A self-calibration method and system for a gate system capable of shielding against multi-source interference.

Abstract

This invention provides a self-calibration method and system for a turnstile system capable of shielding against multi-source interference. Applied to the field of access control technology, the method involves acquiring images using a 3D vision device, detecting the turnstile blades, clustering their spatial distribution, and calculating geometric calibration parameters to achieve channel self-calibration. Based on depth images, background modeling and updating are performed, incorporating static structures into the background to form static shielding areas. The active areas of the blades are determined according to the geometric calibration parameters. When blade movement is detected, dynamic shielding is applied to the foreground within the active area to suppress static appearance interference and blade opening / closing interference. The system includes modules for image acquisition, blade detection and clustering, geometric parameter calculation, background modeling and updating, static appearance shielding, blade movement shielding, and real-time detection optimization. This invention reduces deployment and maintenance costs, decreases false positives and false negatives, and improves the stability and robustness of detection and measurement in complex environments.

Description

Technical Field

[0001] This invention belongs to the field of access detection technology, specifically relating to a self-calibration method and system for a gate system capable of shielding against multi-source interference. Background Technology

[0002] As people increasingly rely on urban public transportation, the efficiency and safety of subways, as an efficient and convenient mode of travel, are of paramount importance. Subway turnstiles, as crucial facilities for passenger entry and exit, directly impact passenger experience and the safety of subway operations with their accurate operational status.

[0003] Currently, to more accurately track passenger positions and status at turnstiles, many subway systems employ advanced vision systems to capture and analyze passenger behavior. These systems typically include high-resolution cameras, depth cameras (such as Time-of-Flight (TOF) technology), image processing software, and related sensors and image processing equipment. By monitoring passenger behavior at turnstiles in real time, these systems can effectively identify whether passengers are passing through the turnstiles correctly and promptly detect any anomalies.

[0004] However, existing vision systems face two core technical challenges in practical applications: first, how to accurately determine the position and orientation of the vision device itself in three-dimensional space; and second, how to effectively eliminate the interference of the structural appearance and fan blade movement of different types of turnstiles on the detection accuracy.

[0005] In existing technologies, the self-localization of vision devices is mostly achieved through the following methods:

[0006] (1) Mechanical positioning device and measuring tools: Traditional measuring tools such as level, rangefinder and angle measuring instrument are used to guide the installers to install the equipment, and manual measurement and adjustment are relied upon to ensure the installation accuracy of the camera.

[0007] (2) Integrated gate support system: The fixed support structure is designed to be integrated with the gate. The installation angle and position of the camera are limited by the pre-designed hardware structure to reduce installation errors.

[0008] (3) Calibration plate calibration: After installation, a pre-set calibration object (such as a checkerboard, calibration plate, etc.) is placed in the channel to help the vision device calculate its accurate position in three-dimensional space using the calibration object with known geometric features.

[0009] (4) Multi-point reference method: The coordinate system is established by using multiple reference points with known locations.

[0010] However, the above positioning method has the following limitations:

[0011] The calibration process is complex and cumbersome: Traditional calibration methods require manual placement of calibration objects (such as checkerboards, calibration boards, etc.), which is time-consuming and prone to errors. It is particularly inconvenient to operate in complex environments such as subways and requires on-site operation by professionals.

[0012] High installation accuracy requirements: Existing integrated gate support systems usually need to be installed in precise positions and angles, which requires a high level of technical expertise from the installers. Due to limitations in the construction conditions at the installation site, installers cannot accurately determine the optimal installation position for the cameras.

[0013] Highly dependent on the environment: The calibration-based method requires placing a reference object in the channel, which is difficult to implement in the actual operating environment and is easily affected by environmental changes.

[0014] High degree of manual intervention: It requires a lot of manual measurement, adjustment and verification work, which increases the complexity and cost of installation and maintenance.

[0015] In addition, most existing technologies shield visual inspection from interference using the following methods:

[0016] (1) Interference elimination method based on region limitation: By manually presetting a fixed detection range and mask area, the gate equipment is excluded, the known equipment interference sources are shielded, and the appearance of the gate is avoided from interfering with the detection results.

[0017] (2) Deep learning-based target detection and filtering method: Using target detection algorithms such as YOLO and SSD, a special fan blade recognition model is trained through a large amount of labeled data to automatically learn the visual features of fan blades and realize intelligent recognition and filtering of different types of fan blades.

[0018] The above-mentioned interference shielding methods have the following shortcomings:

[0019] Poor adaptability: Different channel environments require different turnstile and fan blade models and specifications. Existing fixed mask and preset area methods cannot adapt to diverse turnstile types and require separate configuration for each type of turnstile.

[0020] Unstable detection accuracy: The rapid movement of the fan blades will produce motion blur and occlusion in the image. Traditional image processing algorithms are not adaptable to complex scenes and changing environments, which affects the accurate detection of passengers.

[0021] Fixed area limitation: Existing area limitation methods are too simple and crude, which may miss important detection areas or contain too much interference information, and cannot be dynamically adjusted according to the actual fan blade position.

[0022] High computational requirements and poor real-time performance: Schemes based on large-scale deep learning networks and relying entirely on end-to-end detection have high computational complexity and are difficult to meet real-time requirements while ensuring high frame rates on edge embedded platforms.

[0023] Weak generalization ability: Models trained for specific types of fan blades are difficult to adapt to new types of turnstiles, requiring data to be collected and models to be trained again.

[0024] Therefore, there is an urgent need for a visual inspection method and system for turnstiles that can automatically acquire key geometric parameters of turnstile channels without relying on manual calibration, and can effectively suppress multi-source interference such as the static appearance structure of the turnstile and the dynamic movement of the fan blades, so as to improve the adaptability and detection robustness of different gate types and different installation conditions. Summary of the Invention

[0025] In view of the above-mentioned problems in the prior art, the purpose of this invention is to provide a self-calibration method for gate systems that can achieve multi-source interference shielding, reduce deployment and maintenance costs, reduce false detections and missed detections, and improve the stability and robustness of detection and measurement in complex environments.

[0026] A self-calibration method for a gate system capable of shielding against multi-source interference includes the following steps:

[0027] A 3D vision device installed above the gate acquires a 3D image containing the gate's fan blades, performs target detection on the gate's fan blades in the 3D image, and obtains the fan blade detection result.

[0028] Based on the spatial distribution of multiple fan blades in the fan blade detection results, the fan blades are clustered and grouped, and geometric calibration parameters, including the center of the gate channel, are calculated based on the positions of the grouped fan blades to achieve self-calibration of the gate channel;

[0029] Background modeling is performed based on the depth image acquired by the three-dimensional vision device, and static structures including gate housing and columns are included in the background to form a static shielding area.

[0030] Based on the geometric calibration parameters, the active area corresponding to each fan blade is determined. When the fan blade is detected to be in motion, the foreground within the active area is shielded to form a dynamic shielding area for the fan blade, thereby simultaneously suppressing the interference of the static structure of the gate and the movement of the fan blade on the detection of human bodies in the channel in the same image.

[0031] Preferably, in the foreground area after the combined processing of the static shielding area and the dynamic shielding area of ​​the fan blade, the human body in the channel is detected, and the height of the human body is measured in combination with the geometric calibration parameters and depth information.

[0032] Preferably, when clustering and grouping the fan blades based on the fan blade detection results, the K-means clustering algorithm is used to cluster the center points of the fan blades. When the gate is a flap gate, the number of clusters K is set to 4, and when the gate is a scissor gate, the number of clusters K is set to 2.

[0033] Preferably, after clustering the fan blade center points, outlier data points are removed using the interquartile range method. Fan blade center points outside the preset range are identified as outliers and removed, leaving only a stable set of fan blade trajectory points for calculating geometric calibration parameters.

[0034] Preferably, background modeling includes: continuously acquiring multiple frames of depth images when no passengers are passing through the turnstile channel and the fan blades are stationary; calculating the median value of each pixel position in the time dimension to form a depth median map as a background depth template; during operation, for pixels whose depth difference between each frame of depth image and the background depth template is less than a preset static threshold, they are marked as background pixels and assigned to the static shielding area corresponding to the static structure; when the depth difference is not less than the preset static threshold, the pixel is marked as a foreground pixel, thereby obtaining an initial foreground binary map.

[0035] Preferably, when processing the initial foreground binary image formed by pixels with a depth difference not less than a static threshold, morphological opening and closing operations are performed sequentially, wherein the opening operation adopts the order of erosion followed by dilation, and the closing operation adopts the order of dilation followed by erosion.

[0036] The initial foreground image is divided into multiple candidate target regions by a connected component labeling algorithm, and then noisy regions that do not conform to the shape characteristics of a human body or fan blade are filtered out.

[0037] Preferably, the active area corresponding to each fan blade is a rectangular area or a fan-shaped area based on the center position of the fan blade, and the rectangular area or the fan-shaped area is expanded outward in the horizontal and vertical directions to form a dynamic shielding area for the fan blade.

[0038] Preferably, the inter-frame difference index used to determine whether the fan blades are in motion includes at least one of the following: the average depth difference within the candidate region for dynamic masking of the fan blades, the maximum depth difference, and the ratio of difference pixels.

[0039] When the inter-frame difference index in a certain area exceeds the set fan blade motion threshold, the fan blade is determined to be in motion; conversely, when the inter-frame difference index is continuously below the threshold and remains below it for a certain time window, the fan blade is determined to be in a stationary state.

[0040] Preferably, during long-term system operation, the passenger passage positions detected in the foreground area after joint processing of the static shielding area and the dynamic shielding area of ​​the fan blade are statistically analyzed. When the offset of the statistically obtained passage position distribution relative to the center of the gate channel exceeds a preset threshold, the fan blade-based self-calibration process is automatically triggered to compensate for geometric calibration errors caused by camera installation offset or environmental changes.

[0041] The second objective of this invention is to provide a self-calibration system for a gate system capable of shielding against multi-source interference, comprising:

[0042] A TOF depth camera is used to acquire three-dimensional image data of the turnstile channel;

[0043] An image acquisition module, connected to the TOF depth camera, is used to receive and output a sequence of gate images containing RGB images and depth images;

[0044] The fan blade detection and clustering module is connected to the image acquisition module and is used to perform target detection on the gate fan blades in the gate image sequence and to cluster the fan blade detection results.

[0045] The fan blade and vision module geometric parameter calculation module is connected to the fan blade detection and clustering module. It is used to calculate the geometric calibration parameters, including the gate channel center, channel width and camera installation height, based on the spatial distribution of the grouped fan blades, and generate the corresponding parameter configuration.

[0046] The background modeling and updating module is connected to the image acquisition module and is used to build and update the background model based on the depth image.

[0047] The static gate appearance shielding module is connected to the background modeling and updating module and is used to generate a static shielding area for the static structure including the gate box and columns based on the background model.

[0048] The fan blade motion shielding module is connected to the fan blade detection and clustering module. It is used to determine the active area of ​​each fan blade based on the fan blade geometric calibration parameters, and to dynamically shield the corresponding active area when the fan blade is detected to be in motion.

[0049] The real-time detection optimization module is connected to the static gate appearance shielding module and the fan blade motion shielding module. It is used to extract the foreground and detect human targets in the channel under the joint constraints of the static shielding area and the fan blade dynamic shielding area, and output the target detection results.

[0050] The beneficial effects of this invention are: the gate system self-calibration method and system that can achieve multi-source interference shielding constructs a closed-loop processing system that can achieve self-calibration, static shielding, dynamic shielding, and real-time detection and optimization.

[0051] Using the gate fan blades as a natural calibration object, the spatial topological relationship between the upper and lower, left and right fan blades is automatically established through fan blade detection and clustering. Then, key geometric calibration parameters such as channel center, channel width and camera installation height are calculated online, realizing rapid deployment and adaptive calibration without calibration board, manual measurement and parameter adjustment, significantly reducing installation and maintenance costs and reducing human error.

[0052] Simultaneously, TOF depth images are used for background modeling and updating, automatically absorbing the static appearance structure such as the gate housing and columns as the background to form a static shielding layer. Based on this, the active area of ​​the fan blades is determined according to the geometric calibration results. Combined with the fan blade motion state discrimination, dynamic shielding is implemented in the active area. A multi-source interference joint suppression mechanism with static appearance shielding and fan blade dynamic shielding is constructed, thereby effectively weakening the false foreground and false triggering caused by reflection and occlusion of irregular gate structures and high-speed opening and closing of fan blades. This significantly reduces false detection, false missed detection and jitter in human body detection and height measurement, and improves the stability, accuracy and robustness of detection results.

[0053] In addition, the calibration parameters can be permanently output as a system configuration file and loaded directly later, enabling the system to quickly recover its working state after restarting or slight changes in position, thus improving the gate's universal adaptability, continuous operation capability, and safety management efficiency under complex field conditions. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0055] Figure 1 This is a flowchart of the method of the present invention;

[0056] Figure 2 This is a flowchart of the data acquisition process of this invention;

[0057] Figure 3 This is a flowchart of the background model construction of the present invention;

[0058] Figure 4 This is a flowchart of the dynamic shielding process of the present invention. Detailed Implementation

[0059] Example 1

[0060] A self-calibration method for a gate system capable of shielding against multi-source interference includes the following steps:

[0061] S1. System Initialization and Data Acquisition:

[0062] A 3D vision device is pre-installed above the turnstile. This device includes at least a depth imaging camera, a processor, and a storage module. After the system is powered on, it first loads the camera's intrinsic parameters and factory calibration parameters, and initializes parameters such as exposure, gain, and frame rate to ensure that the image quality under the current ambient lighting conditions meets the requirements of subsequent algorithm processing.

[0063] Subsequently, the system enters calibration mode. Within a time window when the turnstile is in normal working condition but no passengers are passing through the channel, it continuously acquires multiple frames of data streams containing both RGB and depth images. For each frame, a preset object detection network or traditional image processing algorithm is invoked to detect the turnstile blades, obtaining information such as the coordinates of the detection box containing the blade position and the confidence level. These detection results, along with the depth data of the corresponding frame, are stored in a cache queue for subsequent clustering analysis and geometric parameter estimation. If necessary, a minimum acquisition duration or frame count can be set, for example, an acquisition duration of 10–30 seconds or a cumulative frame count of 300–1000 frames, to ensure that the blades are fully sampled throughout the entire opening and closing motion.

[0064] like Figure 2 As shown, in a specific embodiment, the depth images acquired during calibration are used for target detection. When a fan blade is detected, the current detection box is recorded; if no fan blade is detected, depth images are acquired again. Furthermore, it is determined whether the number of image frames for recording the detection box is not less than 200. If it is not less, the clustering analysis process begins; if it is less, depth images are acquired again until the number of image frames for recording the detection box is not less than 200, at which point the clustering analysis process can begin.

[0065] S2. Preprocessing of fan blade inspection results

[0066] Read the fan blade detection boxes for each frame from the buffer queue. For each detection box, calculate the pixel coordinates of its rectangular center point and construct a set of fan blade center points. The expression for the pixel coordinates of the detection box center point is:

[0067]

[0068] in, and The first The coordinates of the top left and bottom right corners of each detection box.

[0069] In one specific embodiment, detection results with low confidence can be filtered, such as retaining only detection boxes with a confidence level greater than a preset threshold (e.g., 0.7 or 0.8), to reduce the interference of false detections on clustering results. For multiple duplicate or highly overlapping detection boxes in the same frame, non-maximum suppression or centroid merging can be used to merge them, ensuring that each fan blade corresponds to only a limited number of center points in each frame, thereby improving the stability of subsequent clustering and statistical analysis.

[0070] S3, Fan-shaped clustering and outlier removal

[0071] The K-means algorithm is preferred for clustering, but it is not limited to this. Other clustering methods such as Gaussian Mixture Model (GMM) and Density Clustering (DBSCAN) can also be used.

[0072] Furthermore, cluster analysis is performed on the preprocessed set of fan blade center points. In the flapping door scenario, the number of clusters K is set to 4 to correspond to the upper and lower fan blades on the left and right sides; in the scissor door scenario, K is set to 2 to correspond to one fan blade on each side; when the system supports other door types, the value of K can be adjusted according to the door type configuration.

[0073] This embodiment uses Let's take an example to illustrate: Set the number of clusters Iterative optimization of the objective function:

[0074]

[0075] in, For the first One cluster, The center of this cluster, These are the normalized pixel coordinates of the center point of the detection box.

[0076] After clustering, outlier data points are removed using the interquartile range (IQR) method. For each cluster... Calculate the pixel coordinates of the center point of the normalized detection box. and Quartiles of coordinates: First quartile: , Third quartile: , Then the interquartile range: , Retain the points that satisfy the following conditions:

[0077]

[0078]

[0079] Center points outside the range are identified as outliers and removed, retaining only the set of center points that are stably distributed and represent the true trajectory of the fan blades.

[0080] The outlier removal operation described above can effectively suppress outliers introduced by factors such as false detection, passenger obstruction, and sudden changes in lighting, thereby improving the reliability of subsequent channel center estimation.

[0081] S4, Channel Center Calculation

[0082] For the door knocking scenario, the average coordinates of the cluster centers of each fan blade are calculated as the stable center position of that fan blade; then the geometric centers of the four fan blade centers are calculated, and after verifying that the four fan blade centers form a valid quadrilateral, the channel center is calculated:

[0083]

[0084] in, For the first The center coordinates of each fan blade.

[0085] Simultaneously, by utilizing the depth information corresponding to these center points, combined with the camera's intrinsic and extrinsic parameters, the channel center can be transformed from the pixel coordinate system to the camera coordinate system or the world coordinate system. This allows for the further calculation of geometric calibration parameters such as channel width, fan blade spacing, camera mounting height, and the horizontal offset of the camera relative to the channel center. This set of geometric parameters can serve as the basis for subsequent human height measurement and the construction of the fan blade shielding area.

[0086] For scissor doors or other door types, the appropriate fan blade combination can be selected based on the door type characteristics for subsequent calculations.

[0087] Furthermore, for each clustering result after outlier removal, gate labels are assigned based on spatial location. First, the clusters are divided into upper fan blade group A and lower fan blade group B according to the y-coordinate of the center point. Within each group, the clusters are then sorted by x-coordinate from smallest to largest. Therefore, the two upper fan blades are labeled (A1, A2), and the two lower fan blades are labeled (B1, B2), thus establishing a fan blade topology consistent with the physical structure of the turnstile. By assigning labels to the fan blades, the aligning turnstiles of the moving turnstiles can be directly blocked during dynamic shielding, which helps improve the efficiency of dynamic shielding.

[0088] S5, Deep Background Modeling and Static Gate Shielding

[0089] like Figure 3 As shown, during the calibration phase or daily work, several time periods are selected in which no passengers pass through the channel and the fan blades are stationary. Multiple frames of depth images are continuously acquired, and statistical analysis is performed on each pixel in the time dimension.

[0090] Structures that do not change over time, such as gate housings, columns, and fixed guardrails, are automatically included in the background model. During runtime, the newly acquired depth image for each frame is compared pixel by pixel with the background depth image. When the difference is less than a preset static threshold, the pixel is marked as a background pixel, and the corresponding pixel can be removed from the foreground mask, thus achieving long-term shielding of static structures such as the gate housing appearance and columns. When the difference exceeds the preset static threshold, the pixel is marked as a foreground pixel, thereby obtaining the initial foreground binary image.

[0091] Specifically, a depth image is defined as... and background model Calculate the foreground mask:

[0092]

[0093] in, This is a static threshold, with a typical value of 50 mm.

[0094] This depth-based background modeling, compared to manually drawing masks or using fixed templates, can adapt to different gate models and appearance structures, improving the versatility and flexibility of deployment.

[0095] Furthermore, considering that depth noise and environmental disturbances may cause isolated noise points and holes in the foreground image, morphological opening operations (erosion followed by dilation) can be performed sequentially on the initial foreground image to remove small-area noise points, and then closing operations (dilation followed by erosion) can be performed to fill small holes in the foreground region.

[0096] Specifically, morphological operations are used to remove noise and fill holes:

[0097]

[0098] in, The structuring element uses a 5×5 rectangular core.

[0099] The processed foreground binary image can be divided into multiple connected components using a connected component labeling algorithm, resulting in a series of candidate target regions. Each region corresponds to the projection of a dynamic object, which in this invention mainly includes the passenger's body and the fan blades in the opening and closing process. For connected components with too small an area or shapes that do not meet expectations, further filtering can be performed using geometric features such as area thresholds and aspect ratios to avoid incorrectly identifying noisy regions as valid candidate targets.

[0100] In one specific implementation, the median of depth values ​​across multiple frames can be calculated for each pixel location to obtain a depth median map as a background model; in another implementation, a background depth image can be constructed using methods such as mode, weighted average, or Gaussian model fitting to adapt to the noise characteristics of different depth sensors.

[0101] S6. Definition of fan blade active area and detection of motion status

[0102] Based on the fan blade center position and channel geometry parameters obtained from S4, an active region covering the possible range of motion of each fan blade is predefined in the image plane or three-dimensional space. This active region can be a rectangular, fan-shaped, or polygonal area with the fan blade center as the base point, and the boundary range is set according to parameters such as fan blade length, opening and closing angle, and installation height. To improve the robustness of shielding, a certain safety margin can be added in both the horizontal and vertical directions based on the fan blade geometry contour, for example, by extending it by 5 to 20 pixels, to form a fan blade shielding candidate region.

[0103] Specifically, based on the calibration results, the active area of ​​each fan blade is defined:

[0104]

[0105] Considering the range of motion, expand the shielding area:

[0106]

[0107] in, , To extend the margin, it adaptively adjusts according to the fan blade type.

[0108] like Figure 4 As shown, during system operation, the inter-frame difference index, such as mean difference, maximum difference, or difference pixel ratio, is calculated by comparing the depth value or gray value of the current frame with that of the previous frame in each fan blade shielding candidate area.

[0109] In this embodiment, fan blade motion is detected by inter-frame difference:

[0110]

[0111] when At that time, it is determined that the fan blades are in motion. The preset blade motion threshold is automatically adjusted based on historical operating data, ensuring that the false detection rate and false negative rate of blade motion detection are within the preset range.

[0112] That is, when the differential index in a certain area exceeds the blade movement threshold, the blade is determined to be in motion; conversely, when the differential index remains below the threshold for a certain period of time, the blade is determined to be stationary.

[0113] Among them, inter-frame difference The calculation is based on the unexpanded active region. The dynamic masking operation is based on the expanded region. .

[0114] The results of the above-mentioned determination of the fan blade motion status can be used for dynamic shielding, as well as for monitoring the operation status of the gate and diagnosing faults.

[0115] Furthermore, in frames where the fan blades are determined to be in motion, logical operations are performed on the shielding candidate region corresponding to the fan blade and the foreground binary image obtained in S5: the preferred method is to force the foreground pixels in the shielding candidate region to be set to zero, thereby removing the foreground part caused by the fan blade movement from the foreground image; for fan blades that are not in motion, the foreground pixels in their shielding candidate region can be retained for subsequent gate status recognition or anomaly detection.

[0116] Meanwhile, with the help of the depth background model generated by S5, static structures such as turnstile housings and columns are marked as background and will not appear in the foreground binary image. Through the superposition of the above static and dynamic shielding, this method can simultaneously suppress static housing appearance interference and dynamic fan blade motion interference in the same foreground image, providing a clean target area for subsequent human body detection and height measurement. Compared with shielding methods that only use fixed masks or are based solely on target detection, this multi-source interference joint suppression method can significantly reduce false detections and false negatives caused by fan blades sweeping across the human body, and improve detection stability under dense traffic conditions.

[0117] S7. Human body detection, height measurement, and parameter saving within the channel.

[0118] After processing by S6, a human detection algorithm or a human contour extraction algorithm is called to detect and segment passengers in the foreground, obtaining human contour boundaries or key point information. For each detected human target, based on its top pixel position in the image, combined with the corresponding depth value and calibration parameters such as camera mounting height and channel center position determined in S4, the image coordinates and depth information are converted into actual spatial height using triangulation or coordinate transformation, thereby obtaining the passenger's height estimate.

[0119] Once the calibration process has been executed multiple times and the results are stable, the latest channel center coordinates, channel width, camera height, fan blade center, fan blade active shielding area parameters, background model update parameters, and shielding enable / disable switch, etc., are packaged into a configuration file or stored in non-volatile memory so that the system can quickly return to normal operation without repeated calibration when it is powered on again.

[0120] In some embodiments, the statistical distribution of human body passage positions can also be monitored during long-term operation. When a significant shift in the center of the channel is detected, a new round of self-calibration process is automatically triggered to compensate for errors caused by equipment offset or environmental changes, thereby further ensuring the detection accuracy and robustness of the system during long-term use.

[0121] Example 2

[0122] like Figure 1 As shown, a gate system self-calibration system capable of shielding multi-source interference includes a TOF depth camera, an image acquisition module, a fan blade detection and clustering module, a fan blade and vision module geometric parameter calculation module, a background modeling and updating module, a static gate appearance shielding module, a fan blade motion shielding module, and a real-time detection and optimization module.

[0123] A TOF depth camera is used to acquire three-dimensional image data of the gate passage.

[0124] The image acquisition module is connected to a TOF depth camera to receive and output a sequence of gate images containing RGB and depth images.

[0125] The fan blade detection and clustering module is connected to the image acquisition module and is used to perform target detection on the fan blades of the gate in the gate image sequence and to cluster the fan blade detection results.

[0126] The fan blade and vision module geometric parameter calculation module is connected to the fan blade detection and clustering module. It is used to calculate the geometric calibration parameters, including the gate channel center, channel width and camera installation height, based on the spatial distribution of the grouped fan blades, and generate the corresponding parameter configuration.

[0127] The background modeling and updating module is connected to the image acquisition module and is used to build and update the background model based on the depth image.

[0128] The static gate appearance shielding module is connected to the background modeling and updating module, and is used to generate a static shielding area for the static structure including the gate box and columns based on the background model.

[0129] The fan blade motion shielding module is connected to the fan blade detection and clustering module. It is used to determine the active area of ​​each fan blade based on the fan blade geometric calibration parameters, and to dynamically shield the corresponding active area when the fan blade is detected to be in motion.

[0130] The real-time detection optimization module is connected to the static gate appearance shielding module and the fan blade motion shielding module. It is used to extract the foreground and detect human targets in the channel under the joint constraints of the static shielding area and the fan blade dynamic shielding area, and output the target detection results.

[0131] Through the collaborative efforts of various modules, a closed-loop processing system capable of self-calibration, static shielding, dynamic shielding, and real-time detection and optimization has been constructed.

[0132] The self-calibration design completely eliminates traditional calibration tools such as checkerboard patterns and calibration boards. Calibration is completed solely by detecting the natural movement characteristics of the fan blades, allowing even non-professionals to install the equipment. This reduces the calibration process, which previously required one hour for a professional engineer, to just five minutes, and it boasts high tolerance for installation position deviations, accurately calibrating even with an installation error of ±10cm. Furthermore, through cluster analysis and statistical optimization, the channel center positioning accuracy reaches ±3cm. Based on fan blade position and depth information, the camera installation height is automatically calculated with an accuracy of ±2cm. It also supports continuous online optimization, constantly improving calibration accuracy as data accumulates.

[0133] The shielding range is dynamically adjusted according to the real-time position of the fan blades to avoid over-shielding or under-shielding, and to effectively shield false targets caused by fan blade movement, reducing the false detection rate from 15% to below 1.5%. At the same time, it can automatically identify and adapt to various fan blade movement modes such as flap gates, scissor gates, swing gates, and wing gates, with high adaptability.

[0134] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A self-calibration method for a gate system capable of shielding against multi-source interference, characterized in that, Includes the following steps: A 3D vision device installed above the gate acquires a 3D image containing the gate's fan blades, performs target detection on the gate's fan blades in the 3D image, and obtains the fan blade detection result. Based on the spatial distribution of multiple fan blades in the fan blade detection results, the fan blades are clustered and grouped, and geometric calibration parameters, including the center of the gate channel, are calculated based on the positions of the grouped fan blades to achieve self-calibration of the gate channel; Background modeling is performed based on the depth image acquired by the three-dimensional vision device, and static structures including gate housing and columns are included in the background to form a static shielding area. Based on the geometric calibration parameters, the active area corresponding to each fan blade is determined. When the fan blade is detected to be in motion, the foreground in the active area is shielded to form a dynamic shielding area for the fan blade, thereby suppressing the interference of the static structure of the gate and the movement of the fan blade on the detection of human body in the channel in the same image. When clustering and grouping the fan blades based on the fan blade detection results, the K-means clustering algorithm is used to cluster the fan blade center points. When the gate is a flap gate, the number of clusters K is set to 4, and when the gate is a scissor gate, the number of clusters K is set to 2. After clustering the fan blade center points, outlier data points are removed using the interquartile range method. Fan blade center points outside the preset range are identified as outliers and removed. Only the stable set of fan blade trajectory points is retained for calculating geometric calibration parameters.

2. The self-calibration method for a gate system capable of shielding multi-source interference according to claim 1, characterized in that, In the foreground area after the combined processing of the static shielding area and the dynamic shielding area of ​​the fan blades, the human body in the channel is detected, and the height of the human body is measured in combination with the geometric calibration parameters and depth information.

3. The self-calibration method for a gate system capable of shielding multi-source interference according to claim 1, characterized in that, The background modeling includes: continuously acquiring multiple frames of depth images when there are no passengers passing through the turnstile channel and the fan blades are stationary; calculating the median value of each pixel position in the time dimension to form a depth median map as a background depth template; during operation, for pixels whose depth difference between each frame of depth image and the background depth template is less than a preset static threshold, they are marked as background pixels and assigned to the static shielding area corresponding to the static structure; when the depth difference is not less than the preset static threshold, the pixel is marked as a foreground pixel, thereby obtaining an initial foreground binary map.

4. The self-calibration method for a gate system capable of multi-source interference shielding according to claim 3, characterized in that, When processing the initial foreground binary image formed by pixels with a depth difference not less than the static threshold, morphological opening and closing operations are performed sequentially, with the opening operation following the order of erosion followed by dilation and the closing operation following the order of dilation followed by erosion. The initial foreground image is divided into multiple candidate target regions by a connected component labeling algorithm, and then noisy regions that do not conform to the shape characteristics of a human body or fan blade are filtered out.

5. The self-calibration method for a gate system capable of shielding multi-source interference according to claim 1, characterized in that, The active area corresponding to each fan blade is a rectangular or fan-shaped area based on the center position of the fan blade, and the rectangular or fan-shaped area is expanded outward in the horizontal and vertical directions to form a dynamic shielding area for the fan blade.

6. The self-calibration method for a gate system capable of shielding multi-source interference according to claim 5, characterized in that, Inter-frame difference metrics used to determine whether the fan blades are in motion include at least one of the following: the average depth difference within the dynamic masking candidate region of the fan blades, the maximum depth difference, and the differential pixel ratio. When the inter-frame difference index in a certain area exceeds the set fan blade motion threshold, the fan blade is determined to be in motion. Conversely, if the inter-frame difference index remains below the threshold for a certain time window, the fan blade is determined to be in a stationary state.

7. The self-calibration method for a gate system capable of shielding multi-source interference according to claim 2, characterized in that, During long-term system operation, the passenger passing positions detected in the foreground area after joint processing of the static shielding area and the dynamic shielding area of ​​the fan blade are statistically analyzed. When the offset of the statistically obtained passing position distribution relative to the center of the gate channel exceeds a preset threshold, the fan blade-based self-calibration process is automatically triggered to compensate for geometric calibration errors caused by camera installation offset or environmental changes.

8. A gate system self-calibration system capable of achieving multi-source interference shielding, used to implement the gate system self-calibration method capable of achieving multi-source interference shielding as described in any one of claims 1 to 7, characterized in that, include: A TOF depth camera is used to acquire three-dimensional image data of the turnstile channel; An image acquisition module, connected to the TOF depth camera, is used to receive and output a sequence of gate images containing RGB images and depth images; The fan blade detection and clustering module is connected to the image acquisition module and is used to perform target detection on the gate fan blades in the gate image sequence and to cluster the fan blade detection results. The fan blade and vision module geometric parameter calculation module is connected to the fan blade detection and clustering module. It is used to calculate the geometric calibration parameters, including the gate channel center, channel width and camera installation height, based on the spatial distribution of the grouped fan blades, and generate the corresponding parameter configuration. The background modeling and updating module is connected to the image acquisition module and is used to build and update the background model based on the depth image. The static gate appearance shielding module is connected to the background modeling and updating module and is used to generate a static shielding area for the static structure including the gate box and columns based on the background model. The fan blade motion shielding module is connected to the fan blade detection and clustering module. It is used to determine the active area of ​​each fan blade based on the fan blade geometric calibration parameters, and to dynamically shield the corresponding active area when the fan blade is detected to be in motion. The real-time detection optimization module is connected to the static gate appearance shielding module and the fan blade motion shielding module. It is used to extract the foreground and detect human targets in the channel under the joint constraints of the static shielding area and the fan blade dynamic shielding area, and output the target detection results.

Images