. Security devices can be categorized into sensing-type security devices and execution-type security devices based on their function. Sensing-type security devices are used to collect environmental characteristic parameters and target biometric data, specifically including multi-sensor arrays and camera modules. The multi-sensor array integrates at least a light sensor, an infrared sensor, a distance sensor, and a microphone array. The light sensor measures the light intensity within the recognition area and outputs a value in lux. The infrared sensor emits and receives infrared signals to detect whether a target object has entered the recognition area. The distance sensor measures the distance between the target object and the access control device. The microphone array collects environmental sound signals and voice data and outputs an environmental noise level in decibels. The camera module collects facial image data and gait sequence data; its image sensor type can be a complementary metal-oxide-semiconductor (CMOS) image sensor with a resolution of at least 1920 pixels multiplied by 1080 pixels. Execution-type security devices execute access control actions based on control commands issued by the edge computing node, specifically including access control devices and alarm devices. The access control device can be implemented as any of the following: a tripod turnout, a wing gate, or a swing gate. Its drive mechanism receives the opening and closing commands from the edge computing node, controlling the raising, lowering, opening, and closing of the gate arm. The access control device also includes a status feedback module to return the current opening and closing status of the gate to the edge computing node. The alarm device can be implemented as an audible and visual alarm. When the access permission level is no access permission, the edge computing node controls the alarm device to emit a buzzer and flash a red light to alert security personnel. The sensing and execution security devices in each security device can be physically integrated into one unit. For example, a multi-sensor array and camera module can be integrated into the upper part of the access control device's housing to form an integrated access control terminal, reducing the complexity of on-site wiring.

[0021] It should be understood that Figure 1 The number of edge computing nodes, security devices, and servers shown is for illustrative purposes only. Depending on implementation needs, there can be any number of diagnostic terminals, networks, and servers.

[0022] The following will be combined with the appendix Figure 2 This application provides a detailed description of a security control method based on environmental perception, as provided in its embodiments. The environmental perception-based security control device in these embodiments can be... Figure 1 The cloud server shown.

[0023] Please see Figure 2 This document provides a flowchart illustrating a security control method based on environmental perception, as described in an embodiment of this application. Figure 2 As shown, the method described in this application embodiment may include the following steps: S1. In response to detecting that a target object has entered the recognition area, environmental feature parameters are collected through the associated multi-sensor array.

[0024] Edge computing nodes use a multi-sensor array to perceive the environmental status in real time. This array includes at least a light sensor, an infrared sensor, and a microphone array, used to collect data on light intensity within the recognition area, the distance between the target object and the security device, and ambient noise levels in decibels, respectively. When a target object enters the recognition area, the infrared detection signal emitted by the infrared sensor is reflected and received by the target object, and the edge computing node generates a detection signal indicating that the target object has entered the recognition area. In response to this detection signal, the edge computing node triggers each sensor in the multi-sensor array to simultaneously collect environmental characteristic parameters at the current moment. Specifically, the light sensor measures and outputs light intensity in lux, the distance sensor measures the distance between the target object and the security device and outputs the distance value in centimeters, and the microphone array collects ambient sound and outputs ambient noise in decibels. The collected light intensity, distance, and ambient noise levels together constitute the environmental characteristic parameters.

[0025] In one example, in an access control system deployed at an office building entrance, when an employee approaches the access control and enters the recognition area within 3 meters of the device, the infrared sensor detects the employee's entry. This triggers a light sensor to measure the ambient light intensity at 500 lux, a distance sensor to measure the distance between the employee and the access control system at 2.5 meters, and a microphone array to measure the ambient noise at 45 decibels. These three parameters together constitute the environmental characteristic parameters of the current scene. Similarly, in a surveillance camera application scenario, when a visitor enters the recognition area covered by the surveillance camera, the infrared sensor detects the target's entry, triggers a light sensor to measure the current light intensity at 800 lux, a distance sensor to measure the distance between the visitor and the surveillance camera at 5 meters, and a microphone array to measure the ambient noise at 55 decibels. These parameters also constitute the environmental characteristic parameters in this scenario.

[0026] S2. Based on environmental characteristic parameters, determine at least two target recognition models to be activated from the preset model configuration library, as well as the target biometric data types corresponding to each target recognition model.

[0027] In this step, the edge computing node uses the environmental feature parameters collected in step S1 to adaptively select the recognition model by querying the model configuration library. The model configuration library is a mapping table pre-built in the local storage of the edge computing node, recording the correspondence between each biometric recognition model and its applicable environmental conditions. Specifically, the biometric recognition models stored in the model configuration library include four types: face recognition, gait recognition, voiceprint recognition, and infrared thermal imaging recognition. Each recognition model is associated with two types of information in the model configuration library. The first type of information is the target biometric data type corresponding to the recognition model; that is, face recognition model corresponds to face image data, gait recognition model corresponds to gait sequence data, voiceprint recognition model corresponds to speech data, and infrared thermal imaging recognition model corresponds to infrared thermal imaging image data. The second type of information is the applicable condition range of the recognition model in one or more environmental dimensions, which at least include the light intensity range, the ambient noise decibel range, and the target object distance range. For example, the facial recognition model is associated with an illumination intensity range of 200 lux to 2000 lux. Within this range, the model maintains high recognition accuracy. Below 200 lux, facial image details are insufficient; above 2000 lux, overexposure occurs, both leading to a significant drop in accuracy. The voiceprint recognition model is associated with an ambient noise decibel range of 0 to 55 decibels. Within this range, speech signal quality is good; above 55 decibels, noise interference intensifies, and recognition accuracy decreases accordingly. Gait recognition and infrared thermal imaging models show significant advantages in low-light conditions, both associated with an illumination intensity range of 0 to 100 lux. Furthermore, the infrared thermal imaging model is also associated with a target object distance range, with an effective recognition distance of 1 to 8 meters. Beyond this distance, the thermal imaging image resolution is insufficient to support reliable recognition.

[0028] When the edge computing node performs model selection, it first extracts three parameter values—light intensity, ambient noise (decibels), and target object distance—from the environmental feature parameters collected in step S1. Then, the edge computing node iterates through each recognition model stored in the model configuration library, comparing the applicable condition range of each recognition model in each environmental dimension with the corresponding parameter value in the current environmental feature parameters. A recognition model is determined to be a candidate model under the current environmental conditions if and only if the applicable condition range of a recognition model in all environmental dimensions includes the corresponding parameter value of the current environmental feature parameters. After the iteration is complete, the edge computing node selects at least two recognition models from the candidate model set as the target recognition models to be activated. If the number of candidate models that meet the applicable conditions exceeds two, selection is made according to a preset priority rule. The priority rule prioritizes the candidate model with the highest historical average confidence score, which is calculated from the confidence scores of each model in historical recognition tasks and stored in the model configuration library, provided that the applicable conditions are met. After selection, the edge computing node determines the target biometric data type corresponding to each target recognition model based on the correspondence recorded in the model configuration library, which serves as the basis for data collection in step S3.

[0029] S3. Based on the determined target biometric data type, collect the target biometric data corresponding to the target object.

[0030] According to the target biometric data types determined in step S2, the edge computing node calls the corresponding data acquisition device to acquire the biometric data of the target object. Specifically, when the target biometric data type is face image data, the edge computing node controls the camera to capture the target object's face and acquire image data containing the face region; when the target biometric data type is gait sequence data, the edge computing node controls the camera to continuously acquire gait sequence data of the target object walking within the recognition area; when the target biometric data type is voice data, the edge computing node controls the microphone to acquire the voice signal emitted by the target object; when the target biometric data type is infrared thermal imaging image data, the edge computing node controls the infrared thermal imaging sensor to acquire the thermal radiation distribution image of the target object; when the target biometric data type is iris image data, the edge computing node controls the iris acquisition camera to acquire the iris texture image of the target object. The acquired target biometric data has a data volume attribute, which includes at least one of data dimension, data frame number, and single frame data size.

[0031] S4. Obtain the available resource pool, which consists of the idle computing resources of each edge computing node in the security system.

[0032] Edge computing nodes send resource query requests to other edge computing nodes in the security system to obtain the computing resource usage status of each edge computing node. Edge computing nodes in the security system are associated with different types of security devices, such as edge computing nodes for access control devices, surveillance cameras, alarm hosts, and vehicle gate devices. Each edge computing node possesses certain computing resources, including a central processing unit (CPU), a graphics processing unit (GPU), and memory resources. When the CPU utilization of an edge computing node is lower than a preset utilization threshold and its available memory capacity is higher than a preset memory threshold, that edge computing node is considered to have idle computing resources. Edge computing nodes aggregate their own idle computing resources with those returned by other edge computing nodes to form an available resource pool, and configure a resource identifier for each idle computing resource in the available resource pool to locate and allocate the corresponding target hardware resources in subsequent steps. The preset utilization threshold can be set to 70%, and the preset memory threshold can be set to 512 megabytes; these thresholds can be flexibly adjusted according to the actual deployment situation.

[0033] S5. For each target recognition model, calculate the reasoning complexity required for the target recognition model to complete the reasoning based on the number of its model parameters and the amount of corresponding target biometric data collected.

[0034] In this step, the edge computing node calculates the inference complexity for each target recognition model. First, the edge computing node retrieves the model configuration file corresponding to the target recognition model from the model configuration library and parses the configuration file to obtain the number of model parameters. The number of model parameters includes the total number of parameters and the number of floating-point operations required for a single inference. Second, the edge computing node obtains the data volume of the corresponding target biometric data collected in step S3. The data volume includes at least one of the data dimension, number of data frames, and single-frame data size of the target biometric data. Finally, the edge computing node calculates the inference complexity based on the number of model parameters and the data volume. The inference complexity is the product of the number of floating-point operations and the data volume, or the weighted sum of the total number of parameters and the data volume. The inference complexity reflects the computational overhead required for the target recognition model to complete the inference task with the current data volume; a higher value indicates more computational resources required.

[0035] S6. Based on the reasoning complexity of each target recognition model, allocate target hardware resources to each target recognition model from the available resource pool.

[0036] Based on the inference complexity of each target recognition model calculated in step S5, the edge computing node matches suitable hardware resources for each target recognition model from the available resource pool constructed in step S4. Specifically, the edge computing node first determines the type and quantity of hardware resources required for each target recognition model based on the inference complexity. The hardware resource type includes at least one of a central processing unit, a graphics processing unit, and a neural network processor, and the resource requirement includes the required number of processor cores and the required memory capacity. Then, the edge computing node traverses all idle computing resources in the available resource pool and filters out candidate hardware resources that meet the hardware resource type and resource requirement. When the same idle computing resource meets the resource requirements of multiple target recognition models, it prioritizes allocating the idle computing resource to the target recognition model with the highest inference complexity. Finally, the edge computing node binds each target recognition model to the filtered candidate hardware resources, generates a mapping relationship between the model and the resource, and determines the candidate hardware resources as the target hardware resources corresponding to each target recognition model.

[0037] S7. Send the biometric data of each target to the target hardware resources allocated to them, execute the inference of each target recognition model in parallel, and obtain the identity recognition results and confidence scores output by each target recognition model.

[0038] Based on the model-resource mapping relationship established in step S6, the edge computing nodes send the biometric data of each target collected in step S3 to the corresponding target hardware resources. Upon receiving the data, each target hardware resource synchronously loads the corresponding target recognition model and executes inference. Each inference task is performed in parallel without blocking the others. After each target recognition model completes inference, it outputs an identity recognition result and its corresponding confidence score. The confidence score is a value between 0 and 1, representing the degree of certainty the target recognition model has about the recognition result; a higher value indicates higher confidence. Because each target recognition model executes in parallel, the latency of multimodal integrated inference is less than the sum of the inference latency when each model is executed serially.

[0039] S8. When there are multiple identity recognition results, the weight coefficient of each identity recognition result in the final decision is determined by using environmental feature parameters and the confidence score corresponding to each identity recognition result.

[0040] In this step, when multiple identity recognition results are obtained in step S7, the edge computing node uses environmental feature parameters to evaluate the environmental adaptability of each identity recognition result and determines the weight coefficient of each identity recognition result in the final decision. Specifically, the edge computing node first extracts the light intensity from the environmental feature parameters and assigns a basic weight to each identity recognition result based on the comparison between the light intensity and a preset light threshold. When the light intensity is greater than the preset high light threshold, the basic weight of the identity recognition result corresponding to the face recognition model is increased because the face recognition accuracy is higher under sufficient light conditions. When the light intensity is less than the preset low light threshold, the basic weight of the identity recognition result corresponding to the gait recognition model or the infrared thermal imaging recognition model is increased because gait recognition and infrared thermal imaging recognition are more stable under low light conditions. Secondly, the edge computing node extracts the environmental noise decibels from the environmental feature parameters and compares the environmental noise decibels with a preset noise threshold. When the environmental noise decibels are greater than the preset noise threshold, the basic weight of the identity recognition result corresponding to the voiceprint recognition model is adjusted downward because a high-noise environment will reduce the accuracy of voiceprint recognition. Finally, the edge computing nodes obtain the confidence scores corresponding to each identity recognition result, compare each confidence score with a preset confidence threshold, and for identity recognition results with confidence scores lower than the preset confidence threshold, adjust their corresponding weight coefficients by a preset ratio to obtain the final weight coefficients of each identity recognition result in the final decision.

[0041] S9. Use weighted coefficients to perform weighted fusion of the confidence scores corresponding to each identity recognition result to obtain a comprehensive confidence score, and determine the final identity recognition result based on the comprehensive confidence score.

[0042] In this step, the edge computing node uses the weight coefficients determined in step S8 to perform weighted fusion of the confidence scores corresponding to each identity recognition result. Specifically, the weighted fusion method involves summing the weight coefficients and corresponding confidence scores for each identity recognition result. The formula is that the overall confidence score equals the sum of the products of each weight coefficient and its corresponding confidence score. After obtaining the overall confidence score, the edge computing node compares it with a preset overall confidence threshold. When the overall confidence score is greater than the preset threshold, the current multimodal fusion recognition result is deemed reliable, and the identity recognition result with the highest confidence score among all identity recognition results is determined as the final identity recognition result. When the overall confidence score is less than or equal to the preset threshold, the current recognition result is deemed unreliable, triggering a re-recognition process or refusing passage.

[0043] S10. Based on the final identity recognition result, control the access control device associated with the edge computing node to perform the corresponding access control action.

[0044] In one possible embodiment, S4, obtaining the available resource pool, includes: S41. Send a resource query request to other edge computing nodes in the security system, excluding the edge computing nodes. The resource query request is used to request each of the other edge computing nodes to return its current processor utilization and available memory capacity.

[0045] In this step, this node initiates resource queries to other edge computing nodes through the communication network within the security system. The edge computing nodes in the security system are interconnected via wired or wireless LANs, forming a distributed computing cluster. This node maintains a list of all edge computing nodes in the system, recording the network address and node identifier of each edge computing node. When executing a resource query, this node iterates through all edge computing nodes in the node list except itself, sending resource query request data packets to each of the other edge computing nodes one by one. The resource query request data packet includes at least the node identifier of this node, a request timestamp, and a request type field, where the request type field indicates that the nature of this query request is a resource status query. After receiving the resource query request, each other edge computing node calls its own system monitoring interface to read the current processor utilization and available memory capacity, and encapsulates these two pieces of data into a resource query response data packet and returns it to this node. Processor utilization refers to the average utilization rate of all processor cores of the edge computing node at the current moment, expressed as a percentage, ranging from 0% to 100%. Available memory capacity refers to the total amount of physical memory that is not currently occupied by the edge computing node, measured in megabytes or gigabytes.

[0046] In one example, node A maintains a list of four other edge computing nodes: node B, node C, node D, and node E. Node A sends resource query request packets sequentially to nodes B, C, D, and E. Each packet carries node A's node identifier A001, a request timestamp of July 5, 2026, at 11:48:50, and a request type field "Resource Status Query." Upon receiving the request, node B immediately reads its processor utilization rate as 30% and its available memory capacity as 2 gigabytes, and encapsulates this data into a resource query response packet, returning it to node A. Nodes C, D, and E also perform the same operation, returning their respective processor utilization rates and available memory capacities to node A.

[0047] S42. Receive the processor utilization and available memory capacity returned by each of the other edge computing nodes, and determine the edge computing nodes whose processor utilization is lower than a preset utilization threshold and whose available memory capacity is higher than a preset memory threshold as target edge computing nodes with idle computing resources.

[0048] In this step, this node receives resource query response data packets from other edge computing nodes and parses out the processor utilization and available memory capacity of each edge computing node. This node compares the parsed values ​​with preset threshold conditions one by one. The preset utilization threshold is used to determine whether the processor is in a low-load state; this node uses a processor utilization rate lower than the preset utilization threshold as a condition for judging that the processor has idle computing power. The preset memory threshold is used to determine whether memory resources are sufficient; this node uses available memory capacity higher than the preset memory threshold as a condition for judging that memory resources are sufficient. Only when an edge computing node simultaneously meets both conditions—processor utilization rate lower than the preset utilization threshold and available memory capacity higher than the preset memory threshold—is this node identified as a target edge computing node, meaning it is considered to have idle computing resources available for use. For edge computing nodes that do not meet either condition, this node excludes them from the range of available resources and does not include them in the subsequent resource aggregation. The preset utilization threshold can be set to 70%, and the preset memory threshold can be set to 512 megabytes. These thresholds can be flexibly adjusted according to the actual hardware configuration and business load of the security system.

[0049] In one example, node A receives reports from node B of 30% processor utilization and 2 gigabytes of available memory; node C reports 45% processor utilization and 1.5 gigabytes of available memory; node D reports 85% processor utilization and 3 gigabytes of available memory; and node E reports 92% processor utilization and 256 megabytes of available memory. The preset utilization threshold is 70%, and the preset memory threshold is 512 megabytes. Node B's processor utilization (30%) is below 70%, and its available memory (2 gigabytes) is above 512 megabytes, satisfying both conditions; therefore, node B is identified as a target edge computing node. Node C's processor utilization (45%) is below 70%, and its available memory (1.5 gigabytes) is above 512 megabytes, also satisfying both conditions; therefore, node C is also identified as a target edge computing node. Node D's processor utilization (85%) is above 70%, failing to meet the processor utilization criteria, and is therefore excluded. Node E's processor utilization (92%) is higher than 70%, and its available memory capacity (256 MB / s) is lower than 512 MB / s; both conditions are not met, so it is excluded. Ultimately, the target edge computing nodes determined by node A are nodes B and C.

[0050] S43. The idle computing resources of the edge computing node itself and the idle computing resources of each target edge computing node are aggregated to generate the available resource pool, and a resource identifier is configured for each idle computing resource in the available resource pool. The resource identifier is used to locate and allocate the corresponding target hardware resources.

[0051] In this step, the node aggregates its own idle computing resources with those of the target edge computing nodes identified in step S42 to construct an available resource pool. First, the node calls its own system monitoring interface to obtain its processor utilization and available memory capacity. When its own processor utilization is lower than a preset utilization threshold and its available memory capacity is higher than a preset memory threshold, the node also includes its own idle computing resources in the available resource pool. Then, the node registers its own and each target edge computing node's idle computing resources item by item. The registration includes the node identifier of the edge computing node to which the idle computing resource belongs, the hardware resource type corresponding to the idle computing resource, the number of processor cores the hardware resource can provide, and the available memory capacity. The hardware resource type includes at least one of a central processing unit, a graphics processing unit, and a neural network processor. Finally, the node generates a globally unique resource identifier for each idle computing resource in the available resource pool. Each resource identifier corresponds one-to-one with an idle computing resource, enabling the rapid location and allocation of the corresponding target hardware resources in subsequent step S6 based on the resource identifier. Resource identifiers can be generated by combining node identifiers with hardware resource types, or by using an incremental sequence number, as long as they can uniquely distinguish each idle computing resource within the available resource pool.

[0052] In one example, node A has a processor utilization of 20% and 3 gigabytes of available memory, both meeting the idle criteria. Therefore, node A includes its own idle computing resources in the aggregation. Node A registers its idle computing resources as follows: node identifier A001, hardware resource type is a graphics processor, number of available processor cores is 8, available memory capacity is 3 gigabytes, and the allocated resource identifier is R001. Node A also registers the idle computing resources of the target edge computing node B as follows: node identifier B001, hardware resource type is a central processing unit, number of available processor cores is 4, available memory capacity is 2 gigabytes, and the allocated resource identifier is R002. Finally, node A registers the idle computing resources of the target edge computing node C as follows: node identifier C001, hardware resource type is a graphics processor, number of available processor cores is 6, available memory capacity is 1.5 gigabytes, and the allocated resource identifier is R003. The three idle computing resources mentioned above together constitute the available resource pool. Each idle computing resource in the available resource pool can be precisely located using resource identifiers R001, R002, and R003. In the subsequent step S6, when it is necessary to allocate target hardware resources for the face recognition model, node A locates its own graphics processor using resource identifier R001 and allocates it to the face recognition model as the target hardware resource. When it is necessary to allocate target hardware resources for the voiceprint recognition model, node A locates the central processing unit of node B using resource identifier R002 and allocates it to the voiceprint recognition model as the target hardware resource.

[0053] In one possible embodiment, see Figure 4 As shown, Figure 4 This is a schematic diagram of the calculation process for inference complexity provided in an embodiment of this application, specifically including the following steps: S51. Obtain the model configuration file corresponding to the target recognition model from the model configuration library, parse the model configuration file to obtain the number of model parameters of the target recognition model, including the total number of model parameters and the number of floating-point operations required for a single inference.

[0054] In this step, the edge computing node obtains the model parameter count for each target recognition model. A model configuration library is pre-stored in the local storage of the edge computing node. This library stores the applicable environmental conditions mapping relationships for each recognition model, as well as the corresponding model configuration file. The model configuration file records the architecture information, parameter information, and computational overhead information of the recognition model. Based on the target recognition model determined in step S2, the edge computing node reads the corresponding model configuration file from the model configuration library, parses the file, and extracts the model parameter count. The model parameter count includes two core metrics: the total number of parameters, i.e., the total number of trainable parameters constituting the recognition model, reflecting the model's size; and the number of floating-point operations required for a single inference, i.e., the number of floating-point operations required for the recognition model to perform a complete forward inference on a single input data set, reflecting the model's computational overhead. The total number of parameters and the number of floating-point operations together constitute a complete description of the model parameter count, providing basic data for calculating the inference complexity in the subsequent step S53.

[0055] In one example, after determining the activation of the face recognition model and the voiceprint recognition model in step S2, the edge computing node retrieves the model configuration files for these two models from the model configuration library. For the face recognition model, parsing its model configuration file reveals a total of 25 million parameters and 500 million floating-point operations per inference iteration. For the voiceprint recognition model, parsing its model configuration file reveals a total of 5 million parameters and 100 million floating-point operations per inference iteration. The difference in the number of model parameters reflects that the face recognition model has a larger model size and higher computational overhead than the voiceprint recognition model.

[0056] S52. Obtain the data volume of the corresponding target biometric data collected. The data volume includes at least one of the data dimension, data frame number, and single frame data size of the target biometric data.

[0057] In this step, the edge computing node obtains the data volume for each target biometric data collected in step S3. Data volume is a quantitative indicator measuring the size of the target biometric data, specifically including at least one of data dimension, number of data frames, and single-frame data size. Data dimension refers to the number of dimensions of the target biometric data in the feature space. For example, when face image data is represented in pixel matrix form, its data dimension is the image height multiplied by the image width multiplied by the number of color channels. The number of data frames refers to the number of consecutive frames contained in the target biometric data. For static image data, the number of frames is 1; for video sequences or gait sequences, the number of frames is the total number of consecutive image frames contained in the sequence. Single-frame data size refers to the number of bytes occupied by a single frame of data in the storage medium. The edge computing node selects one or more of the data dimension, number of data frames, and single-frame data size to characterize the data volume based on the specific type of the target biometric data, so as to provide accurate input parameters for calculating the inference complexity in the subsequent step S53.

[0058] In one example, the face image data corresponding to the face recognition model is a single-frame static image with one data frame and a single-frame data size of 2 megabytes (2,097,152 bytes). The data dimension is calculated as image height (1080 pixels) multiplied by image width (1920 pixels) multiplied by the number of color channels (3), resulting in 6,220,800 dimensions. For the speech data corresponding to the voiceprint recognition model, this data is a time-series signal with one data frame and a single-frame data size of 96 kilobytes (98,304 bytes). The data dimension is calculated as sampling rate (16,000 Hz) multiplied by duration (3 seconds), resulting in 48,000 dimensions. These differences in data volume and data dimension result in varying computational loads for different target recognition models during inference.

[0059] S53. Based on the number of model parameters and the amount of data, calculate the inference complexity of the target recognition model. The inference complexity is the product of the number of floating-point operations and the amount of data, or the weighted sum of the total number of parameters and the amount of data.

[0060] In this step, the edge computing node uses the number of model parameters obtained in step S51 and the amount of data obtained in step S52 to calculate the inference complexity of each target recognition model. Inference complexity is a comprehensive indicator that measures the computational cost required for a target recognition model to complete an inference task with a given amount of data. The higher the inference complexity value, the more computational resources the target recognition model needs to complete inference with the current amount of data. The edge computing node can calculate the inference complexity in two ways. The first way is to directly multiply the number of floating-point operations by the amount of data, that is, multiply the computational cost of a single inference by the data size to obtain the total computational cost. The second way is to perform a weighted summation of the total number of parameters and the amount of data, that is, assign weight coefficients to the total number of parameters and the amount of data respectively before summing. This method is suitable for scenarios where the number of model parameters and the amount of data have different degrees of influence on the inference cost. In both calculation methods, the number of floating-point operations, the total number of parameters, and the amount of data are represented using Arabic numerals, and multiplication and addition operations are described in written form. Based on the inference complexity of each target recognition model, the edge computing node can accurately determine the relative demand for hardware resources of each model, providing a quantitative basis for the reasonable allocation of target hardware resources within the available resource pool in step S6.

[0061] In one example, using the first calculation method, the face recognition model requires 500 million floating-point operations and 2 megabytes of data, resulting in an inference complexity of 500 million multiplied by 2 megabytes, or 1 multiplied by 10 to the power of 15 floating-point operations per byte. The voiceprint recognition model requires 100 million floating-point operations and 96 kilobytes of data, resulting in an inference complexity of 100 million multiplied by 96 kilobytes, or 9.6 multiplied by 10 to the power of 12 floating-point operations per byte. Comparing the two inference complexity values, the face recognition model's inference complexity is approximately 104 times that of the voiceprint recognition model, indicating that the face recognition model requires more computing resources. Therefore, in step S6, high-performance hardware resources such as graphics processing units (GPUs) should be prioritized for the face recognition model, while basic hardware resources such as central processing units (CPUs) should be allocated to the voiceprint recognition model. If the second calculation method is adopted, and the weighting coefficient of the total number of parameters is set to 0.6 and the weighting coefficient of the data volume is set to 0.4, then the inference complexity of the face recognition model is equal to 0.6 multiplied by 25 million plus 0.4 multiplied by 2 trillion, which is 15 million plus 0.8 trillion. The inference complexity of the voiceprint recognition model is equal to 0.6 multiplied by 5 million plus 0.4 multiplied by 96 thousand, which is 3 million plus 38.4 thousand. There is also a significant difference between the two, and the conclusion of the resource allocation strategy is consistent with the first method.

[0062] In one possible embodiment, allocating target hardware resources from the available resource pool for each target recognition model based on the inference complexity of each target recognition model includes: S61. Based on the inference complexity of each target recognition model, determine the type and amount of hardware resources required for each target recognition model. The type of hardware resources includes at least one of a central processing unit, a graphics processing unit, and a neural network processor. The amount of resources required includes the number of processor cores and the memory capacity required.

[0063] Based on the inference complexity of each target recognition model calculated in step S5, the edge computing nodes determine the specific hardware resource requirements of each target recognition model. Inference complexity is a comprehensive indicator that measures the computational overhead required for a target recognition model to complete its inference task; a higher value indicates higher computational and memory requirements. The edge computing nodes pre-maintain a mapping strategy between inference complexity and hardware resource requirements, specifying recommended hardware resource types and resource requirements for different levels of inference complexity. Specifically, when the inference complexity of a target recognition model is below a first preset threshold, the model is classified as a lightweight inference task, requiring only CPU resources (2 cores) and 256 megabytes of memory. When the inference complexity is between the first and second preset thresholds, the model is classified as a medium-level inference task, requiring recommended GPU resources (4 cores) and 1 gigabyte of memory. When the inference complexity exceeds the second preset threshold, the target recognition model is classified as a heavyweight inference task, and it is recommended to allocate neural network processor resources or high-performance graphics processor resources, requiring 8 processor cores and at least 2 gigabytes of memory. The specific values ​​of the first and second preset thresholds can be flexibly adjusted and calibrated according to the hardware configuration of the edge computing nodes in the security system.

[0064] In one example, the inference complexity of the face recognition model is 1 × 10¹⁵ floating-point operations per second, exceeding the second preset threshold, classifying it as a heavyweight inference task. The edge computing node determines that the required hardware resource type is a graphics processing unit (GPU), requiring 8 processor cores and 2 gigabytes of memory. The inference complexity of the voiceprint recognition model is 9.6 × 10¹² floating-point operations per second, falling between the first and second preset thresholds, classifying it as a medium-level inference task. The edge computing node determines that the required hardware resource type is a central processing unit (CPU), requiring 4 processor cores and 512 megabytes of memory. The inference complexity of the gait recognition model is 8 × 10¹³ floating-point operations per second, also classifying it as a medium-level inference task. The edge computing node determines that the required hardware resource type is a graphics processing unit (GPU), requiring 4 processor cores and 1 gigabyte of memory.

[0065] S62. Traverse each idle computing resource in the available resource pool, filter out candidate hardware resources that meet the hardware resource type and the resource requirement, and when the same idle computing resource meets the resource requirements of multiple target recognition models at the same time, give priority to allocating the idle computing resource to the target recognition model with the highest inference complexity.

[0066] The edge computing nodes iterate through and match the available resource pool constructed in step S4. Each idle computing resource in the available resource pool has corresponding resource attribute information, including hardware resource type, number of available processor cores, and available memory capacity. The edge computing nodes compare the hardware resource type and resource requirement determined for each target recognition model in sub-step S61 with the resource attribute information of each idle computing resource in the available resource pool. When the hardware resource type of an idle computing resource matches the hardware resource type required by the target recognition model, and the number of available processor cores of the idle computing resource is greater than or equal to the required number of processor cores, and the available memory capacity is greater than or equal to the required memory capacity, the idle computing resource is marked as a candidate hardware resource that meets the resource requirements of the target recognition model. During the iteration process, the same idle computing resource may simultaneously meet the resource requirements of multiple target recognition models. At this point, the edge computing nodes adopt a priority allocation strategy, sorting each target recognition model from high to low inference complexity, and prioritizing the allocation of the idle computing resources to the target recognition model with the highest inference complexity, so as to ensure that the inference task with the most urgent computing power demand receives priority resource guarantee.

[0067] In one example, the available resource pool contains three idle computing resources: a graphics processing unit (GPU) on node A with 16 available cores and 3 gigabytes of available memory; a central processing unit (CPU) on node B with 8 available cores and 2 gigabytes of available memory; and a GPU on node C with 4 available cores and 1 gigabyte of available memory. The edge computing nodes compare the requirements of the face recognition model with these three idle computing resources. They find that the GPU on node A meets the requirements, while the GPU on node C does not meet the requirements due to insufficient cores, and the CPU on node B does not meet the requirements due to hardware type mismatch. Therefore, the candidate hardware resource for the face recognition model is the GPU on node A. Similarly, comparing the requirements of the voiceprint recognition model with these three idle computing resources, they find that the CPU on node B meets the requirements, making it the candidate hardware resource for the voiceprint recognition model. Finally, comparing the requirements of the gait recognition model with these three idle computing resources, they find that the GPU on node C meets the requirements, making it the candidate hardware resource for the gait recognition model. In this example, no single idle computing resource simultaneously meets the needs of multiple target recognition models; each model receives independently allocated candidate hardware resources. In another scenario, if node A's graphics processor simultaneously meets the needs of both a face recognition model and a gait recognition model, the edge computing node will prioritize allocating node A's graphics processor to the face recognition model, which has higher inference complexity. The gait recognition model will then have to wait or be allocated other idle computing resources.

[0068] S63. Bind each of the target recognition models to the selected candidate hardware resources to generate a model resource mapping table, and determine the candidate hardware resources as the target hardware resources corresponding to each of the target recognition models according to the model resource mapping table.

[0069] After completing the resource filtering in sub-step S62, the edge computing node formally binds each target recognition model to its corresponding candidate hardware resource. Specifically, the binding operation involves creating a model-resource mapping table in the edge computing node's memory. This table records the correspondence between the model identifier of each target recognition model and the resource identifier of the candidate hardware resource allocated to it, in key-value pairs. The model identifier is a unique number of the target recognition model in the model configuration library, and the resource identifier is a unique number of the idle computing resource in the available resource pool; the two are linked one-to-one through the mapping table. After the mapping table is established, the edge computing node performs an integrity check to confirm that each target recognition model is bound to at least one candidate hardware resource, and that no two different target recognition models are bound to the same candidate hardware resource. Upon successful verification, the edge computing node formally identifies the candidate hardware resources in the mapping table as the target hardware resources for the target recognition models and sends resource locking requests to the edge computing nodes where each target hardware resource resides, notifying them to reserve the corresponding processor cores and memory space to prepare for the execution of the parallel inference task in subsequent step S7. Meanwhile, edge computing nodes persistently store the mapping table in local logs for subsequent tracing and troubleshooting.

[0070] In one example, the edge computing node creates a model resource mapping table in memory. The table records the following: the face recognition model is bound to the graphics processor (GPU) of node A, the voiceprint recognition model is bound to the central processing unit (CPU) of node B, and the gait recognition model is bound to the GPU of node C. The mapping table has three records, with each target recognition model corresponding to a unique candidate hardware resource; there are no missing or duplicate bindings. After the edge computing node verifies the mapping, it designates the GPU of node A as the target hardware resource for the face recognition model, the CPU of node B as the target hardware resource for the voiceprint recognition model, and the GPU of node C as the target hardware resource for the gait recognition model. Subsequently, the edge computing node sends resource locking requests to nodes A, B, and C, informing each node to reserve the corresponding processor cores and memory space for executing the inference tasks of the corresponding target recognition models.

[0071] In one possible embodiment, when multiple identity recognition results exist, determining the weight coefficient of each identity recognition result in the final decision using the environmental feature parameters and the confidence score corresponding to each identity recognition result includes: S81. Extract the light intensity from the environmental feature parameters, and assign basic weights to each of the identity recognition results based on the comparison result of the light intensity and the preset light threshold. When the light intensity is greater than the preset high light threshold, increase the basic weight of the identity recognition result corresponding to the face recognition model. When the light intensity is less than the preset low light threshold, increase the basic weight of the identity recognition result corresponding to the gait recognition model.

[0072] The edge computing node extracts light intensity from the environmental feature parameters collected in step S1, using light intensity as the core criterion for judging the quality of the current visual conditions. Light intensity is measured in lux; a higher value indicates a brighter environment, and a lower value indicates a darker environment. The edge computing node has two preset judgment benchmarks: a high light threshold and a low light threshold. The high light threshold is typically set to 300 lux, and the low light threshold is typically set to 50 lux. These thresholds can be flexibly adjusted according to the lighting conditions of the actual deployment scenario.

[0073] When the light intensity exceeds a preset high light threshold, it indicates sufficient ambient light, resulting in high-quality raw data for visual biometric recognition models and more reliable inference results. Therefore, edge computing nodes increase the base weight of the identity recognition results corresponding to visual recognition models, including face recognition and iris recognition models. When the light intensity is below a preset low light threshold, it indicates insufficient ambient light, significantly reducing the accuracy of visual recognition models. In this case, edge computing nodes increase the base weight of the identity recognition results corresponding to gait recognition models, as gait recognition models are independent of lighting conditions and exhibit better recognition stability in low-light environments than visual recognition models. When the light intensity is between the preset low light threshold and the preset high light threshold, the base weights of each recognition model are allocated according to a default ratio, which can be set based on the historical accuracy of each model under standard lighting conditions.

[0074] For example, in a security scenario deploying both face recognition and gait recognition models, the preset high-light threshold is 300 lux, and the preset low-light threshold is 50 lux. When the ambient light intensity is 500 lux in a recognition task, exceeding the high-light threshold, the edge computing node increases the base weight of the face recognition model to 0.8 and decreases the base weight of the gait recognition model to 0.2. In another recognition task, with an ambient light intensity of only 20 lux, below the low-light threshold, the edge computing node increases the base weight of the gait recognition model to 0.75 and decreases the base weight of the face recognition model to 0.25. When the light intensity is 150 lux, falling between the two thresholds, the edge computing node allocates the base weights according to the default ratio: 0.6 for the face recognition model and 0.4 for the gait recognition model.

[0075] S82. Extract the ambient noise decibels from the environmental feature parameters, compare the ambient noise decibels with a preset noise threshold, and when the ambient noise decibels are greater than the preset noise threshold, adjust the basic weight of the identity recognition result corresponding to the voiceprint recognition model downward.

[0076] The edge computing node extracts the ambient noise in decibels from the environmental feature parameters collected in step S1, using this decibel level as the core criterion for judging the quality of the current ambient hearing conditions. Ambient noise in decibels is measured in decibels; a higher value indicates a noisier environment, while a lower value indicates a quieter environment. The edge computing node has a preset noise threshold, typically set at 60 decibels, which can be flexibly adjusted according to the acoustic environment of the actual deployment scenario.

[0077] When the ambient noise level exceeds a preset noise threshold, it indicates strong background noise interference in the current environment. This significantly affects the voice signal acquisition quality of the voiceprint recognition model, leading to a decrease in the accuracy of voiceprint feature extraction. Therefore, the edge computing nodes adjust the base weights of the identity recognition results corresponding to the voiceprint recognition model. The adjustment is achieved by multiplying the base weights of the voiceprint recognition model by a correction coefficient, which decreases as the ambient noise level exceeds the preset noise threshold. When the ambient noise level is less than or equal to the preset noise threshold, it indicates that the current ambient noise level is within an acceptable range, and the recognition performance of the voiceprint recognition model is not significantly affected. The edge computing nodes maintain the base weights of the voiceprint recognition model unchanged.

[0078] For example, in a security scenario where a face recognition model and a voiceprint recognition model work together, the preset noise threshold is 60 dB. After allocation in step S81, the base weight of the face recognition model is 0.7, and the base weight of the voiceprint recognition model is 0.3. The current ambient noise level is 75 dB, exceeding the preset noise threshold of 60 dB by 15 dB. The edge computing node multiplies the base weight of the voiceprint recognition model by a correction factor of 0.5, resulting in a reduced base weight of 0.15. To maintain a total weight of 1, the edge computing node proportionally transfers the reduced weight of 0.15 to the face recognition model, adjusting its weight to 0.85. When the ambient noise level is 45 dB, below the preset noise threshold, the base weight of the voiceprint recognition model remains unchanged at 0.3.

[0079] S83. Obtain the confidence score corresponding to each of the identity recognition results, compare each of the confidence scores with a preset confidence threshold, and for the identity recognition results with a confidence score lower than the preset confidence threshold, reduce the corresponding weight coefficient by a preset ratio to obtain the final weight coefficient of each of the identity recognition results in the final decision.

[0080] Edge computing nodes incorporate confidence scores as a feedback mechanism for weight adjustment. The confidence score is a value between 0 and 1 output by each target recognition model during the inference process in step S7, directly reflecting the degree of confidence each target recognition model has in its own inference results. Edge computing nodes have a preset confidence threshold, typically set to 0.8, and a preset reduction ratio, typically set to 50%.

[0081] After obtaining the confidence scores for each identity recognition result, the edge computing node compares each confidence score with a preset confidence threshold. When the confidence score of an identity recognition result is lower than the preset confidence threshold, it indicates that the target recognition model has a low degree of confidence in the inference result, and the credibility of the identity recognition result is insufficient. The edge computing node then lowers the weight coefficient corresponding to the identity recognition result by a preset proportion. When the confidence score of an identity recognition result is greater than or equal to the preset confidence threshold, it indicates that the target recognition model has a high degree of confidence in the inference result, and the edge computing node keeps the weight coefficient corresponding to the identity recognition result unchanged. The resulting weight surplus is proportionally transferred to identity recognition results with confidence scores higher than the preset confidence threshold. After this adjustment, the weight coefficients corresponding to each identity recognition result are the final weight coefficients, which are used for the weighted fusion calculation in step S9.

[0082] For example, the preset confidence threshold is set to 0.8, and the preset reduction ratio is set to 50%. After adjustment in step S82, the weight coefficient of the face recognition model is 0.85, and the weight coefficient of the voiceprint recognition model is 0.15. In step S7, the confidence score output by the face recognition model is 0.92, and the confidence score output by the voiceprint recognition model is 0.65. The confidence score of the face recognition model (0.92) is greater than the preset confidence threshold of 0.8, so the weight coefficient remains unchanged at 0.85. The confidence score of the voiceprint recognition model (0.65) is less than the preset confidence threshold of 0.8, indicating that the voiceprint recognition model has a low confidence in the recognition result. The edge computing node reduces the weight coefficient of the voiceprint recognition model by 50% from 0.15, resulting in a reduced weight coefficient of 0.075. The resulting 0.075 weight margin is transferred to the face recognition model, adjusting the final weight coefficient of the face recognition model to 0.925. Ultimately, the final weight coefficient of the face recognition model is 0.925, the final weight coefficient of the voiceprint recognition model is 0.075, and the sum of the two final weight coefficients is 1.

[0083] In one possible embodiment, the step of weighting and fusing the confidence scores corresponding to each of the identity recognition results using the weighting coefficients to obtain a comprehensive confidence score, and determining the final identity recognition result based on the comprehensive confidence score, includes: S91. The weight coefficients corresponding to each of the identity recognition results are weighted and summed with the corresponding confidence scores to obtain the comprehensive confidence score.

[0084] The edge computing nodes perform a weighted summation operation on the weight coefficients corresponding to each identity recognition result determined in step S8 and the confidence scores corresponding to each identity recognition result obtained in step S7. The formula for the weighted summation is that the overall confidence score equals the sum of the products of each weight coefficient and its corresponding confidence score, i.e., the overall confidence score equals weight coefficient 1 multiplied by confidence score 1 plus weight coefficient 2 multiplied by confidence score 2, and so on, until all identity recognition results are included in the calculation. This weighted summation process integrates the independent judgment results of multiple recognition models into a unified quantitative index. The overall confidence score ranges from 0 to 1, with a higher value indicating higher overall recognition credibility after multimodal fusion. The weight coefficients play a regulatory role in the fusion process, allowing recognition models that perform more stably under adverse environmental conditions to have a greater influence on the final result, while the influence of recognition models that are more susceptible to environmental interference is correspondingly weakened.

[0085] For example, assuming three target recognition models are activated in the current scenario: face recognition, gait recognition, and voiceprint recognition, the weight coefficients determined in step S8 are 0.5 for the face recognition model, 0.3 for the gait recognition model, and 0.2 for the voiceprint recognition model. The confidence scores for each model output in step S7 are 0.88, 0.76, and 0.82, respectively. The edge computing node calculates the overall confidence score according to the formula: 0.5 multiplied by 0.88 plus 0.3 multiplied by 0.76 plus 0.2 multiplied by 0.82, which equals 0.440 plus 0.228 plus 0.164, resulting in a final overall confidence score of 0.832. S92. The comprehensive confidence score is compared with the preset comprehensive confidence threshold. When the comprehensive confidence score is greater than the preset comprehensive confidence threshold, the current multimodal fusion recognition result is determined to be credible. The identity recognition result with the highest confidence score among all the identity recognition results is determined as the final identity recognition result.

[0086] The edge computing node compares the comprehensive confidence score calculated in sub-step S91 with a preset comprehensive confidence threshold to determine whether the multimodal fusion recognition result meets the credibility standard. The preset comprehensive confidence threshold is a pre-configured judgment threshold, typically set between 0.70 and 0.85, and its specific value can be adjusted according to the security level requirements of the security scenario. When the comprehensive confidence score is greater than the preset comprehensive confidence threshold, it indicates that the comprehensive judgment of multiple recognition models after their respective weight adjustments has reached a sufficient level of credibility, and the edge computing node considers the current multimodal fusion recognition result credible. After determining credibility, the edge computing node sorts the confidence scores corresponding to each identity recognition result and selects the identity recognition result with the highest confidence score as the final identity recognition result. The logic for using the highest confidence score as the final identity recognition result is that, under the premise of comprehensive credibility, the identity information output by the recognition model with the highest confidence among all recognition models has the highest individual credibility, making it reasonable to use it as the final output result.

[0087] For example, the overall confidence score calculated in sub-step S91 is 0.832, and the preset overall confidence threshold is 0.75. Since 0.832 is greater than 0.75, the edge computing node determines that the current multimodal fusion recognition result is reliable. At this time, the confidence score of the face recognition model is 0.88, the confidence score of the gait recognition model is 0.76, and the confidence score of the voiceprint recognition model is 0.82. Among the three, the confidence score of the face recognition model is the highest at 0.88. Therefore, the edge computing node determines the identity recognition result "Employee Zhang San" output by the face recognition model as the final identity recognition result.

[0088] S93. When the overall confidence score is less than or equal to the preset overall confidence threshold, the current recognition result is determined to be unreliable, triggering a re-recognition process or refusing passage.

[0089] When the comprehensive confidence score calculated in sub-step S91 is less than or equal to the preset comprehensive confidence threshold, it indicates that the comprehensive judgment after fusion of multiple recognition models has failed to meet the credibility standard, and the current recognition result has significant uncertainty and should not be directly used as the basis for passage decision. At this time, the edge computing node determines that the current recognition result is unreliable. For unreliable situations, the edge computing node can adopt one of two processing methods according to a preset strategy. The first processing method is to trigger a re-identification process, that is, control the multi-sensor array to re-collect the biometric data of the target object and re-execute the recognition process from steps S3 to S9 to try to obtain a more reliable recognition result. The maximum number of re-identification attempts is usually set to 3 times; after exceeding the limit, it automatically switches to the second processing method. The second processing method is to directly refuse passage, that is, the edge computing node controls the access control device to keep the gate closed and does not perform any opening action, while simultaneously sending an alarm notification to the security system management platform, indicating that there is an abnormal identity recognition event.

[0090] For example, suppose that in a certain recognition process, the comprehensive confidence score calculated by sub-step S91 is 0.68, while the preset comprehensive confidence threshold is 0.75. Since 0.68 is less than 0.75, the edge computing node determines that the current recognition result is unreliable. The edge computing node first checks that the current number of consecutive re-recognitions is 1, which is less than the upper limit of 3. Therefore, it triggers the re-recognition process, controls the camera and microphone to re-collect the facial image data and voice data of the target object, and re-executes the subsequent recognition steps. If the comprehensive confidence score is still lower than 0.75 after re-recognition, and the number of consecutive re-recognitions has reached the upper limit of 3, the edge computing node implements a denial-of-access policy, keeps the access control device's gate in the closed state, and sends an alarm notification to the security system management platform. The alarm content is: Node A detected an identity recognition anomaly at 12:15:33 on July 5, 2026. The comprehensive confidence score of 0.68 is lower than the threshold of 0.75, and access has been denied. Please have on-site security personnel verify.

[0091] In one possible embodiment, controlling the access control device associated with the edge computing node to perform corresponding access control actions based on the final identity recognition result includes: S101. The final identity recognition result is compared with the authorized identity database pre-stored on the local edge computing node to determine whether the final identity recognition result matches any authorized identity record in the authorized identity database, and the access permission level corresponding to the target object is determined. The access permission level includes full access permission, time-limited access permission, and no access permission.

[0092] After obtaining the final identity verification result, the edge computing node first compares it with the locally stored authorized identity database. The authorized identity database is pre-stored on the edge computing node's local storage medium. This database contains multiple authorized identity records, each containing at least the authorized person's identity identifier and corresponding access permission level. The identity identifier can be the person's name, employee ID, identification number, or a feature vector code mapped from biometric data. During the comparison, the edge computing node matches the identity identifier in the final identity verification result with each authorized identity record in the authorized identity database. If the identity identifier in the final identity verification result matches the identity identifier in a certain authorized identity record, the match is considered successful, and the corresponding access permission level is read from that authorized identity record. There are three types of access permission levels: full access permission means the target object is qualified to pass at any time; time-limited access permission means the target object is only qualified to pass within a specified time period; and no access permission means the target object has not been authorized to pass. If the identity identifier in the final identity recognition result does not find a matching record in the authorized identity database, the target object's access permission level is determined to be no access permission by default.

[0093] For example, the authorized identity database stored locally on the edge computing node contains the following three authorized identity records: Record 1 is "Employee Zhang San, full access permission", Record 2 is "Visitor Li Si, time-limited access permission, access period is 9:00 to 18:00", and Record 3 is "Employee Wang Wu, full access permission". The final identity recognition result determined in step S9 is "Employee Zhang San". The edge computing node compares "Employee Zhang San" with each record in the authorized identity database and finds that the identity identifier matches that of Record 1, indicating a successful match. The access permission level is then read as full access permission from Record 1. If the final identity recognition result is "Zhao Liu", but there is no record for "Zhao Liu" in the authorized identity database, then Zhao Liu's access permission level is determined to be no access permission.

[0094] S102. Based on the access permission level, determine the access control action type to be executed by the access control device. When the access permission level is full access permission, the access control action type is to open the gate and maintain the preset access duration. When the access permission level is time-limited access permission, the access control action type is to open the gate and maintain the access duration. When the access permission level is no access permission, the access control action type is to keep the gate closed and trigger an alarm prompt.

[0095] The edge computing node, based on the access permission level determined in sub-step S101, queries the corresponding access control action type from a preset control action mapping table. The control action mapping table, stored on the edge computing node's local storage, defines the correspondence between each access permission level and access control action type. When the access permission level is full access, the edge computing node determines the access control action type as opening the gate and maintaining a preset access duration. The preset access duration is a pre-configured fixed duration, typically set between 5 and 10 seconds, sufficient for one target object to complete passage while preventing tailgating. When the access permission level is time-limited access, the edge computing node first obtains the current system time and determines whether it falls within the target object's access time period. If the current system time is within the access time period, the determined access control action type is opening the gate and maintaining a limited duration, typically set to 3 to 5 seconds, shorter than the preset access duration corresponding to full access. If the current system time is not within the permitted passage period, the edge computing node will adjust the access permission level to "no access permission" and execute the action type corresponding to "no access permission". When the access permission level is "no access permission", the access control action type determined by the edge computing node is to keep the gate closed and trigger an alarm. The alarm may include sending an alarm message to the security management center, triggering an audible and visual alarm device, or pushing a notification to the terminal device of on-site security personnel.

[0096] In one example, target Zhang San has full access permission. The edge computing node determines the access control action type as opening the gate and maintaining access for a preset duration of 8 seconds based on the control action mapping table. Target Li Si has limited access permission, with access hours from 9:00 to 18:00. The current system time is 11:48, which falls within the access period. The edge computing node determines the access control action type as opening the gate and maintaining access for a limited duration of 4 seconds. If the current system time is 20:00, Li Si's access permission level is adjusted to no access permission, and the access control action type is keeping the gate closed and triggering an alarm. Target Zhao Liu has no access permission. The edge computing node determines the access control action type as keeping the gate closed and triggering an alarm. The alarm includes sending a message to the security management center stating "Unauthorized personnel attempted to pass" and triggering a flashing red warning light on the access control device.

[0097] S103. Send a control command corresponding to the access control action type to the access control device, and simultaneously generate an access event record. The access event record includes at least the final identity recognition result, the access control action type, the execution timestamp, and the node identifier of the edge computing node.

[0098] After determining the access control action type, the edge computing node generates a control command corresponding to that action type and sends it to the access control device's controller via a communication interface. The communication interface can be any of RS485 bus, CAN bus, Ethernet, or wireless communication. The control command carries gate action parameters, including the gate opening command, gate holding duration, and the execution sequence of the gate closing command. Upon receiving the control command, the access control device's controller parses the gate action parameters and drives the gate actuator to complete the specified action. Simultaneously with the execution of the control command, the edge computing node generates an access event record and writes it to a local log file or uploads it to the security system's central management platform. The access event record contains at least four pieces of information: the first is the final identification result, used to record the identity of the target object for this access; the second is the access control action type, used to record the handling method for this access request; and the third is the execution timestamp, used to record the precise time when the control command was issued, in the format of year-month-day hour:minute:second. The fourth item is the node identifier of the edge computing node, which records the unique number of the edge computing node that performed this access control, facilitating the location of the problematic node when the security system contains multiple edge computing nodes. In addition, the access event log can be expanded to include auxiliary information such as the access control device number, on-site captured image, and passage direction to enhance the completeness and traceability of the event log.

[0099] For example, if an edge computing node determines that the access control action type is to open the gate and maintain access for a preset duration of 8 seconds, it will generate a control command. The control command will include the gate action parameters: open, maintain open for 8 seconds, and close. The edge computing node will send the control command to the access control controller via an Ethernet interface. The controller will then open the gate and automatically close it after 8 seconds. Simultaneously, the edge computing node will generate an access event record with the following content: "Final identification result: Employee Zhang San, Access control action type: Full access - Gate open for 8 seconds, Execution timestamp: 2026-07-05 11:48:51, Node ID: NODE-A-01". If the access control action type is "No access permission triggers alarm", the access event record will be: "Final identification result: Zhao Liu, Access control action type: No access permission - Maintain closed and alarm, Execution timestamp: 2026-07-05 11:50:03, Node ID: NODE-A-01".

[0100] In one possible embodiment, determining at least two target recognition models to be activated from a preset model configuration library based on the environmental feature parameters, and the target biometric data type corresponding to each target recognition model, includes: Step S21: Extract the light intensity from the environmental feature parameters, compare the light intensity with a preset light threshold, and when the light intensity is greater than the preset light threshold, determine the face recognition model as a target recognition model; when the light intensity is less than the preset light threshold, remove the face recognition model from the candidate target recognition model set, and determine the gait recognition model as a target recognition model.

[0101] In this sub-step, the edge computing node extracts the key environmental indicator of light intensity from the environmental feature parameters collected in step S1, and filters the vision-based biometric recognition model based on the light intensity. Light intensity directly determines the clarity of the image captured by the camera and the richness of available feature points, making it a core environmental factor affecting the accuracy of the face recognition model. The edge computing node compares the extracted light intensity with a preset light threshold, which is calibrated based on the minimum lighting conditions under which the face recognition model can maintain a recognition accuracy of over 90% in historical tests, typically set to 50 lux. When the light intensity is greater than the preset light threshold, the edge computing node determines that the current ambient lighting conditions meet the requirements for the normal operation of the face recognition model and identifies the face recognition model as a target recognition model. When the light intensity is less than the preset light threshold, it indicates insufficient ambient lighting, and the number of extractable facial feature points in the face image will be significantly reduced. In this case, the edge computing node removes the face recognition model from the candidate target recognition model set and instead identifies the gait recognition model as a target recognition model. Gait recognition models identify targets by analyzing their walking posture and body contour features. They are independent of lighting conditions and can maintain a high recognition accuracy even in low-light environments.

[0102] For example, in an access control system deployed at a parking garage entrance, an edge computing node extracts an illumination intensity of 30 lux from environmental feature parameters, which is lower than the preset illumination threshold of 50 lux. In this case, the edge computing node determines that the current ambient light is insufficient and unsuitable for using a face recognition model, removing it from the candidate target recognition model set, while simultaneously identifying a gait recognition model as one of the target recognition models. At the office building entrance, however, the illumination intensity is 500 lux, exceeding the preset illumination threshold, so the edge computing node directly identifies the face recognition model as a target recognition model.

[0103] Step S22: Extract ambient noise decibels from the environmental feature parameters, compare the ambient noise decibels with a preset noise threshold, and when the ambient noise decibels are less than the preset noise threshold, determine the voiceprint recognition model as a target recognition model; when the ambient noise decibels are greater than or equal to the preset noise threshold, remove the voiceprint recognition model from the candidate target recognition model set.

[0104] In this sub-step, the edge computing node extracts the ambient noise decibels from environmental feature parameters and filters the voiceprint recognition model based on the noise level. The core working principle of the voiceprint recognition model is to extract the speaker's acoustic features from the speech signal. Ambient noise directly interferes with the acquisition quality of the speech signal, causing the acoustic features to be submerged or distorted. The edge computing node compares the ambient noise decibels with a preset noise threshold, which is calibrated based on the maximum noise level at which the voiceprint recognition model can maintain a recognition accuracy of over 90% in historical tests, and is usually set to 65 decibels. When the ambient noise decibels are less than the preset noise threshold, it indicates that the current ambient noise level is low and the quality of the speech signal meets the requirements of the voiceprint recognition model. The edge computing node then identifies the voiceprint recognition model as a target recognition model. When the ambient noise decibels are greater than or equal to the preset noise threshold, it indicates that the current ambient noise level is too high, and the effective acoustic features in the speech signal will be severely interfered with by the noise. In this case, the edge computing node removes the voiceprint recognition model from the candidate target recognition model set and no longer considers it as a candidate model for the current recognition cycle.

[0105] For example, at the entrance of a building facing the street, an edge computing node extracts an ambient noise level of 72 decibels from environmental feature parameters, which is higher than the preset noise threshold of 65 decibels. In this case, the edge computing node determines that the current ambient noise is too high, making it difficult for the voiceprint recognition model to function properly, and removes the voiceprint recognition model from the candidate target recognition model set. In a quiet office building lobby, the ambient noise level is 40 decibels, which is lower than the preset noise threshold, and the edge computing node recognizes the voiceprint recognition model as a target recognition model.

[0106] Step S23: Based on the finally determined target recognition models, read the unique target biometric data type corresponding to each target recognition model from the model configuration library, and generate a correspondence table between target recognition models and target biometric data types.

[0107] In this sub-step, after completing the model selection in steps S21 and S22, the edge computing node configures the data types associated with the finally determined target recognition models. The model configuration library pre-stores a unique correspondence between each biometric recognition model and the biometric data types it needs to collect. For example, a face recognition model uniquely corresponds to face image data, a gait recognition model uniquely corresponds to gait sequence data, a voiceprint recognition model uniquely corresponds to voice data, and an infrared thermal imaging recognition model uniquely corresponds to infrared thermal imaging image data. The edge computing node iterates through each of the finally determined target recognition models, reading the unique target biometric data type corresponding to each model from the model configuration library, and compiles the reading results into a correspondence table between target recognition models and target biometric data types. This correspondence table will serve as the basis for guiding data collection in the subsequent step S3. Based on the target biometric data types recorded in this table, the edge computing node selectively calls the corresponding sensors to collect biometric data of the target object.

[0108] For example, suppose that after filtering in steps S21 and S22, the final target recognition models are a face recognition model and a voiceprint recognition model. The edge computing node reads the unique face image data corresponding to the face recognition model and the unique voice data corresponding to the voiceprint recognition model from the model configuration library, generating the following correspondence table: the face recognition model corresponds to the face image data, and the voiceprint recognition model corresponds to the voice data. This correspondence table will guide step S3 to call the camera and microphone to collect the face image data and voice data of the target object, respectively. If the final target recognition model is an infrared thermal imaging recognition model and a gait recognition model, then the correspondence table is: the infrared thermal imaging recognition model corresponds to the infrared thermal imaging image data, and the gait recognition model corresponds to the gait sequence data. Step S3 will then call the infrared camera and the visible light camera to collect the corresponding data accordingly.

[0109] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0110] Please see Figure 4 This illustration shows a schematic diagram of an environmentally aware security control device provided in an exemplary embodiment of this application, hereinafter referred to as device 4. Device 4 can be implemented as all or part of a security device through software, hardware, or a combination of both. Device 4 includes: Acquisition module 401 is used to acquire environmental feature parameters through an associated multi-sensor array in response to detecting that a target object has entered the recognition area; The determining module 402 is used to determine, based on the environmental feature parameters, at least two target recognition models to be activated from a preset model configuration library, and the target biometric data types corresponding to each target recognition model; The acquisition module 401 is further configured to collect target biometric data corresponding to the target object based on the determined target biometric data type. The acquisition module 401 is also used to acquire an available resource pool, which consists of the idle computing resources of each edge computing node in the security system. The calculation module 403 is used to calculate the reasoning complexity required for the target recognition model to complete the reasoning for each target recognition model, based on the number of its model parameters and the amount of corresponding target biometric data collected. The allocation module 404 is used to allocate target hardware resources to each target recognition model from the available resource pool according to the reasoning complexity of each target recognition model. The inference module 405 is used to send the biometric data of each target to the target hardware resources allocated to them, and execute the inference of each target recognition model in parallel to obtain the identity recognition result and confidence score output by each target recognition model. The calculation module 403 is also used to determine the weight coefficient of each identity recognition result in the final decision by using the environmental feature parameters and the confidence score corresponding to each identity recognition result when there are multiple identity recognition results. The weighting module 406 is used to perform weighted fusion of the confidence scores corresponding to each of the identity recognition results using the weight coefficients to obtain a comprehensive confidence score, and to determine the final identity recognition result based on the comprehensive confidence score. The control module 407 is used to control the access control device associated with the edge computing node to perform corresponding access control actions based on the final identity recognition result.

[0111] For further details regarding the implementation of the above-mentioned technical solutions by each module in the security control device based on environmental perception, please refer to the description of a security control method based on environmental perception provided in the above-mentioned embodiments of the invention, which will not be repeated here.

[0112] It should be noted that the device 4 provided in the above embodiments, when executing a security control method based on environmental perception, is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the above functions. In addition, the security control device based on environmental perception provided in the above embodiments and the security control method embodiment based on environmental perception belong to the same concept, and the implementation process is detailed in the method embodiment, which will not be repeated here.

[0113] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0114] See Figure 5 The diagram shown is a schematic of a computer storage medium provided in an embodiment of this application. The computer storage medium can store multiple instructions (i.e., ... Figure 5 The computer program shown above), the instructions are adapted to be loaded and executed as described above. Figure 2 The method steps of the illustrated embodiment can be found in the following documentation for detailed execution. Figure 2 The specific details of the illustrated embodiments will not be elaborated here.

[0115] This application also provides a computer program product that stores at least one instruction, which is loaded and executed to implement an environment-aware security control method as described in the above embodiments.

[0116] Please see Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 6 As shown, the electronic device 600 may include: at least one 601, at least one network interface 604, a user interface 603, a memory 605, and at least one communication bus 602.

[0117] The communication bus 602 is used to enable communication between these components.

[0118] The user interface 603 may include input units such as a mouse and keyboard.

[0119] The network interface 604 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0120] 601 may include one or more processing cores. 601 connects various parts within the electronic device 600 using various interfaces and lines, and performs various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 605, and by calling data stored in memory 605. Optionally, 601 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). 601 may integrate one or more of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also not be integrated into 601 and may be implemented as a separate chip.

[0121] The memory 605 may include random access memory (RAM) or read-only memory. Optionally, the memory 605 may include a non-transitory computer-readable storage medium. The memory 605 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 605 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 605 may also be at least one storage device located remotely from the aforementioned 601. Figure 6 As shown, the memory 605, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and application programs.

[0122] exist Figure 6 In the illustrated electronic device 600, the user interface 603 is primarily used to provide an input interface for the user and to acquire user input data; while 601 can be used to call the application program stored in the memory 605 and specifically execute, for example... Figure 2 The method shown can be referred to for details. Figure 2 As shown, it will not be elaborated further here.

[0123] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory, or random access memory, etc.

[0124] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A security control method based on environmental perception, characterized in that, include: ‌ In response to the detection of a target object entering the recognition area, environmental feature parameters are collected through an associated multi-sensor array; Based on the environmental feature parameters, at least two target recognition models to be activated are determined from the preset model configuration library, as well as the target biometric data types corresponding to each target recognition model; Based on the determined target biometric data type, collect the target biometric data corresponding to the target object; Obtain an available resource pool, which consists of the idle computing resources of each edge computing node in the security system; For each target recognition model, the reasoning complexity required for the target recognition model to complete the reasoning is calculated based on the number of its model parameters and the amount of corresponding target biometric data collected. Based on the reasoning complexity of each target recognition model, target hardware resources are allocated to each target recognition model from the available resource pool; The biometric data of each target are sent to the target hardware resources allocated to them, and the inference of each target recognition model is executed in parallel to obtain the identity recognition result and its confidence score output by each target recognition model. When there are multiple identity recognition results, the weight coefficient of each identity recognition result in the final decision is determined by using the environmental feature parameters and the confidence score corresponding to each identity recognition result. The confidence scores corresponding to each of the identity recognition results are weighted and fused using the weight coefficients to obtain a comprehensive confidence score, and the final identity recognition result is determined based on the comprehensive confidence score. Based on the final identity recognition result, the access control device associated with the edge computing node is controlled to perform corresponding access control actions.

2. The method according to claim 1, characterized in that, The process of obtaining an available resource pool includes: Send resource query requests to other edge computing nodes in the security system, excluding edge computing nodes. The resource query requests are used to request each of the other edge computing nodes to return its current processor utilization and available memory capacity. Receive the processor utilization and available memory capacity returned by each of the other edge computing nodes, and determine the edge computing nodes whose processor utilization is lower than a preset utilization threshold and whose available memory capacity is higher than a preset memory threshold as target edge computing nodes with idle computing resources; The idle computing resources of the edge computing node itself and the idle computing resources of each target edge computing node are aggregated to generate the available resource pool, and a resource identifier is configured for each idle computing resource in the available resource pool. The resource identifier is used to locate and allocate the corresponding target hardware resources.

3. The method according to claim 1, characterized in that, For each target recognition model, the inference complexity required for the model to complete the inference is calculated based on the number of model parameters and the amount of corresponding target biometric data collected, including: Obtain the model configuration file corresponding to the target recognition model from the model configuration library, parse the model configuration file to obtain the model parameter count of the target recognition model, the model parameter count includes the total number of model parameters and the number of floating-point operations required for a single inference; The amount of data collected for the corresponding target biometric data is obtained, wherein the amount of data includes at least one of the data dimension, number of data frames, and size of a single frame of the target biometric data; The inference complexity of the target recognition model is calculated based on the number of model parameters and the amount of data. The inference complexity is the product of the number of floating-point operations and the amount of data, or the weighted sum of the total number of parameters and the amount of data.

4. The method according to claim 1, characterized in that, The step of allocating target hardware resources from the available resource pool for each target recognition model based on the inference complexity of each target recognition model includes: Based on the inference complexity of each target recognition model, the type and amount of hardware resources required for each target recognition model are determined. The type of hardware resources includes at least one of a central processing unit, a graphics processing unit, and a neural network processor. The amount of resources required includes the number of processor cores and the required memory capacity. Traverse all idle computing resources in the available resource pool, filter out candidate hardware resources that meet the hardware resource type and the resource requirement, and when the same idle computing resource meets the resource requirements of multiple target recognition models at the same time, give priority to allocating the idle computing resource to the target recognition model with the highest inference complexity; Each target recognition model is bound to the selected candidate hardware resources to generate a model resource mapping table, and the candidate hardware resources are determined as the target hardware resources corresponding to each target recognition model according to the model resource mapping table.

5. The method according to claim 1, characterized in that, When multiple identity recognition results exist, the weight coefficient of each identity recognition result in the final decision is determined using the environmental feature parameters and the confidence score corresponding to each identity recognition result, including: Illumination intensity is extracted from the environmental feature parameters. Based on the comparison between the illumination intensity and the preset illumination threshold, basic weights are assigned to each of the identity recognition results. Specifically, when the illumination intensity is greater than the preset high illumination threshold, the basic weight of the identity recognition result corresponding to the face recognition model is increased. When the illumination intensity is less than the preset low illumination threshold, the basic weight of the identity recognition result corresponding to the gait recognition model is increased. The environmental noise decibels are extracted from the environmental feature parameters. The environmental noise decibels are compared with a preset noise threshold. When the environmental noise decibels are greater than the preset noise threshold, the basic weight of the identity recognition result corresponding to the voiceprint recognition model is adjusted downward. Obtain the confidence score corresponding to each of the identity recognition results, compare each confidence score with a preset confidence threshold, and for identity recognition results with a confidence score lower than the preset confidence threshold, reduce the corresponding weight coefficient by a preset ratio to obtain the final weight coefficient of each of the identity recognition results in the final decision.

6. The method according to claim 1, characterized in that, The step of weighting and fusing the confidence scores corresponding to each identity recognition result using the weighting coefficients to obtain a comprehensive confidence score, and determining the final identity recognition result based on the comprehensive confidence score, includes: The weight coefficients corresponding to each of the identity recognition results are weighted and summed with the corresponding confidence scores to obtain the comprehensive confidence score; The comprehensive confidence score is compared with a preset comprehensive confidence threshold. When the comprehensive confidence score is greater than the preset comprehensive confidence threshold, the current multimodal fusion recognition result is determined to be credible. The identity recognition result with the highest confidence score among all the identity recognition results is determined as the final identity recognition result. When the overall confidence score is less than or equal to the preset overall confidence threshold, the current recognition result is determined to be unreliable, triggering a re-recognition process or refusing passage.

7. The method according to claim 1, characterized in that, The step of controlling the access control device associated with the edge computing node to perform corresponding access control actions based on the final identity recognition result includes: The final identity recognition result is compared with the authorized identity database pre-stored on the local edge computing node to determine whether the final identity recognition result matches any authorized identity record in the authorized identity database, and to determine the access permission level corresponding to the target object. The access permission level includes full access permission, time-limited access permission, and no access permission. Based on the access permission level, the access control device is determined to perform the type of access control action. When the access permission level is full access permission, the access control action is to open the gate and maintain the preset access time. When the access permission level is time-limited access permission, the access control action is to open the gate and maintain the access time for a limited period of time. When the access permission level is no access permission, the access control action is to keep the gate closed and trigger an alarm. Send a control command corresponding to the access control action type to the access control device, and simultaneously generate an access event record. The access event record includes at least the final identity recognition result, the access control action type, the execution timestamp, and the node identifier of the edge computing node.

8. The method according to claim 1, characterized in that, The step of determining at least two target recognition models to be activated from a preset model configuration library based on the environmental feature parameters, and the target biometric data type corresponding to each target recognition model, includes: Illumination intensity is extracted from the environmental feature parameters. The illumination intensity is compared with a preset illumination threshold. When the illumination intensity is greater than the preset illumination threshold, the face recognition model is determined to be a target recognition model. When the illumination intensity is less than the preset illumination threshold, the face recognition model is removed from the candidate target recognition model set, and the gait recognition model is determined to be a target recognition model. The environmental noise decibels are extracted from the environmental feature parameters. The environmental noise decibels are compared with a preset noise threshold. When the environmental noise decibels are less than the preset noise threshold, the voiceprint recognition model is determined to be a target recognition model. When the environmental noise decibels are greater than or equal to the preset noise threshold, the voiceprint recognition model is removed from the candidate target recognition model set. Based on the final determined target recognition models, the target biometric data type uniquely corresponding to each target recognition model is read from the model configuration library, and a correspondence table between target recognition models and target biometric data types is generated.

9. A security control device based on environmental perception, characterized in that, The device includes: The acquisition module is used to collect environmental feature parameters through an associated multi-sensor array in response to the detection of a target object entering the recognition area; The determination module is used to determine, based on the environmental feature parameters, at least two target recognition models to be activated from a preset model configuration library, and the target biometric data types corresponding to each target recognition model; The acquisition module is also used to collect target biometric data corresponding to the target object according to the determined target biometric data type; The acquisition module is also used to acquire an available resource pool, which consists of the idle computing resources of each edge computing node in the security system. The calculation module is used to calculate the reasoning complexity required for each target recognition model to complete the reasoning, based on the number of its model parameters and the amount of corresponding target biometric data collected. The allocation module is used to allocate target hardware resources to each target recognition model from the available resource pool according to the inference complexity of each target recognition model. The inference module is used to send the biometric data of each target to the target hardware resources allocated to them, and execute the inference of each target recognition model in parallel to obtain the identity recognition result and confidence score output by each target recognition model. The calculation module is also used to determine the weight coefficient of each identity recognition result in the final decision by using the environmental feature parameters and the confidence score corresponding to each identity recognition result when there are multiple identity recognition results. The weighting module is used to perform weighted fusion of the confidence scores corresponding to each of the identity recognition results using the weight coefficients to obtain a comprehensive confidence score, and to determine the final identity recognition result based on the comprehensive confidence score. The control module is used to control the access control device associated with the edge computing node to perform corresponding access control actions based on the final identity recognition result.

10. An electronic device, characterized in that, include: A processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and execute the steps of the environment-aware security control method as described in any one of claims 1 to 9.

Images