Access control rapid release method based on attention mechanism improvement

Abstract

The invention provides an improved access control rapid release method based on an attention mechanism, and belongs to the technical field of computer vision and artificial intelligence, and the method comprises the steps: carrying out the rapid processing of continuous image frames collected by a camera through a parallel visual processing module and an attention weighted reasoning mechanism, and obtaining a target type and confidence; multi-frame confidence coefficient weighted fusion is carried out through a time sequence attention mechanism so as to improve the recognition stability; and a conflict punishment mechanism and security level judgment logic (such as blacklist interception and a VIP priority strategy) are introduced on the basis of the fusion confidence coefficient to make a decision. A visual identification method based on an improved YOLOv7 embedded SimAM attention mechanism is supplemented with a risk level strategy judgment and time sequence reasoning mechanism, high-confidence identification of vehicles and faces is achieved, and gate lifting response is rapidly controlled. The method has millisecond-level response capability, extremely low false recognition rate and flexible safety strategy support, realizes quick, safe and stable brake bar lifting control, and reduces the false opening and closing rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and artificial intelligence, and particularly relates to a gate pass quick release method based on an improved attention mechanism, which is suitable for a gate control scene in an enterprise gate pass system, and particularly suitable for quick pole lifting response optimization under vehicle and face recognition BACKGROUND

[0002] With the growing demand for enterprise security and commuting management, intelligent gate systems based on visual recognition gradually replace traditional card swiping or password verification methods. However, in the prior art, due to factors such as light, angle, occlusion and algorithm accuracy, the recognition response speed is unstable, the misrecognition rate is high, and the passing efficiency is low, which seriously affects the user experience and system operation efficiency. Current mainstream methods such as YOLOv5 / YOLOv7 target detection framework have achieved remarkable results in general target recognition, but direct application in the gate pass scene still has adaptability problems. Especially in complex light, motion blur, occlusion and other environments, the target feature extraction capability decreases, which affects the overall recognition accuracy and the decision robustness of the gate pole lifting control. Therefore, how to improve the recognition ability of the visual recognition model for faces, license plates and other targets, reduce misjudgment and omission, and shorten the delay time from recognition to pole lifting, has become the research focus in this field. SUMMARY

[0003] The present application provides a gate pass quick release method based on an improved attention mechanism, which solves the problems of slow gate response, unstable recognition, lack of security strategy and other problems in the prior art, and realizes:

[0004] Real-time requirement: completing the processing of the input video frame and the gate pole lifting decision within 300 milliseconds;

[0005] High-precision recognition: high-precision visual detection for single target object to minimize misjudgment and omission;

[0006] Stability enhancement: introducing cross-frame timing attention mechanism to weight and fuse multiple frame results, improving the stability and reliability of the recognition output;

[0007] Security level control: decision constraint through conflict punishment mechanism and security level judgment logic (including blacklist interception and VIP priority release) to ensure system security.

[0008] The technical solution of the present application is:

[0009] The application discloses an improved access control fast release method based on an attention mechanism, which utilizes a parallel visual processing module and an attention weighted inference mechanism to quickly process continuous image frames collected by a camera to obtain a target category and a confidence; a time sequence attention mechanism is used to weight and fuse multiple frame confidences to improve recognition stability; and a conflict penalty mechanism and a safety level judgment logic (such as blacklist interception and VIP priority strategy) are introduced on the basis of the fused confidences to make decisions. The visual recognition method based on the improved YOLOv7 embedded SimAM attention mechanism, supplemented by a risk level strategy determination and a time sequence inference mechanism, realizes high-confidence recognition of vehicles and faces, and quickly controls the lifting of the gate machine. The application has millisecond-level response capability, extremely low misrecognition rate and flexible safety strategy support, realizes fast and safe and stable gate lever lifting control, and reduces the misopening and misclosing rate.

[0010] Further,

[0011] Parallel visual processing: using a camera arranged at the entrance area of the gate machine to continuously collect video frames, using multi-core parallel computing or GPU acceleration means to process the captured image frames;

[0012] Target detection and feature extraction are respectively performed on the current frame and the adjacent frame, and the detection model used combines an attention mechanism to highlight the target area features to speed up the key area recognition.

[0013] Further,

[0014] The camera collected image is subjected to size normalization, illumination normalization and angle correction operations to improve the robustness and consistency of subsequent recognition.

[0015] The working steps are as follows:

[0016] Collect the access control scene image, pre-process and construct the visual input;

[0017] Vehicle and face target detection is performed based on the improved YOLOv7 network, wherein the network is embedded with a non-parameter attention mechanism SimAM;

[0018] According to the confidence and time interval of the detection output, it is judged whether the gate machine lifting is triggered or not;

[0019] Under the condition of meeting the safety strategy, the gate machine is quickly controlled to execute actions.

[0020] Further,

[0021] The license plate and face area in the image are synchronously detected, feature extraction and classification are respectively performed, and corresponding confidences C car , C face are outputted.

[0022] The vehicle and face recognition confidence thresholds are respectively θ1 and θ2, when both are met and the recognition time interval Δt is less than the preset time T_max, the gate lifting rod is triggered.

[0023] Further,

[0024] The rule-based safety judgment logic is set: the blacklist and VIP list are configured in the system in advance; for the detected target, firstly, whether the identity or feature is in the blacklist is queried; if it is a blacklist member, the system is marked as high risk, and the lifting rod is forcibly rejected; if it is a VIP personnel or vehicle, it is marked as low risk, and the release instruction is preferentially issued, which is not limited by the confidence threshold; otherwise, according to the comparison between the fusion confidence S and the preset threshold, or through the risk level function R=f(S), whether the lifting rod is lifted or not is decided.

[0025] Further,

[0026] The risk level function is defined as:

[0027]

[0028] Wherein, R_{\text{max}} is the highest risk value, R_{\text{min}} is the lowest risk value; when R is lower than the preset risk threshold, the lifting rod instruction is sent to the gate control module, otherwise the gate rod is kept locked.

[0029] Further,

[0030] After obtaining the judgment result, the gate is controlled to execute corresponding actions according to the decision result.

[0031] Further,

[0032] When it is judged to allow release, the gate rod is quickly lifted up; when it is judged to reject release, the gate rod is kept or lowered again; the whole process adopts parallel process design: while waiting for the execution of the previous frame decision, the next frame data is processed in parallel, so as to further compress the overall delay.

[0033] The model can be deployed on an edge device, the recognition and lifting rod delay is less than 50ms, and the model volume is less than 70MB.

[0034] The beneficial effects of the present application are

[0035] The response delay is realized, which is much better than the traditional scheme;

[0036] After introducing SimAM and weighted confidence fusion, the average recognition accuracy (mAP) is improved by more than 8%;

[0037] The conflict penalty function greatly reduces the misrecognition rate and improves the system robustness;

[0038] Supports flexible risk strategies (VIP priority / blacklist blocking) to improve security management.

[0039] A complete software upgrade was implemented, ensuring compatibility with the existing gate hardware system. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the workflow of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0042] This invention discloses an improved access control fast release method based on an attention mechanism, which can achieve high-precision visual recognition and control of a single target while ensuring a response time of less than 300 milliseconds. The method utilizes a parallel visual processing module and an attention-weighted inference mechanism to rapidly process continuous image frames captured by a camera, obtaining the target category and confidence level. A temporal attention mechanism is used to weighted fuse the confidence levels of multiple frames to improve recognition stability. Furthermore, a conflict penalty mechanism and security level judgment logic (such as blacklist blocking and VIP priority strategies) are introduced based on the fused confidence levels for decision-making. Based on an improved YOLOv7 visual recognition method embedding the SimAM attention mechanism, supplemented by risk level strategy judgment and a temporal inference mechanism, high-confidence recognition of vehicles and faces is achieved, and the gate opening response is quickly controlled. The millisecond-level response capability, extremely low false recognition rate, and flexible security strategies enable this invention to achieve fast, safe, and stable gate opening control, reducing the rate of false opening and closing.

[0043] Specifically as follows:

[0044] Parallel visual processing: The method uses cameras deployed in the gate entrance area to continuously capture video frames, and employs multi-core parallel computing or GPU acceleration to perform visual processing on the captured image frames. Specifically, target detection and feature extraction are performed on the current frame and adjacent frames respectively. The detection model used combines attention mechanisms (such as self-attention, spatial channel attention, etc.) to highlight the features of the target area, thereby accelerating the identification of key areas.

[0045] Image preprocessing module: Performs operations such as size normalization, illumination normalization, and angle correction on images captured by access control cameras to improve the robustness and consistency of subsequent recognition.

[0046] YOLOv7 network structure enhanced by timing attention: SimAM modules are embedded in the Backbone and Head structures of YOLOv7 to introduce spatial non-parametric attention mechanisms.

[0047]

[0048] wherein, α represents the current channel mean, β is an adjustable factor, t is the current target feature response, σ is a Sigmoid activation function, and is an element-wise multiplication.

[0049] Multi-modal fusion recognition module: The model simultaneously detects the license plate and face regions in the image, adopts a double-branch inference mechanism for feature extraction and classification, and outputs the corresponding confidence C car , C face . The joint confidence S is obtained by weighted fusion

[0050] S = w1c car +w2c face , w1+w2 = 1

[0051] Conflict penalty mechanism: To improve the stability of the output results, when the processor detects that the classification results between consecutive frames are inconsistent (such as an object being classified into different categories in different frames), the conflict penalty mechanism is triggered. The pole decision and conflict penalty mechanism are as follows:

[0052] If C car > θ1 and C face > θ2, the "candidate pole lifting" state is entered;

[0053] If △t < Tmax, the system judges that it is highly reliable, and the pole is lifted;

[0054] Otherwise, if the consecutive misrecognition n > N, the alarm is triggered and the exception is recorded.

[0055] The vehicle and face recognition confidence thresholds are θ1 = 0.8 and θ2 = 0.85, respectively. When both are satisfied and the recognition time interval △t is less than the preset T_max, the gate pole is lifted.

[0056] The penalty function P is introduced: P = λ1(1-c car ) 2 + λ2(1-c face ) 2 .

[0057] Security level judgment: the fused confidence S is used to judge the security level of the target and finally decide whether to lift the barrier. The invention designs a rule-based security judgment logic: a blacklist and a VIP list are pre-configured in the system. For the detected target, first query whether its identity or feature is in the blacklist; if it is a blacklist member, the system marks it as high risk and forcibly refuses to lift the barrier; if it is identified as a VIP person or vehicle, it is marked as low risk and the release instruction is issued preferentially, not subject to the general confidence threshold. Otherwise, according to the comparison between the fused confidence S and the preset threshold, or through the risk level function R = f (S) to decide whether to lift the barrier. The risk level function can be defined as:

[0058]

[0059] Where R_{\text{max}} is the highest risk value and R_{\text{min}} is the lowest risk value. When R is lower than the preset risk threshold, send the lift barrier instruction to the gate control module, otherwise keep the barrier locked.

[0060] Execution control: after obtaining the above judgment result, the controller controls the gate to perform corresponding actions according to the decision result. When it is judged to allow release, control the barrier to rise quickly; when it is judged to refuse release, keep or lower the barrier again. The whole process adopts parallel process design: while waiting for the execution of the last frame decision, the next frame data is processed immediately in parallel to further compress the overall delay.

[0061] The above only describes the preferred embodiments of the present application, which are used to illustrate the technical solutions of the present application, and are not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A fast access control method based on an improved attention mechanism, characterized in that, By utilizing a parallel visual processing module and an attention-weighted inference mechanism, continuous image frames captured by a camera are processed to obtain the target category and confidence level. A temporal attention mechanism is used to weight and fuse the confidence levels of multiple frames to improve recognition stability. Based on the fused confidence levels, a conflict penalty mechanism and a security level judgment logic are introduced for decision-making. A visual recognition method based on an improved YOLOv7 embedded SimAM attention mechanism, supplemented by risk level strategy judgment and temporal inference mechanism, achieves high-confidence recognition of vehicles and faces and quickly controls the gate opening response.

2. The method according to claim 1, characterized in that, Parallel visual processing: Video frames are continuously captured using cameras placed in the gate entrance area, and the captured image frames are processed visually using multi-core parallel computing or GPU acceleration. Target detection and feature extraction are performed on the current frame and adjacent frames respectively. The detection model used combines an attention mechanism to highlight the features of the target region, so as to accelerate the identification of key regions.

3. The method according to claim 2, characterized in that, The images captured by the camera are normalized in size, illumination, and angle to improve the robustness and consistency of subsequent recognition.

4. The method according to claim 3, characterized in that, The work steps are as follows: Collect images of the access control scene, perform preprocessing to construct visual input; Vehicle and face target detection is performed based on an improved YOLOv7 network, wherein the network is embedded with a parameterless attention mechanism, SimAM. Based on the confidence level and time interval of the detection output, determine whether to trigger the gate to lift. Under the premise of meeting the security policy, quickly control the gate to perform actions.

5. The method according to claim 4, characterized in that, Simultaneously detect license plate and face regions in the image, perform feature extraction and classification respectively, and output the corresponding confidence scores Ccar and Cface; Set the confidence thresholds for vehicle and face recognition to θ1 and θ2, respectively. When both are satisfied and the recognition time interval Δt is less than the preset time T_max, the gate is triggered to lift.

6. The method according to claim 5, characterized in that, Set up rule-based security judgment logic: Pre-configure blacklists and VIP lists in the system; for detected targets, first query whether their identity or characteristics are in the blacklist; if they are blacklist members, the system marks them as high risk and forcibly refuses to raise the barrier; if they are identified as VIP personnel or vehicles, they are marked as low risk and priority is given to issuing a passage instruction, without being restricted by the confidence threshold; otherwise, the system compares the fused confidence S with the preset threshold, or decides whether to raise the barrier through the risk level function R = f(S).

7. The method according to claim 6, characterized in that, The risk level function is defined as follows: Where R_{max}} is the highest risk value and R_{min}} is the lowest risk value; when R is lower than the preset risk threshold, a gate control module is sent a gate lifting command; otherwise, the gate remains locked.

8. The method according to claim 6 or 7, characterized in that, After obtaining the judgment result, the gate is controlled to perform the corresponding action based on the decision result.

9. The method according to claim 8, characterized in that, When the decision is made to allow passage, the control gate is raised quickly; when the decision is made to refuse passage, the gate is held or lowered again. The entire process adopts a parallel workflow design: while waiting for the previous frame's decision to be executed, the next frame's data is processed in parallel in real time to further reduce the overall latency.

10. The method according to claim 9, characterized in that, The model is deployed on edge devices with recognition and lever raising latency of less than 50ms and model size of less than 70MB.

Images