Automatic door anti-pinch control method, device and storage medium based on artificial intelligence

Abstract

The present invention relates to an automatic door anti-pinch control method, device and storage medium based on artificial intelligence. The method comprises acquiring original image data through an image acquisition device arranged at a specific position of the automatic door, and performing region extraction on the original image data to obtain corresponding real-time image data; inputting the real-time image data into a preset artificial intelligence deep neural network model for target detection to obtain a target recognition result; dividing the operation area based on the target recognition result to obtain a corresponding multi-level protection area; setting the automatic door trigger parameter based on the multi-level protection area to obtain the corresponding regional trigger parameter; performing gate control judgment on the target recognition result according to the regional trigger parameter and the multi-level protection area, and executing a preset door control action to obtain a corresponding execution result; adjusting the regional trigger parameter in real time based on the execution result to obtain the corresponding system parameter. The present invention can realize the intelligent control of the automatic door.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic door anti-pinch technology, and in particular to an automatic door anti-pinch control method, device and storage medium based on artificial intelligence. Background Art

[0002] As an indispensable intelligent facility in modern buildings, automatic doors have been widely used in commercial centers, hospitals, office buildings and other places. With the continuous improvement of people's requirements for safety and convenience, how to realize intelligent anti-pinch control of automatic doors and ensure the safety of personnel passage has become one of the key research topics. At present, the mainstream anti-pinch control method of automatic doors mainly relies on a single detection method such as infrared sensors or pressure sensors. This method has problems such as large detection blind spots and poor environmental adaptability. At the same time, traditional anti-pinch control strategies often use fixed trigger parameters and cannot be dynamically adjusted according to different scenarios and usage conditions. This may lead to insufficient reliability and flexibility of anti-pinch control, affecting the overall safety performance of the automatic door system. Summary of the invention

[0003] The main purpose of the present invention is to provide an automatic door anti-pinch control method, device and storage medium based on artificial intelligence, which can realize intelligent control of automatic doors and avoid problems such as false triggering or untimely response that may occur under traditional control methods.

[0004] To achieve the above object, the present invention provides an automatic door anti-pinch control method based on artificial intelligence, comprising:

[0005] The original image data is acquired by an image acquisition device arranged at a specific position of the automatic door, and a region is extracted from the original image data to obtain corresponding real-time image data;

[0006] Inputting the real-time image data into a preset artificial intelligence deep neural network model for target detection to obtain a target recognition result;

[0007] Based on the target recognition result, the operation area is divided to obtain corresponding multi-level protection areas;

[0008] Automatic door triggering parameters are set based on the multi-level protection areas to obtain corresponding area triggering parameters;

[0009] Performing gate control judgment on the target recognition result according to the area trigger parameter and the multi-level protection area, and executing a preset gate control action to obtain a corresponding execution result;

[0010] The area trigger parameters are adjusted in real time based on the execution results to obtain corresponding system parameters.

[0011] Furthermore, the original image data is obtained by an image acquisition device arranged at a specific position of the automatic door, and the region of the original image data is extracted to obtain corresponding real-time image data, including:

[0012] Performing position analysis on the image acquisition device according to preset door parameters to obtain corresponding installation position coordinate data;

[0013] Performing shooting angle analysis on the image acquisition device according to the installation position coordinate data to obtain corresponding image acquisition angle parameters;

[0014] Performing perspective transformation on the original image data according to the image acquisition angle parameter to obtain corresponding image perspective data;

[0015] Performing regional segmentation on the image perspective data according to a preset door area template to obtain a corresponding target area image;

[0016] Performing edge detection processing on the target area image to obtain corresponding edge feature data;

[0017] The target area image is morphologically processed according to the edge feature data to obtain the real-time image data.

[0018] Furthermore, the real-time image data is input into a preset artificial intelligence deep neural network model for target detection to obtain a target recognition result, including:

[0019] Inputting the real-time image data into the artificial intelligence deep neural network model for standardized enhancement processing to obtain enhanced image data;

[0020] Performing feature complexity calculation on the enhanced image data to obtain a feature complexity parameter;

[0021] The feature complexity parameter is compared according to a preset recognition threshold, and when the feature complexity parameter is greater than the preset recognition threshold, the artificial intelligence deep neural network model is switched to a dual-model recognition mode;

[0022] When switching to the dual-model recognition mode, the enhanced image data is subjected to feature extraction and fusion processing by two parallel convolutional neural networks to obtain an initial recognition result of the dual-model recognition;

[0023] When the feature complexity parameter is less than the preset recognition threshold, the artificial intelligence deep neural network model switches to a single model recognition mode;

[0024] When switching to the single model recognition mode, feature extraction is performed on the enhanced image data through three 3×3 convolutional layers to obtain a corresponding output feature map;

[0025] Perform BatchNorm processing on the output feature map of the BatchNorm layer to obtain a reconstructed feature map;

[0026] Performing nonlinear mapping on the reconstructed feature map through a ReLU activation function to obtain a corresponding mapping feature map;

[0027] Downsampling the mapped feature map through a maximum pooling layer to obtain a primary feature map;

[0028] The primary feature map is feature mapped through two 1×1 convolutional layers, and feature filtered through a Dropout layer to obtain dimension-reduced feature data;

[0029] Performing feature classification on the dimension-reduced feature data through a fully connected layer to obtain the initial recognition result of a single model;

[0030] The recognition result output by the artificial intelligence deep neural network model is reliability verified to obtain the target recognition result.

[0031] Furthermore, the feature extraction and fusion processing of the enhanced image data by two parallel convolutional neural networks to obtain the initial recognition result of dual-model recognition includes:

[0032] The first branch of convolutional neural network processing includes:

[0033] Processing the enhanced image data by a first residual block to obtain first-level feature data;

[0034] Processing the first-level feature data by a second residual block to obtain second-level feature data;

[0035] Processing the second-level feature data by a third residual block to obtain third-level feature data;

[0036] Processing the third-level feature data by a fourth residual block to obtain a first feature map;

[0037] Performing attention mechanism processing on the first feature map to obtain first enhanced feature data;

[0038] The second branch of convolutional neural network processing includes:

[0039] Processing the enhanced image data through a first depthwise separable convolution block to obtain first layer feature data;

[0040] Processing the first layer feature data through a second depthwise separable convolution block to obtain second layer feature data;

[0041] Processing the second layer feature data by a third depth-wise separable convolution block to obtain a third layer feature data;

[0042] Processing the third layer feature data through a fourth depth-wise separable convolution block to obtain fourth layer feature data;

[0043] Processing the fourth layer feature data through a fifth depth-wise separable convolution block to obtain a second feature map;

[0044] Performing pyramid pooling processing on the second feature map to obtain second enhanced feature data;

[0045] Performing feature map splicing on the first enhanced feature data and the second enhanced feature data to obtain fused feature data;

[0046] The fused feature data is subjected to feature classification through a fully connected layer to obtain the initial recognition result of dual-model recognition.

[0047] Furthermore, the operation area is divided based on the target recognition result to obtain corresponding multi-level protection areas, including:

[0048] Extracting target coordinates from the target recognition result to obtain real-time target coordinate data;

[0049] Performing target object motion trajectory analysis based on the real-time target coordinate data to obtain a motion parameter set of the target object;

[0050] Performing double threshold partitioning processing on the motion parameter set to obtain a start frame initial area and an anti-pinch frame initial area;

[0051] Performing grid mapping processing according to the initial area of ​​the start frame and the initial area of ​​the anti-pinch frame to obtain a basic protection grid;

[0052] Performing double-layer density calculation on the basic protection grid to obtain a start frame density matrix and an anti-pinch frame density matrix;

[0053] Performing region expansion processing on the initial region of the start frame according to the start frame density matrix to obtain a start frame expansion region;

[0054] Performing dynamic shrinkage processing on the initial area of ​​the anti-pinch frame according to the anti-pinch frame density matrix to obtain the core area of ​​the anti-pinch frame;

[0055] Performing boundary fusion processing on the start frame extension area and the anti-pinch frame core area to obtain a regional transition zone;

[0056] The start frame extension area and the anti-pinch frame core area are subjected to regional integration processing according to the regional transition zone to obtain the multi-level protection area.

[0057] Furthermore, the automatic door trigger parameter setting based on the multi-level protection area to obtain the corresponding area trigger parameter includes:

[0058] Perform hierarchical preset of regional trigger parameters for the working area to obtain corresponding start trigger parameters and anti-pinch trigger parameters;

[0059] Performing unified sensitivity calibration on the start trigger parameter and the anti-pinch trigger parameter to obtain a corresponding sensitivity coefficient, wherein the value range of the sensitivity coefficient is 0-100;

[0060] Performing weighted calculation on the start trigger parameter and the anti-pinch trigger parameter according to the sensitivity coefficient to obtain a weighted trigger parameter, wherein the weight of the start trigger parameter is 0.3, and the weight of the anti-pinch trigger parameter is 0.7;

[0061] Performing timing analysis on the weighted trigger parameters to obtain timing correlation parameters including trigger time interval, duration and response speed;

[0062] Performing threshold optimization on the weighted trigger parameter according to the timing correlation parameter to obtain an optimized trigger threshold;

[0063] Adaptively adjusting the optimization trigger threshold to obtain a dynamic trigger parameter;

[0064] The dynamic trigger parameter is subjected to safety redundancy processing to obtain the regional trigger parameter, wherein the safety redundancy processing includes setting a backup trigger parameter.

[0065] Further, the gate control determination is performed on the target recognition result according to the area trigger parameter and the multi-level protection area, and a preset gate control action is executed to obtain a corresponding execution result, including:

[0066] Extracting target features from the target recognition results to obtain a target feature data set;

[0067] Scoring the importance of the multi-level protection areas according to the target feature data set to obtain a regional scoring matrix;

[0068] Calibrate the regional trigger parameters according to the regional scoring matrix to obtain a gated trigger threshold;

[0069] Performing time series analysis on the target feature data set in the multi-level protection area to obtain target motion trajectory prediction data;

[0070] Calculate the collision risk according to the target motion trajectory prediction data and the gate trigger threshold to obtain a collision risk level;

[0071] Selecting a corresponding door control action strategy according to the collision risk level to obtain a door control instruction sequence;

[0072] Sorting the execution priority of the gated instruction sequence to obtain a final execution instruction;

[0073] According to the final execution instruction, the automatic door is controlled to execute the corresponding action to obtain the door control execution state;

[0074] Performing state analysis on the gated execution state to obtain the execution result.

[0075] Further, the real-time adjustment of the area trigger parameter based on the execution result to obtain the corresponding system parameter includes:

[0076] Performing execution data analysis on the execution result to obtain corresponding execution sequence data;

[0077] Perform cross deviation verification on the regional trigger parameters according to the execution sequence data to obtain a corresponding parameter deviation matrix;

[0078] Performing eigendecomposition on the parameter deviation matrix to obtain a corresponding deviation feature group;

[0079] Performing nonlinear mapping on the area trigger parameters according to the deviation feature group to obtain corresponding mapping parameters;

[0080] Adaptively adjust the area trigger parameter according to the mapping parameter to obtain a corresponding adjustment parameter;

[0081] Performing an index evaluation on the adjustment parameter according to a preset evaluation rule to obtain a corresponding evaluation index;

[0082] Performing bidirectional verification on the adjustment parameters according to the evaluation index to obtain corresponding verification parameters;

[0083] Incrementally optimizing the verification parameters to obtain corresponding optimization parameters;

[0084] The area trigger parameters are recursively updated according to the optimization parameters to obtain corresponding system parameters.

[0085] The present invention also provides an automatic door anti-pinch control device based on artificial intelligence, which is applied to any one of the automatic door anti-pinch control methods based on artificial intelligence described above, comprising:

[0086] An acquisition module, the acquisition module is used to acquire original image data through an image acquisition device arranged at a specific position of the automatic door, and perform region extraction on the original image data to obtain corresponding real-time image data;

[0087] An analysis module, wherein the analysis module is used to input the real-time image data into a preset artificial intelligence deep neural network model for target detection to obtain a target recognition result;

[0088] An association module, the association module is used to divide the operation area based on the target recognition result to obtain corresponding multi-level protection areas;

[0089] A processing module, the processing module is used to set automatic door trigger parameters based on the multi-level protection area to obtain corresponding area trigger parameters;

[0090] A control module, wherein the control module performs a gate control judgment on the target recognition result according to the area trigger parameter and the multi-level protection area, and executes a preset gate control action to obtain a corresponding execution result;

[0091] An execution module is used to adjust the area trigger parameters in real time based on the execution result to obtain corresponding system parameters.

[0092] The present invention also provides a storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute any of the above methods.

[0093] The present invention provides an automatic door anti-pinch control method, device and storage medium based on artificial intelligence, which have the following beneficial effects:

[0094] By dividing the target recognition results into multiple protection areas and setting the corresponding regional trigger parameters based on this, differentiated protection control of different areas is achieved. Compared with the traditional fixed parameter control method, it has higher control accuracy and flexibility, and can better adapt to the security protection needs in different scenarios. Based on the regional trigger parameters and multi-level protection areas, door control judgment and execution of corresponding control actions can realize intelligent control of automatic doors, avoid problems such as false triggering or untimely response that may occur under traditional control methods, and significantly improve the safety performance of the system. Through real-time analysis and feedback of the execution results, the regional trigger parameters are dynamically adjusted, so that the system can adaptively optimize the control strategy according to the actual usage situation, which not only improves the environmental adaptability of the system, but also realizes the continuous optimization of anti-pinch control, ensuring the stable and reliable operation of the automatic door system in various application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 This is a flow chart of an automatic door anti-pinch control method based on artificial intelligence provided by the present invention;

[0096] Figure 2 It is a structural diagram of an automatic door anti-pinch control device based on artificial intelligence provided by the present invention.

[0097] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0098] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0099] The present invention is further described below in conjunction with the accompanying drawings and specific implementation methods.

[0100] Reference Figure 1 As shown, an automatic door anti-pinch control method based on artificial intelligence includes:

[0101] Step S1: acquiring original image data through an image acquisition device arranged at a specific position of the automatic door, and performing region extraction on the original image data to obtain corresponding real-time image data;

[0102] Step S2: input the real-time image data into a preset artificial intelligence deep neural network model for target detection to obtain a target recognition result;

[0103] Step S3: Divide the operation area based on the target recognition result to obtain the corresponding multi-level protection area;

[0104] Step S4: setting automatic door trigger parameters based on the multi-level protection area to obtain corresponding area trigger parameters;

[0105] Step S5: performing gate control judgment on the target recognition result according to the regional trigger parameter and the multi-level protection area, and executing the preset gate control action to obtain the corresponding execution result;

[0106] Step S6: Based on the execution result, the regional trigger parameters are adjusted in real time to obtain corresponding system parameters.

[0107] Based on the above steps, the detailed process is as follows:

[0108] Step S1: Install a high-quality image acquisition device at a key position of the automatic door. The image acquisition device can be an infrared camera or a depth camera. These devices are usually installed above or on both sides of the door frame to ensure that the activity range of the door area can be fully covered. The acquisition device continuously captures scene images, usually acquiring raw image data at a frequency of 25-30 frames per second. The acquired raw image data will then undergo a series of preprocessing steps, including image denoising, illumination compensation, and distortion correction to improve image quality. According to the needs of the actual application scenario, the processed image is subjected to regional extraction, focusing on the activity area within about 2-3 meters before and after the automatic door. The region extraction process adopts an image segmentation algorithm, using traditional image processing methods such as edge detection, region growing, etc., and can also use a semantic segmentation method based on deep learning.

[0109] Step S2: Use pre-trained deep learning models, such as mainstream target detection networks such as YOLO, Faster R-CNN or SSD. These models are usually trained with a large amount of data and can accurately identify common targets such as humans, children, pets, and luggage. When real-time image data is input into the model, the network will automatically extract image features to locate and classify the target. The network output results include the detected target category, location coordinates, bounding box, and confidence score. In order to improve the real-time and accuracy of detection, the model is often optimized in a targeted manner, including model pruning, quantization and other operations to ensure smooth operation on embedded devices. Establish a target tracking mechanism to calculate the target's motion trajectory and speed by comparing the target position changes between adjacent frames. This information has an important impact on the subsequent protection area division and trigger parameter setting.

[0110] Step S3: Based on the target recognition results obtained, the door area is divided into multiple security protection levels according to the location, movement state and importance of different targets. The usual practice is to divide the area into three levels: emergency protection area, warning area and observation area. The emergency protection area is usually located within the direct action range of the door leaf, about 0.5 meters; the warning area is located within 1-1.5 meters outside the emergency protection area; the observation area extends to 2-3 meters outside the warning area. The range of each area is dynamically adjusted according to the movement direction and speed of the target. For example, when a fast-moving target is detected, the range of the warning area will be appropriately expanded. In addition, different types of targets may correspond to different protection strategies. For example, for children or the elderly, the system will adopt a larger safety margin and expand the range of the protection area.

[0111] Step S4: Based on the multi-level protection areas divided in step S3, configure corresponding trigger parameters for each area. These parameters mainly include door speed control parameters, response time thresholds, and safety distance thresholds. In the emergency protection area, the most stringent trigger parameters are set, such as reducing the door speed to the minimum (usually not exceeding 0.3 m / s), and the response time threshold is set within 100 milliseconds to ensure timely response to emergencies. In the warning area, relatively mild parameter settings are used, the door speed can be appropriately increased (0.3-0.5 m / s), and the response time threshold can be extended to 200-300 milliseconds. The observation area uses the most relaxed parameter settings, which are mainly used for pre-judgment and early warning.

[0112] Step S5: Compare the target recognition results with the area trigger parameters in real time to make a gating decision. When a target is detected to enter a certain protection area, determine whether the type, location, and motion characteristics of the target trigger the parameter threshold of the area. If the trigger condition is met, the corresponding preset gating action is executed immediately. These actions include: emergency stop (when a target is detected in the emergency protection area), deceleration operation (when the target is in the warning area), normal operation or start-up (when the target is only moving in the observation area), etc. The system will also consider the movement trend of the target and make predictive control in advance. For example, when it is detected that the target is approaching the warning area quickly, the door speed is reduced in advance to reserve sufficient response time for possible emergencies. After each control action is executed, the action type, response time, execution effect and other information are recorded to provide a basis for subsequent parameter optimization.

[0113] Step S6: Continuously analyze the execution results and evaluate the rationality and effectiveness of the current parameter settings. The main evaluation indicators include: false alarm rate (the frequency of triggering stop or deceleration when it is not necessary), missed alarm rate (the situation where it should be triggered but is not triggered), response time (the time interval from detecting the target to executing the action), etc. Based on these evaluation results, the regional trigger parameters are automatically adjusted. For example, if it is found that the false alarm rate is high in a certain period of time, the trigger threshold of this period is appropriately increased; if it is found that the missed alarm rate of a certain type of target is high, the corresponding trigger threshold will be lowered. The system will also establish a parameter learning mechanism, and gradually optimize the parameter settings by analyzing a large amount of historical data, so that the system can maintain a high traffic efficiency while ensuring safety. This continuous parameter optimization ensures that the system can continuously adapt to changes in the actual use environment and provide optimal anti-pinch protection.

[0114] The present invention provides an artificial intelligence-based automatic door anti-pinch control method, which realizes differentiated protection control of different areas by dividing the target recognition results into multi-level protection areas and setting corresponding regional trigger parameters based on this. Compared with the traditional fixed parameter control method, it has higher control accuracy and flexibility, and can better adapt to the safety protection needs in different scenarios. Based on the regional trigger parameters and multi-level protection areas, door control judgment is performed and corresponding control actions are executed, which can realize intelligent control of automatic doors, avoid problems such as false triggering or untimely response that may occur in traditional control methods, and significantly improve the safety performance of the system. Through real-time analysis and feedback of the execution results, the regional trigger parameters are dynamically adjusted, so that the system can adaptively optimize the control strategy according to the actual usage situation, which not only improves the environmental adaptability of the system, but also realizes the continuous optimization of anti-pinch control, ensuring the stable and reliable operation of the automatic door system in various application scenarios.

[0115] In one embodiment, the original image data is obtained by an image acquisition device arranged at a specific position of the automatic door, and the original image data is subjected to region extraction to obtain corresponding real-time image data, including:

[0116] The image acquisition device is fixed at a preset position of the automatic door through a pan / tilt bracket.

[0117] When determining the installation position of the camera, use precise positioning and combine the preset door parameters for position analysis. Establish a standard three-dimensional coordinate system, where the X axis is parallel to the ground and the door, the Y axis is perpendicular to the ground, the Z axis is perpendicular to the door surface, and the center point of the door is set as the origin (0, 0, 0). The ideal installation position coordinates (x, y, z) of the camera are calculated by the least squares method. The x value range is 20-50mm, the y value range is 1800-2000mm, the z value range is ±30mm, the installation height is 1.2-1.5 times the door height, and the horizontal distance is 0.8-1.2 times the door width.

[0118] When determining the shooting angle of the camera, an optimization calculation is performed based on the coordinate data of the installation position. The two key angle parameters of the camera are determined by projection matrix analysis. The effective range of the horizontal rotation angle α is ±20°, and the effective range of the pitch angle β is -5° to -15°.

[0119] After acquiring the original image, perspective transformation processing is performed. Coordinate mapping is established based on the pre-calibrated internal and external parameters of the camera, and the 3×3 perspective transformation matrix H is used for image correction. The 4-point perspective transformation algorithm is used to convert the oblique view into a front view image. The least squares method is used to calculate the transformation matrix to achieve geometric correction.

[0120] Then, the region segmentation process is performed and the preset door area template is loaded. The template contains the door contour feature vector and key point coordinate information, defines the geometric features of the door area (boundary contour, area range, aspect ratio, etc.), and extracts the target area image through the template matching algorithm.

[0121] In the edge detection processing stage, the improved Canny operator is used for processing, Gaussian filtering (kernel size 5×5) is implemented, the image gradient is calculated, non-maximum suppression is performed, and an adaptive threshold is used (the low threshold is 1 / 3 of the high threshold). The high threshold is dynamically calculated by the OTSU algorithm.

[0122] Morphological processing is performed using multi-level morphological operations, and dilation operations (using 3×3 rectangular structure elements, iterated 2 times) and erosion operations (using 3×3 rectangular structure elements, iterated 1 time) are performed. The noise is eliminated by combining opening and closing operations, edge holes are filled, and the integrity of the target area is maintained.

[0123] System performance indicators include: overall preprocessing time is controlled within 50ms, dual-model recognition time is 200ms, single-model recognition time is 100ms, and system sensitivity range is 0-100 (default value 50, step size 10).

[0124] This embodiment realizes a stable and reliable automatic door anti-pinch control preprocessing system through precise image acquisition layout, systematic image processing flow and reliable parameter setting. It has strong environmental adaptability and anti-interference ability, and the processing accuracy and real-time performance meet the actual application requirements, providing a reliable data basis for subsequent target detection and anti-pinch control. It also supports real-time parameter adjustment through wireless communication modules, further improving the flexibility and adaptability of the system. All parameter settings have been verified by experiments, fully considering the real-time requirements of the system while ensuring processing accuracy, laying a solid foundation for realizing reliable anti-pinch control.

[0125] In one embodiment, real-time image data is input into a preset artificial intelligence deep neural network model for target detection to obtain target recognition results, including:

[0126] When the real-time image data is processed by standardization enhancement, the normalization method is used to scale the image pixel values ​​to the [0, 1] interval, and the image contrast is improved through the histogram equalization technology. The enhanced image data has more prominent target features and more uniform brightness distribution, which is conducive to subsequent feature extraction.

[0127] The feature complexity calculation is based on the gray-level co-occurrence matrix of the enhanced image data. The feature complexity parameter is obtained by calculating the statistical features of the image, such as entropy, contrast, and correlation. This parameter reflects the complexity of the image content and is used for the adaptive selection of subsequent recognition modes. The preset recognition threshold is set to 0.75, which represents the cutoff point of medium complexity.

[0128] When the feature complexity parameter is greater than 0.75, the system switches to dual-model recognition mode. In this mode, two parallel convolutional neural networks use the ResNet50 and DenseNet121 architectures to extract features from enhanced image data. The feature maps of the two networks are fused by weighted averaging, and the weight coefficient is determined by the network confidence. The fused features are passed through the corresponding classifier to obtain the initial results of dual-model recognition.

[0129] When the feature complexity parameter is less than 0.75, the system switches to single model recognition mode. In this mode, three 3×3 convolutional layers extract features from the image in turn, with the number of convolution kernels in each layer being 64, 128, and 256 respectively. The extracted feature map is normalized by the BatchNorm layer to eliminate data distribution offset. The normalized feature map introduces nonlinear characteristics through the ReLU activation function to enhance the model's expressiveness.

[0130] The feature map activated by ReLU is downsampled by 2×2 by the maximum pooling layer, which compresses the size of the feature map while retaining the significant features. The downsampled feature map is compressed in the channel dimension by two 1×1 convolution layers, with the number of convolution kernels being 128 and 64 respectively. The Dropout layer randomly masks features with a probability of 0.5 to prevent overfitting. Finally, the features are mapped to the predefined category space through the fully connected layer to obtain the initial recognition results of the single model.

[0131] When verifying the reliability of the recognition results output by the model, two dimensions are used: confidence threshold and temporal consistency. The confidence threshold is set to 0.9, and recognition results below this threshold will be marked as unreliable. Temporal consistency requires that the recognition results of three consecutive frames remain consistent, otherwise they will also be marked as unreliable. The results that have been verified for reliability are used as the final target recognition results for subsequent anti-pinch control decisions.

[0132] This embodiment can achieve all-round monitoring of the traffic area by arranging an image acquisition device at a specific position of the automatic door to obtain real-time image data, and combining the deep neural network model to perform target detection and reliability verification, effectively overcoming the detection blind spot problem of the traditional single sensor, and significantly enhancing the intelligence and accuracy of the automatic door anti-pinch control. The method introduces the feature complexity parameter as the basis for switching the recognition mode, realizes the adaptive conversion of the single model and dual model recognition modes, and improves the system's adaptability to different scenes. In the dual model recognition mode, the feature fusion mechanism of the two parallel convolutional neural networks fully utilizes the advantages of different networks and enhances the recognition robustness in complex scenes. In the single model recognition mode, the multi-level feature extraction and dimensionality reduction processing effectively reduce the consumption of computing resources and improve the recognition efficiency of simple scenes. The dual verification mechanism based on the confidence threshold and time consistency effectively filters unreliable recognition results and improves the reliability of target detection. Overall, this method realizes the reasonable allocation of computing resources while ensuring the recognition accuracy, and greatly improves the reliability and accuracy of anti-pinch control.

[0133] In one embodiment, two parallel convolutional neural networks are used to perform feature extraction and fusion processing on the enhanced image data to obtain an initial recognition result of dual-model recognition, including:

[0134] The first branch convolutional neural network uses a residual block structure for feature extraction. The residual block consists of multiple convolutional layers, batch normalization layers, and activation function layers. The convolution kernel size of the first residual block is set to 7×7, the step size is 2, and the number of channels is 64. The input enhanced image data is processed to obtain the first-level feature data. The convolution kernel size of the second residual block is set to 3×3, the step size is 2, and the number of channels is 128. The first-level feature data is processed to obtain the second-level feature data. The convolution kernel size of the third residual block is set to 3×3, the step size is 2, and the number of channels is 256. The second-level feature data is processed to obtain the third-level feature data. The convolution kernel size of the fourth residual block is set to 3×3, the step size is 2, and the number of channels is 512. The third-level feature data is processed to obtain the first feature map. The first feature map is processed by an attention mechanism, and the attention mechanism includes a channel attention module and a spatial attention module. The channel attention module uses global average pooling and global maximum pooling to extract features, and the spatial attention module uses a convolutional layer and a sigmoid activation function to generate attention weights to obtain the first enhanced feature data.

[0135] The second branch convolutional neural network uses a deep separable convolutional block structure for feature extraction. The first deep separable convolutional block contains a deep convolutional layer and a point-by-point convolutional layer. The convolution kernel size of the deep convolutional layer is 3×3, the step size is 2, and the number of channels is 32. The input enhanced image data is processed to obtain the first layer of feature data. The deep convolutional layer of the second deep separable convolutional block has a convolution kernel size of 3×3, a step size of 2, and a number of channels of 64. The first layer of feature data is processed to obtain the second layer of feature data. The deep convolutional layer of the third deep separable convolutional block has a convolution kernel size of 3×3, a step size of 2, and a number of channels of 128. The second layer of feature data is processed to obtain the third layer of feature data. The deep convolutional layer of the fourth deep separable convolutional block has a convolution kernel size of 3×3, a step size of 2, and a number of channels of 256. The third layer of feature data is processed to obtain the fourth layer of feature data. The convolution kernel size of the depth convolution layer of the fifth depth-separable convolution block is 3×3, the step size is 2, and the number of channels is 512. The fourth layer feature data is processed to obtain the second feature map. The second feature map is subjected to pyramid pooling, which includes pooling operations of four scales: 1×1, 2×2, 3×3, and 6×6, to obtain the second enhanced feature data.

[0136] The dual-model recognition mechanism fuses the features extracted by the two branch networks. The first enhanced feature data and the second enhanced feature data are spliced ​​in the channel dimension to obtain fused feature data. The fully connected layer performs feature classification on the fused feature data. The number of neurons in the fully connected layer is set to 1024, and the ReLU activation function is used. The number of neurons in the output layer is the same as the number of classification categories. The corresponding activation function is used to obtain the initial recognition result of the dual-model recognition. The dual-model recognition mechanism makes full use of the feature extraction capability of the residual network and the computational efficiency of the deep separable convolution to improve the recognition performance of the model.

[0137] This embodiment uses a dual-branch parallel convolutional neural network design, and uses residual blocks and depth-separable convolution blocks for feature extraction, respectively, which effectively combines the advantages of the two network structures and improves the comprehensiveness and accuracy of feature extraction. In the first branch, the attention mechanism is introduced to adjust the weights of the channel and spatial dimensions of the feature map, which enhances the expressiveness of key features and suppresses the interference of irrelevant features. In the second branch, a pyramid pooling structure is used to obtain target information at different scales through multi-scale feature extraction, thereby improving the adaptability of the model to targets of different sizes. The features extracted by the two branch networks are complementarily enhanced through splicing fusion and classification of the fully connected layers, overcoming the limitations of a single network structure. This method reduces the computational complexity while ensuring recognition accuracy, and provides efficient and reliable technical support for automatic door anti-pinch control.

[0138] In one embodiment, based on the target recognition result, the operation area is divided to obtain corresponding multi-level protection areas, including:

[0139] During the target coordinate extraction process, for each recognized target object, the center point coordinates (x, y) of its bounding box and the width and height information (w, h) of the bounding box are extracted to form target coordinate data. By comparing the target coordinate data between consecutive frames, motion parameters such as the motion speed, motion direction, and acceleration of the target object are calculated to form a set of motion parameters. When the speed of the target object exceeds 0.5 m / s or the acceleration exceeds 0.3 m / s², it is determined to be in a fast motion state.

[0140] The dual-threshold partitioning process uses two preset thresholds: the start threshold T1 and the anti-pinch threshold T2, where T1 < T2. According to the comparison of the parameter values in the set of motion parameters with these two thresholds, when the parameter value is greater than T1, it is divided into the initial start box area, and when the parameter value is greater than T2, it is divided into the initial anti-pinch box area. A grid coordinate system is established for the entire operation area, with a grid size of 10 cm × 10 cm, and the initial start box area and the initial anti-pinch box area are mapped into this grid system to generate a basic protection grid.

[0141] On the basic protection grid, the density of target objects in each grid is calculated. The start box density matrix records the number of target objects with a speed greater than T1 in each grid, and the anti-pinch box density matrix records the number of target objects with a speed greater than T2 in each grid. The area expansion process expands the initial start box area, and the expansion distance is proportional to the density. The greater the density, the greater the expansion distance, and the maximum expansion distance is 30 cm. The dynamic contraction process contracts the initial anti-pinch box area towards the area with a higher density according to the distribution characteristics of the anti-pinch box density matrix, and the contraction radius is inversely proportional to the density.

[0142] The boundary fusion process establishes a 5-cm-wide transition zone between the expanded start box area and the core anti-pinch box area. The area integration process integrates the expanded start box area, the transition zone, and the core anti-pinch box area into three-level protection areas: the observation area (expanded start box area), the warning area (transition zone), and the emergency protection area (core anti-pinch box area). The observation area triggers the normal operation of the automatic door, the warning area triggers the deceleration and stop of the automatic door, and the emergency protection area triggers the emergency stop and retraction of the automatic door.

[0143] This embodiment realizes adaptive division of the protection area through dual threshold partitioning and grid density calculation method, so that the protection area can be dynamically adjusted according to the motion state of the target object. The regional expansion and contraction processing mechanism ensures the reasonable distribution of the protection area and avoids misjudgment caused by too large or too small protection area. Setting the regional transition zone realizes the smooth transition between different protection areas and effectively reduces the control jitter caused by the sudden change of the protection area. The three-level protection area division method enables the automatic door to adopt corresponding control strategies according to the position of the target object in different areas, which not only ensures safety but also improves the flexibility of door control. The overall solution not only ensures the safe operation of the automatic door, but also avoids the inconvenience caused by excessive protection, and achieves a good balance between safety and practicality.

[0144] In one embodiment, automatic door triggering parameters are set based on the multi-level protection area to obtain corresponding area triggering parameters, including:

[0145] When setting the automatic door trigger parameters based on multi-level protection areas, the area trigger parameters are graded and preset for the operating area. The area trigger parameters are divided into start trigger parameters and anti-pinch trigger parameters. The start trigger parameters include the distance range for the automatic door to start, the speed of people approaching, the density of people and other parameters. The anti-pinch trigger parameters include the anti-pinch area distance, the speed of people moving, the human body pressure and other parameters.

[0146] During sensitivity calibration, the sensitivity of the start trigger parameter and the anti-pinch trigger parameter are unified to the standard range of 0-100, and each parameter is standardized to have the same numerical benchmark. During the calibration process, the larger the sensitivity coefficient, the more sensitive the system response, and the smaller the sensitivity coefficient, the slower the system response. The calibration result directly affects the trigger effect of the system.

[0147] The start trigger parameter and the anti-pinch trigger parameter are weighted, and the anti-pinch trigger parameter is given a weight of 0.7 and the start trigger parameter is given a weight of 0.3. The weighted calculation ensures that the anti-pinch function occupies a more important position in the entire control process and improves the safety of the system.

[0148] During the timing analysis, the trigger time interval, duration and response speed of the weighted trigger parameters are analyzed to obtain the timing correlation parameters. The timing correlation parameters reflect the time characteristics of the trigger event, including the minimum interval between two triggers, the maximum duration of a single trigger, and the time delay from the system receiving the signal to responding.

[0149] Based on the timing correlation parameters, the threshold of the weighted trigger parameters is optimized. By adjusting the trigger threshold, the system has better response characteristics while ensuring safety. The optimization process comprehensively considers the timeliness of the trigger and the probability of false triggering, and achieves the best control effect in actual operation.

[0150] In the adaptive adjustment stage, the optimization trigger threshold is dynamically adjusted according to environmental changes and usage conditions to form dynamic trigger parameters. The adaptive adjustment mechanism enables the system to adapt to different usage environments and scenarios, improving the flexibility of control.

[0151] Safety redundancy processing ensures system reliability by setting backup trigger parameters and enabling backup parameters when the main trigger parameters fail. The backup trigger parameters are appropriately adjusted based on the main trigger parameters, slightly increasing the trigger threshold while ensuring safety, and ultimately forming regional trigger parameters with redundant protection.

[0152] This embodiment realizes precise adjustment of the anti-pinch control of the automatic door by presetting the regional trigger parameters of the operating area in a graded manner, combined with sensitivity calibration and weighted calculation, and effectively improves the safety and reliability of the system. Through unified sensitivity calibration and differentiated weight distribution, the anti-pinch function occupies a dominant position in the control process, significantly reducing the risk of personal injury. The system adopts a combination of timing analysis and threshold optimization to ensure that the trigger parameters can maintain the best state under different working environments, thereby improving the accuracy and adaptability of the control. The adaptive adjustment mechanism enables the system to dynamically adjust the trigger parameters according to actual usage, enhancing the environmental adaptability of the system. The safety redundancy processing establishes a dual protection mechanism by setting the backup trigger parameters, which can maintain the normal operation of the system even when the main trigger parameters fail, greatly improving the overall safety and reliability.

[0153] In one embodiment, a gate control determination is performed on the target recognition result according to the area trigger parameter and the multi-level protection area, and a preset gate control action is executed to obtain a corresponding execution result, including:

[0154] The importance of multi-level protection areas is scored based on the relative position between the protection area and the automatic door and the safety risk level. The scoring adopts a quantitative standard of 0-100. The closer the area is to the automatic door, the higher the score is, and the higher the safety risk is. The regional scoring matrix is ​​obtained by weighted average calculation, and each element in the matrix corresponds to the score value of a protection area.

[0155] During the calibration of regional trigger parameters, the trigger thresholds of different regions are set according to the regional scoring matrix. The higher the scoring region, the lower the trigger threshold, so as to achieve timely response to high-risk areas. The gated trigger threshold includes parameters in multiple dimensions such as distance threshold, speed threshold and acceleration threshold.

[0156] Model the target feature data and predict the target's future trajectory. The prediction model input includes the target's historical position, speed and other time series data, and the output is the target's position prediction value at a future time point, generating complete trajectory prediction data.

[0157] The collision risk calculation is based on the cross analysis of the target prediction trajectory and the automatic door motion trajectory. The minimum distance between the two in the time and space dimensions is calculated, and the collision risk level is quantified by combining factors such as relative speed and acceleration. The risk level is divided into three levels: low, medium, and high, corresponding to different control strategies.

[0158] The door control action strategy includes three modes: normal operation, deceleration operation, and emergency stop. When the observation area detects that the collision risk level is low risk, the door control action strategy executed is the normal operation strategy to maintain the normal operation of the automatic door; when the warning area detects that the collision risk level is medium risk, the door control action strategy executed is the deceleration operation strategy, which controls the automatic door to decelerate and stop; when the emergency protection area detects that the collision risk level is high risk, the door control action strategy executed is the emergency stop strategy, which controls the automatic door to emergency stop and retreat. Each strategy corresponds to a set of specific control instruction sequences.

[0159] The execution priority sorting adopts a multi-objective decision-making algorithm, which comprehensively considers safety, efficiency and stability. The highest priority instruction will be selected as the final execution instruction. The execution instruction is converted into a specific motor control signal through the control system to achieve precise control of the automatic door.

[0160] The state analysis module monitors the door control execution status in real time, including parameters such as door position, speed, motor current, etc. By comparing the deviation between the execution instruction and the actual state, the control effect is evaluated and the execution result is generated. The execution result is used for system performance evaluation and optimization improvement.

[0161] This embodiment achieves differentiated management of different risk areas and improves the accuracy of anti-pinch control by combining the importance scoring mechanism of multi-level protection areas. Based on the multi-dimensional trigger threshold and risk level division, a complete anti-pinch strategy system is established to make the control more flexible and reliable. By executing the priority sorting mechanism, while ensuring safety, it takes into account both operational efficiency and stability, thereby improving the user experience. The real-time status analysis and feedback mechanism enable the system to have self-optimization capabilities and continuously improve the accuracy and reliability of anti-pinch control. The overall solution makes full use of artificial intelligence technology to significantly improve the intelligence level and safety performance of automatic door anti-pinch control.

[0162] In one embodiment, the area trigger parameters are adjusted in real time based on the execution results to obtain corresponding system parameters, including:

[0163] In the execution data analysis phase, the system cross-compares the execution sequence data with the preset regional trigger parameters, calculates the deviation between the two, and generates a parameter deviation matrix. The matrix records the degree of difference between the actual execution parameters and the expected parameters.

[0164] When the parameter deviation matrix is ​​subjected to eigendecomposition, the singular value decomposition method is used to extract the main eigenvectors to form a deviation feature group, which reflects the main deviation characteristics during the system operation process.

[0165] In the parameter mapping stage, the system uses a neural network model to perform nonlinear mapping on the deviation feature group, establish a mapping relationship from feature space to parameter space, and obtain mapping parameters. This mapping process is completed based on a pre-trained deep learning model.

[0166] In the adaptive adjustment processing stage, the system dynamically adjusts the regional trigger parameters based on the mapping parameters and historical operation data to generate adjustment parameters. The adjustment process uses an adaptive control algorithm to ensure the stability of parameter adjustment.

[0167] The indicator evaluation for adjustment parameters adopts a multi-dimensional evaluation system, including response time, adjustment accuracy, stability and other indicators. The evaluation rules are based on the scoring criteria set by expert experience, assigning different weights to each indicator and calculating the comprehensive evaluation score.

[0168] During the two-way verification process, the adjustment parameters are compared with the historical optimal parameters and the predicted values ​​of the theoretical model respectively. Through the two dimensions of forward verification and reverse verification, the reliability of the parameters is ensured and the verification parameters are generated.

[0169] In the incremental optimization stage, the gradient descent algorithm is used to fine-tune the parameters based on the verification, find the local optimal solution, and obtain the optimized parameters. The optimization objective function includes two aspects: system performance and stability.

[0170] Parameters are updated recursively, and the optimized parameters are gradually replaced with the original parameters according to the preset update rules. During the update process, parameter change rate limits are set to prevent sudden changes in parameters from causing system instability. Recursive updates ensure that system parameters are continuously optimized as the operating environment changes.

[0171] This embodiment achieves the accuracy and reliability of automatic door anti-pinch control by systematically analyzing the execution results and adjusting the parameters. The cross-deviation verification mechanism based on the execution sequence data effectively identifies and quantifies the parameter deviation, providing an accurate basis for subsequent adjustments. The combined application of feature decomposition and nonlinear mapping makes the adjustment of regional trigger parameters more adaptive and scientific. The adaptive adjustment processing mechanism ensures the real-time and accuracy of parameter adjustment, and at the same time, through the indicator evaluation of preset evaluation rules, the controllability of the adjustment effect is guaranteed. The introduction of the two-way verification mechanism enhances the stability and reliability of parameter adjustment. Through the coordination of incremental optimization and recursive update, the continuous optimization and dynamic adjustment of system parameters are achieved, and the overall performance of the automatic door anti-pinch control system is improved. This method not only improves the accuracy of anti-pinch control, but also reduces the system misjudgment rate, and has strong practical value and promotion significance.

[0172] Reference Figure 2 As shown, the present invention also provides an automatic door anti-pinch control device based on artificial intelligence, which is applied to any one of the above-mentioned automatic door anti-pinch control methods based on artificial intelligence, including:

[0173] The acquisition module is used to obtain the original image data through the image acquisition device arranged at a specific position of the automatic door, and perform area extraction on the original image data to obtain the corresponding real-time image data;

[0174] An analysis module is used to input real-time image data into a preset artificial intelligence deep neural network model for target detection to obtain target recognition results;

[0175] An association module is used to divide the operation area based on the target recognition result to obtain the corresponding multi-level protection area;

[0176] A processing module, the processing module is used to set automatic door trigger parameters based on multi-level protection areas to obtain corresponding area trigger parameters;

[0177] A control module, which performs gate control judgment on the target recognition result according to the regional trigger parameters and the multi-level protection area, and executes the preset gate control action to obtain the corresponding execution result;

[0178] The execution module is used to adjust the regional trigger parameters in real time based on the execution results to obtain corresponding system parameters.

[0179] The present invention provides an artificial intelligence-based automatic door anti-pinch control device, which realizes differentiated protection control of different areas by dividing the target recognition results into multiple protection areas and setting corresponding area trigger parameters based on the division. Compared with the traditional fixed parameter control method, it has higher control accuracy and flexibility, and can better adapt to the safety protection needs in different scenarios. By making door control decisions based on the area trigger parameters and the multiple protection areas and executing corresponding control actions, the intelligent control of the automatic door can be realized, avoiding the problems of false triggering or untimely response that may occur in the traditional control method, and significantly improving the safety performance of the system. Through real-time analysis and feedback of the execution results, the area trigger parameters are dynamically adjusted, so that the system can adaptively optimize the control strategy according to the actual usage situation, which not only improves the environmental adaptability of the system, but also realizes the continuous optimization of anti-pinch control, ensuring the stable and reliable operation of the automatic door system in various application scenarios.

[0180] The present invention also provides a storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute any of the above methods.

[0181] It should be noted that technicians in the relevant technical field can clearly understand that for the convenience and conciseness of description, the specific working process of the system and each module described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0182] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An automatic door anti-pinch control method based on artificial intelligence, characterized in that: include: The original image data is acquired by an image acquisition device arranged at a specific position of the automatic door, and a region is extracted from the original image data to obtain corresponding real-time image data; Inputting the real-time image data into a preset artificial intelligence deep neural network model for target detection to obtain a target recognition result; Based on the target recognition result, the operation area is divided to obtain corresponding multi-level protection areas; Automatic door triggering parameters are set based on the multi-level protection areas to obtain corresponding area triggering parameters; Performing gate control judgment on the target recognition result according to the area trigger parameter and the multi-level protection area, and executing a preset gate control action to obtain a corresponding execution result; The area trigger parameters are adjusted in real time based on the execution results to obtain corresponding system parameters.

2. The automatic door anti-pinch control method based on artificial intelligence according to claim 1 is characterized in that: The original image data is obtained by an image acquisition device arranged at a specific position of the automatic door, and the region of the original image data is extracted to obtain corresponding real-time image data, including: Performing position analysis on the image acquisition device according to preset door parameters to obtain corresponding installation position coordinate data; Performing shooting angle analysis on the image acquisition device according to the installation position coordinate data to obtain corresponding image acquisition angle parameters; Performing perspective transformation on the original image data according to the image acquisition angle parameter to obtain corresponding image perspective data; Performing regional segmentation on the image perspective data according to a preset door area template to obtain a corresponding target area image; Performing edge detection processing on the target area image to obtain corresponding edge feature data; The target area image is morphologically processed according to the edge feature data to obtain the real-time image data.

3. The automatic door anti-pinch control method based on artificial intelligence according to claim 1 is characterized in that: The step of inputting the real-time image data into a preset artificial intelligence deep neural network model for target detection to obtain a target recognition result includes: Inputting the real-time image data into the artificial intelligence deep neural network model for standardized enhancement processing to obtain enhanced image data; Performing feature complexity calculation on the enhanced image data to obtain a feature complexity parameter; The feature complexity parameter is compared according to a preset recognition threshold, and when the feature complexity parameter is greater than the preset recognition threshold, the artificial intelligence deep neural network model is switched to a dual-model recognition mode; When switching to the dual-model recognition mode, the enhanced image data is subjected to feature extraction and fusion processing by two parallel convolutional neural networks to obtain an initial recognition result of the dual-model recognition; When the feature complexity parameter is less than the preset recognition threshold, the artificial intelligence deep neural network model switches to a single model recognition mode; When switching to the single model recognition mode, feature extraction is performed on the enhanced image data through three 3×3 convolutional layers to obtain a corresponding output feature map; Perform BatchNorm processing on the output feature map of the BatchNorm layer to obtain a reconstructed feature map; Performing nonlinear mapping on the reconstructed feature map through a ReLU activation function to obtain a corresponding mapping feature map; Downsampling the mapped feature map through a maximum pooling layer to obtain a primary feature map; The primary feature map is feature mapped through two 1×1 convolutional layers, and feature filtered through a Dropout layer to obtain dimension-reduced feature data; Performing feature classification on the dimension-reduced feature data through a fully connected layer to obtain the initial recognition result of a single model; The recognition result output by the artificial intelligence deep neural network model is reliability verified to obtain the target recognition result.

4. The automatic door anti-pinch control method based on artificial intelligence according to claim 3 is characterized in that: The step of performing feature extraction and fusion processing on the enhanced image data by two parallel convolutional neural networks to obtain the initial recognition result of dual-model recognition includes: The first branch of convolutional neural network processing includes: Processing the enhanced image data by a first residual block to obtain first-level feature data; Processing the first-level feature data by a second residual block to obtain second-level feature data; Processing the second-level feature data by a third residual block to obtain third-level feature data; Processing the third-level feature data by a fourth residual block to obtain a first feature map; Performing attention mechanism processing on the first feature map to obtain first enhanced feature data; The second branch of convolutional neural network processing includes: Processing the enhanced image data through a first depthwise separable convolution block to obtain first layer feature data; Processing the first layer feature data through a second depthwise separable convolution block to obtain second layer feature data; Processing the second layer feature data by a third depth-wise separable convolution block to obtain a third layer feature data; Processing the third layer feature data through a fourth depth-wise separable convolution block to obtain fourth layer feature data; Processing the fourth layer feature data through a fifth depth-wise separable convolution block to obtain a second feature map; Performing pyramid pooling processing on the second feature map to obtain second enhanced feature data; Performing feature map splicing on the first enhanced feature data and the second enhanced feature data to obtain fused feature data; The fused feature data is subjected to feature classification through a fully connected layer to obtain the initial recognition result of dual-model recognition.

5. The automatic door anti-pinch control method based on artificial intelligence according to claim 1 is characterized in that: The operation area is divided based on the target recognition result to obtain corresponding multi-level protection areas, including: Extracting target coordinates from the target recognition result to obtain real-time target coordinate data; Performing target object motion trajectory analysis based on the real-time target coordinate data to obtain a motion parameter set of the target object; Performing double threshold partitioning processing on the motion parameter set to obtain a start frame initial area and an anti-pinch frame initial area; Performing grid mapping processing according to the initial area of ​​the start frame and the initial area of ​​the anti-pinch frame to obtain a basic protection grid; Performing double-layer density calculation on the basic protection grid to obtain a start frame density matrix and an anti-pinch frame density matrix; Performing region expansion processing on the initial region of the start frame according to the start frame density matrix to obtain a start frame expansion region; Performing dynamic shrinkage processing on the initial area of ​​the anti-pinch frame according to the anti-pinch frame density matrix to obtain the core area of ​​the anti-pinch frame; Performing boundary fusion processing on the start frame extension area and the anti-pinch frame core area to obtain a regional transition zone; The start frame extension area and the anti-pinch frame core area are subjected to regional integration processing according to the regional transition zone to obtain the multi-level protection area.

6. The automatic door anti-pinch control method based on artificial intelligence according to claim 1 is characterized in that: The automatic door trigger parameter setting based on the multi-level protection area to obtain the corresponding area trigger parameter includes: Perform hierarchical preset of regional trigger parameters for the working area to obtain corresponding start trigger parameters and anti-pinch trigger parameters; Performing unified sensitivity calibration on the start trigger parameter and the anti-pinch trigger parameter to obtain a corresponding sensitivity coefficient, wherein the value range of the sensitivity coefficient is 0-100; Performing weighted calculation on the start trigger parameter and the anti-pinch trigger parameter according to the sensitivity coefficient to obtain a weighted trigger parameter, wherein the weight of the start trigger parameter is 0.3, and the weight of the anti-pinch trigger parameter is 0.7; Performing timing analysis on the weighted trigger parameters to obtain timing correlation parameters including trigger time interval, duration and response speed; Performing threshold optimization on the weighted trigger parameter according to the timing correlation parameter to obtain an optimized trigger threshold; Adaptively adjusting the optimization trigger threshold to obtain a dynamic trigger parameter; The dynamic trigger parameter is subjected to safety redundancy processing to obtain the regional trigger parameter, wherein the safety redundancy processing includes setting a backup trigger parameter.

7. The automatic door anti-pinch control method based on artificial intelligence according to claim 1 is characterized in that: The gate control determination is performed on the target recognition result according to the area trigger parameter and the multi-level protection area, and a preset gate control action is executed to obtain a corresponding execution result, including: Extracting target features from the target recognition results to obtain a target feature data set; Scoring the importance of the multi-level protection areas according to the target feature data set to obtain a regional scoring matrix; Calibrate the regional trigger parameters according to the regional scoring matrix to obtain a gated trigger threshold; Performing time series analysis on the target feature data set in the multi-level protection area to obtain target motion trajectory prediction data; Calculate the collision risk according to the target motion trajectory prediction data and the gate trigger threshold to obtain a collision risk level; Selecting a corresponding door control action strategy according to the collision risk level to obtain a door control instruction sequence; Sorting the execution priority of the gated instruction sequence to obtain a final execution instruction; According to the final execution instruction, the automatic door is controlled to execute the corresponding action to obtain the door control execution state; Performing state analysis on the gated execution state to obtain the execution result.

8. The automatic door anti-pinch control method based on artificial intelligence according to claim 1 is characterized in that: The real-time adjustment of the area trigger parameter based on the execution result to obtain the corresponding system parameter includes: Performing execution data analysis on the execution result to obtain corresponding execution sequence data; Perform cross deviation verification on the regional trigger parameters according to the execution sequence data to obtain a corresponding parameter deviation matrix; Performing eigendecomposition on the parameter deviation matrix to obtain a corresponding deviation feature group; Performing nonlinear mapping on the area trigger parameters according to the deviation feature group to obtain corresponding mapping parameters; Adaptively adjust the area trigger parameter according to the mapping parameter to obtain a corresponding adjustment parameter; Performing an index evaluation on the adjustment parameter according to a preset evaluation rule to obtain a corresponding evaluation index; Performing bidirectional verification on the adjustment parameters according to the evaluation index to obtain corresponding verification parameters; Incrementally optimizing the verification parameters to obtain corresponding optimization parameters; The area trigger parameters are recursively updated according to the optimization parameters to obtain corresponding system parameters.

9. An automatic door anti-pinch control device based on artificial intelligence, characterized in that: The artificial intelligence-based automatic door anti-pinch control method applied to any one of claims 1 to 8 above comprises: An acquisition module, the acquisition module is used to acquire original image data through an image acquisition device arranged at a specific position of the automatic door, and perform region extraction on the original image data to obtain corresponding real-time image data; An analysis module, wherein the analysis module is used to input the real-time image data into a preset artificial intelligence deep neural network model for target detection to obtain a target recognition result; An association module, the association module is used to divide the operation area based on the target recognition result to obtain corresponding multi-level protection areas; A processing module, the processing module is used to set automatic door trigger parameters based on the multi-level protection area to obtain corresponding area trigger parameters; A control module, wherein the control module performs a gate control judgment on the target recognition result according to the area trigger parameter and the multi-level protection area, and executes a preset gate control action to obtain a corresponding execution result; An execution module is used to adjust the area trigger parameters in real time based on the execution result to obtain corresponding system parameters.

10. A storage medium, characterized in that: Computer instructions are stored, and the computer instructions are used to make a computer execute the method according to any one of claims 1 to 8.

Images