A non-inductive traffic management method and system

Abstract

The present application relates to the field of image processing, in particular to a non-sensing passing management method and system, wherein the method comprises: collecting a depth image of a passing area, and segmenting a foreground target from the depth image; for each pixel point in the foreground target area, calculating a depth fluctuation value in the neighborhood window of the point; screening out structural abnormal points; calculating the spatial dispersion of the structural abnormal points relative to the geometric centroid of the foreground target; determining the morphological abnormality factor of the foreground target; calculating a tailing potential score; combining the tailing potential score with at least one auxiliary feature to form an enhanced feature vector; and inputting the enhanced feature vector into a pre-trained random forest classifier to obtain a final tailing behavior judgment result; and in response to the tailing behavior judgment result being tailing, executing a management strategy. The method of the present application can greatly improve the accuracy of tailing behavior recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing. More particularly, the present application relates to a non-sensing passage management method and system. BACKGROUND

[0002] Non-sensing passage systems aim to improve the efficiency and experience of personnel passing through, which usually capture image information of the passage area by sensors such as cameras, and confirm the passage authority through identity recognition technology (such as face recognition), and automatically control the gate to open. In actual application, especially in subway stations, office building lobbies and other peak hours, the flow of people is huge and dense, and it is easy to occur "tail" behavior of unauthorized personnel following authorized personnel to pass through illegally before the gate is closed.

[0003] To address this challenge, existing technologies usually use a classifier based on machine learning, such as a random forest model, to determine whether tailing occurs. Such a method first extracts a set of artificially designed features from sensor data, such as: the number of heads in the detection area, the pixel area of the human body contour, the distance between the centers of the human body contour, the average passing speed, etc. Then, input these features into a pre-trained random forest model, and output the classification result of "normal passage" or "tailing" by the model. When tailing occurs, the alarm device is controlled to alarm.

[0004] However, the above method based on random forest and artificially designed features has significant technical limitations. The core technical problem is that this method relies heavily on a set of fixed features designed based on prior knowledge. This "feature engineering" method lacks sufficient expressiveness and flexibility, and when faced with the ever-increasing and increasingly complex tailing strategies in the real world (for example, tailers use large luggage or heavy clothing to form a single contour with the person in front, two people walk closely together or adopt deceptive gait to imitate single person passing), there is a serious problem of insufficient recognition and generalization ability. Since the algorithm cannot automatically learn and discover deeper, more abstract features that reflect the inherent nature of "multiple entities" from raw data, it has a serious lack of generalization ability for complex tailing scenarios that have not appeared in training samples, resulting in low accuracy of tailing behavior recognition results, and high false negative rate and false positive rate of alarm devices. SUMMARY

[0005] To solve the technical problems of low accuracy of tailing behavior recognition results, high false negative rate and false positive rate of alarm devices in the existing non-sensing passage management method, the present application provides solutions in the following aspects.

[0006] In a first aspect, the present application provides a non-inductive access management method, comprising: collecting a depth image of an access area, and segmenting a foreground target from the depth image; for each pixel point in the foreground target region, calculating the depth fluctuation value in the neighborhood window of the point;

[0007] Screening the pixel points with the depth fluctuation value in the neighborhood window greater than the fluctuation value threshold from the pixel points in the foreground target region, and taking the screened pixel points as structure abnormal points; calculating the spatial dispersion of the structure abnormal points relative to the geometric centroid of the foreground target; determining the morphological abnormality factor of the foreground target, which is used to represent the deviation degree of the size of the foreground target relative to the standard size of the foreground target; the standard size of the foreground target refers to the size of the foreground target corresponding to the access channel when only one person passes through;

[0008] Taking the spatial dispersion as the basis score, and weighting the basis score by using a non-linear amplification function positively correlated with the morphological abnormality factor, to obtain a tailing potential score;

[0009] Combining the tailing potential score with at least one auxiliary feature to form an enhanced feature vector; and inputting the enhanced feature vector into a pre-trained random forest classifier to obtain a final tailing behavior judgment result; in response to the tailing behavior judgment result being tailing, executing a management strategy, the management strategy including: triggering an audible and light alarm, keeping the gate closed, and sending alarm information to a back-end security center.

[0010] The beneficial effects are: the present application fuses the internal structure dispersion and the external morphological abnormality factor to generate a high-discrimination "tailing potential score", and by constructing the tailing potential score, the feature engineering is upgraded from relying on simple geometric features designed by artificial to dynamic features combining the internal structure complexity of the target and the global morphology, which fundamentally improves the expression ability of the features. The method can effectively identify various complex tailing behaviors disguised by clothes, luggage, close posture, etc., greatly improves the generalization ability and recognition accuracy of unobserved tailing modes, greatly improves the accuracy of tailing behavior recognition, and thus makes the execution of the management strategy more accurate and robust. In addition, by deeply analyzing the internal depth structure of the foreground target, the present method can effectively distinguish between "a single user with a large body type or carrying large items" and "two independent individuals in close contact", avoiding the misjudgment caused by the traditional method due to the excessive area of the contour. This ensures safety while reducing unnecessary alarms, ensuring the access efficiency and experience of normal users. Furthermore, the method of the present application mainly relies on the 3D geometric structure and depth information of the target, and is not sensitive to environmental factors such as light changes, clothing colors, and shadows, and has stronger environmental adaptability and robustness compared to methods relying on RGB images.

[0011] Preferably, the depth fluctuation value in the neighborhood window of the pixel point (i, j) is calculated by the following expression:

[0012] ;

[0013] where D ldf (i, j) represents the depth fluctuation value in the neighborhood window of the pixel point (i, j), W ij is the neighborhood window of size N x N centered at the point (i, j), d(p, q) is the depth value of any point (p, q) in the neighborhood window, μ ij is the average depth value of all pixel points in the neighborhood window W ij .

[0014] Preferably, the spatial dispersion degree is calculated by the following expression:

[0015]

[0016] where I sad represents the spatial dispersion degree of the pixel points in the structural abnormal point set relative to the geometric centroid of the foreground target, p k is the pixel coordinate of the kth structural abnormal point, M represents the number of pixel points in the structural abnormal point set, is the pixel coordinate of the geometric centroid of the entire foreground target region.

[0017] Preferably, the morphological abnormality factor of the foreground target is calculated by the following expression:

[0018] ;

[0019] where F represents the morphological abnormality factor of the foreground target, represents the pixel area of the foreground target, represents the standard size of the foreground target.

[0020] Preferably, the trailing potential score is calculated by the following expression:

[0021]

[0022] where S represents the trailing potential score, represents the spatial dispersion degree of the pixel points in the structural abnormal point set relative to the geometric centroid of the foreground target, represents the morphological abnormality factor of the foreground target, is a positive weight coefficient, is a maximum function, represents the logarithm with base 10.

[0023] The beneficial effect is that the formula takes the spatial dispersion of the structural abnormal points relative to the geometric centroid of the foreground target as the basic score. When the target shape is normal ( ≤1), the trailing potential score is 0, indicating no trailing suspicion. When the target shape is beyond the standard ( >1), The greater the value of the log term, the greater the value of the trailing potential score, thereby producing an exponential amplification effect on This means that for an oversized target contour, even if only moderate structural dispersion is detected inside it, its final trailing potential score will be significantly raised. Therefore, the calculation expression of the trailing potential score of the present application can more accurately calculate the trailing potential score.

[0024] Preferably, the neighborhood window is a 5x5 window.

[0025] The beneficial effect is that if the neighborhood window size is too small, it is too sensitive to sensor noise, which may produce artifacts and cannot accurately characterize the real surface structure of the foreground target. If the neighborhood window size is too large, it will be over-smoothed, causing key subtle depth mutation information such as human body contact and occlusion to be blurred, thereby reducing the distinguishing degree of the feature. Therefore, setting the size of the neighborhood window to 5x5 can effectively suppress noise and retain key details.

[0026] Preferably, the determination method of the fluctuation value threshold comprises:

[0027] Collecting multiple depth images of single normal passing; the collected depth image data includes images of people of different heights, different body shapes and different clothing passing through the passing area;

[0028] Extracting the foreground target area of each depth image and calculating the depth fluctuation value in the neighborhood window of each pixel point in each foreground target area;

[0029] Taking the 95% quantile of the calculated depth fluctuation values of all pixel points as the fluctuation value threshold.

[0030] The beneficial effect is that: under normal circumstances, when a single person passes, most of the surface of the body is smooth (the depth fluctuation value of the foreground area pixel point of the corresponding depth image is small), when the clothes of the person are wrinkled, the edge of the backpack and the arm swing, it will cause the local depth fluctuation value to be larger, in addition, when two people body contact and other tailing behaviors occur, it will also cause the local depth fluctuation value to be larger, but the local depth fluctuation value corresponding to the tailing behavior will be much larger than the local depth fluctuation value caused by the normal wrinkles of the clothes, the edge of the backpack and the arm swing; by taking the 95% quantile of the depth fluctuation value of all pixel points as the fluctuation value threshold, a robust normal baseline is established, which can avoid misjudging these local depth slight fluctuations as abnormal, and can effectively filter out most of the low-amplitude fluctuations caused by normal wrinkles of clothes and the like, and only retain the significant structural abnormalities caused by limb interlacing, human contact and the like, thereby improving the signal-to-noise ratio of the calculated tailing potential score.

[0031] Preferably, the method for obtaining the standard size of the foreground target comprises:

[0032] Collect depth images of normal passing of people with different heights and different body shapes, thereby obtaining a plurality of normal passing depth images;

[0033] Extract the foreground target from each normal passing depth image respectively and calculate the pixel area of the foreground target;

[0034] Average the calculated pixel area of the foreground target, thereby obtaining the standard size of the foreground target.

[0035] Preferably, the auxiliary features include at least one of the average passing speed of the foreground target within a preset time length and the smoothness of the motion trajectory of the geometric centroid of the foreground target.

[0036] The beneficial effect is that: the passing speed and the trajectory smoothness belong to kinematic features, by introducing the passing speed and the trajectory smoothness, the system adds time dimension judgment basis. This can effectively distinguish between "two people walking side by side" and "a stationary wide object", because the former has clear passing speed and coherent trajectory, while the latter does not. Complementing these dynamic features with tailing potential scores can further exclude potential misjudgment scenarios, making the final tailing detection result more reliable.

[0037] In a second aspect, the present application provides an inconsiderate passing management system, comprising a processor and a memory, the memory stores computer program instructions, when the computer program instructions are executed by the processor, the inconsiderate passing management method of the present application is realized.

[0038] In summary, the present application has the advantages that the inapparent passage management method and system can greatly improve the accuracy of tailing behavior recognition, and thus make the execution of management strategy more accurate and robust. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a flow chart of an inapparent passage management method according to an embodiment of the present application;

[0040] Figure 2 is a structural schematic diagram of an inapparent passage management system according to an embodiment of the present application. DETAILED DESCRIPTION

[0041] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0042] Embodiment of inapparent passage management method:

[0043] As shown in Figure 1 the inapparent passage management method of the present application comprises:

[0044] S101, collecting a depth image of a passage area, and segmenting a foreground target from the depth image;

[0045] A 3D depth camera, such as a time-of-flight (ToF) camera or a structured light camera, can be deployed above or obliquely above the passage gate channel. The camera continuously collects a depth image sequence of the passage area at a top-down angle. Each pixel value in the depth image directly corresponds to the physical distance between the point in the scene and the camera.

[0046] In general, the background of the depth image is the ground, the gate wall and other objects.

[0047] In this embodiment, segmenting the foreground target from the depth image comprises: performing a pixel-by-pixel difference comparison between the current frame depth image and the background depth image, and identifying the pixel points with a difference exceeding a preset distance threshold as the foreground, thereby generating a foreground target containing only the passage personnel; the background depth image is a depth image not containing the passage personnel. In other embodiments, a neural network model can also be used to identify the foreground target and segment it from the depth image.

[0048] S102, for each pixel point in the foreground target area, calculating a depth fluctuation value in the neighborhood window of the point;

[0049] In this embodiment, the neighborhood window is an N x N window, where N is a small odd number; preferably, the neighborhood window in this embodiment uses a 5 x 5 window.

[0050] If N is too small (for example, 3), the sensor noise is too sensitive, which may produce artifacts and cannot stably reflect the real surface structure of the foreground target; if N is too large (for example, 9 or more), it will be excessively smoothed, resulting in the key subtle depth mutation information such as human body contact and occlusion being blurred, thereby reducing the distinguishing degree of the features. Therefore, setting N to 5 can achieve a good balance between effectively suppressing noise and retaining key details.

[0051] In the embodiment, for a pixel point (i, j), the calculation expression of the depth fluctuation value in the neighborhood window thereof is as follows:

[0052] ;

[0053] In the formula, D ldf (i, j) represents the depth fluctuation value in the neighborhood window of the pixel point (i, j), W ij is a neighborhood window of N x N size with the pixel point (i, j) as the center, d(p, q) is the depth value of any point (p, q) in the neighborhood window, μ ij is the average depth value of all pixel points in the neighborhood window W ij .

[0054] The formula is used to measure the local surface smoothness of each point in the foreground target. When the neighborhood depth values of a point change gently (for example, in the flat back area), the standard deviation, that is, the Dldf value, will be low; otherwise, when the neighborhood depth values change dramatically due to clothing wrinkles, interlaced limbs, or occlusion between human bodies, the Dldf value will be significantly increased.

[0055] The depth of a single continuous human body surface changes relatively smoothly and regularly in a local range; when two or more human bodies are close or partially overlapped, a large number of irregular and dramatic depth mutations will inevitably occur in the contact and occlusion areas. By calculating the depth fluctuation value in the neighborhood window of the pixel point in the foreground target region, the "roughness" of the internal structure can be quantified.

[0056] In other embodiments, the depth fluctuation value in the neighborhood window of the pixel point can also be calculated in other suitable manners.

[0057] In S103, the structure abnormal points are screened according to the depth fluctuation value, specifically, the pixel points with the depth fluctuation value greater than the fluctuation value threshold in the neighborhood window are screened from the pixel points in the foreground target region, and the screened pixel points are taken as the structure abnormal points.

[0058] If the depth fluctuation value of a certain pixel point neighborhood range is large, it indicates that the clothes of the passerby in the foreground target region have wrinkles or the probability of having a trailing person is large, and therefore this pixel point is determined as a structure abnormal point; if the depth change of a certain pixel point neighborhood range is relatively smooth and the change amplitude is small, it indicates that the clothes of the passerby in the foreground target region are flat and the probability of having a trailing person is large, and therefore this pixel point is determined as a structure normal point.

[0059] S104, calculating a spatial dispersion degree of the structure abnormal points relative to a geometric centroid of the foreground target;

[0060] For a single passerby, even if he carries a backpack or his clothes have wrinkles, the region with large depth fluctuation tends to be concentrated in a single position. However, for a trailing scene, there are likely to be multiple dispersed and unrelated regions with large depth fluctuation (for example, the contact position of the shoulders of two people, and the occlusion position of the head of the latter and the back of the former). This step can quantify the dispersion degree of the spatial distribution of the region with large depth fluctuation by calculating the spatial dispersion degree of the pixel points in the structure abnormal point set relative to the geometric centroid of the foreground target, and thus helps to determine whether a trailing phenomenon occurs.

[0061] In this embodiment, the calculation expression of the spatial dispersion degree of the structure abnormal points relative to the geometric centroid of the foreground target is:

[0062]

[0063] In the formula, I sad represents the spatial dispersion degree of the pixel points in the structure abnormal point set relative to the geometric centroid of the foreground target, p k is the pixel coordinate of the kth structure abnormal point, and M represents the number of pixel points in the structure abnormal point set, is the pixel coordinate of the geometric centroid of the entire foreground target region.

[0064] This formula calculates the root mean square of the distances of all structure abnormal points to the geometric centroid of the foreground region. When all abnormal points are closely distributed around the centroid (corresponding to a single passerby), the value of I sad is small. When the abnormal points are dispersed in multiple different positions of the foreground region (corresponding to a trailing person), the root mean square of their distances to the overall centroid increases, thereby causing the value of I sad to significantly increase. The higher the value of I sad , the more likely it is that the foreground target includes multiple passersby.

[0065] S105, obtaining a shape abnormality factor of the foreground target, specifically: determining the shape abnormality factor of the foreground target, the shape abnormality factor being used to represent the deviation degree of the size of the foreground target relative to the standard size of the foreground target; the standard size of the foreground target refers to the size of the foreground target when only one person passes through the passageway.

[0066] If the spatial dispersion of the structural abnormal points of the foreground target relative to the geometric centroid of the foreground target is large, but the area of the foreground target is small, it indicates that the case of large spatial dispersion of the structural abnormal points is likely caused by sensor noise; if the area of the foreground target is large and the internal dispersion is also high, it is extremely likely caused by the trailing phenomenon. By calculating the morphological abnormality factor of the foreground target, it is helpful to more accurately determine whether there is a trailing phenomenon.

[0067] In the embodiment, the expression for calculating the morphological abnormality factor of the foreground target is:

[0068] ;

[0069] In the formula, the morphological abnormality factor of the foreground target is represented by F, the pixel area of the foreground target is represented by A, and the standard size of the foreground target is represented by S.

[0070] S106, calculating the trailing potential score, specifically: taking the spatial dispersion as the base score, and using a non-linear amplification function positively correlated with the morphological abnormality factor to weight the base score, thereby obtaining the trailing potential score;

[0071] In the embodiment, the expression for calculating the trailing potential score is:

[0072] ;

[0073] In the formula, the trailing potential score is represented by T, the spatial dispersion of the pixel points in the set of structural abnormal points relative to the geometric centroid of the foreground target is represented by D, the morphological abnormality factor of the foreground target is represented by F, is a positive weight coefficient, is a maximum function, and the logarithm with base 10 is represented by log.

[0074] S107, obtaining an enhanced feature vector and combining a random forest classifier to obtain a trailing behavior judgment result, specifically: combining the trailing potential score with at least one auxiliary feature to form an enhanced feature vector; and inputting the enhanced feature vector into a pre-trained random forest classifier to obtain the final trailing behavior judgment result;

[0075] Since the trailing potential score already contains deep insights into complex trailing patterns, the random forest model can make more accurate and robust judgments based on this high-discrimination feature.

[0076] In the embodiment, the auxiliary features include at least one of an average passing speed of the foreground target in a preset time length and a smoothness of a motion trajectory of a geometric center of the foreground target.

[0077] In the embodiment, the preset time length can be 3s.

[0078] The random forest classifier is trained in the following manner: a large number of labeled passing data samples are collected, including various normal passing and complex tailing scenes, an enhanced feature vector is calculated for each sample, and the random forest classifier is trained. As an exemplary implementation, the random forest model can include 100 decision trees. To prevent overfitting, the maximum depth of each tree can be set to 10, and the minimum number of samples in the leaf node is required to be 5. The setting of these hyperparameters ensures that the model has good generalization performance while maintaining high classification ability.

[0079] S108, according to the judgment result, execute the management strategy, specifically: in response to the tailing behavior judgment result is tailing, then execute the management strategy, the management strategy includes: trigger sound and light alarm, keep the gate closed and send alarm information to the background security center.

[0080] In one embodiment, the method for determining the fluctuation value threshold comprises:

[0081] S201, collect depth images of a plurality of different persons passing through the gate alone in different situations to form a data set; the plurality of different persons include persons of different heights and different body shapes, and the situations include the clothing of the person, whether the person carries small items, and the type of small items carried by the person;

[0082] Because the clothing of the person in the depth image will cause the depth fluctuation value in the neighborhood window of the pixel points in the extracted foreground target region to change, in the embodiment, to improve the accuracy of the fluctuation value threshold calculation, the clothing types of the persons in the plurality of depth images in the generated data set should be sufficient, for example: T-shirt, thick coat, and wrinkled windbreaker. The depth fluctuation value corresponding to the extracted foreground target region when the person in the depth image carries small items will be different from the depth fluctuation value corresponding to the extracted foreground target region when the person does not carry small items, and the person carrying different small items in the depth image will also cause the depth fluctuation value corresponding to the extracted foreground target region to be different, therefore, the generated data set should include depth images of persons not carrying small items passing through the gate alone, and depth images of persons carrying small items passing through the gate alone. The depth images of persons carrying small items passing through the gate alone should cover multiple types, and the types of small items carried by the persons in the different types of depth images are different.

[0083] S202, extract the foreground target region of each depth image, and calculate the depth fluctuation value of each pixel point in the neighborhood window of each foreground target region;

[0084] S203, take the 95% quantile of the calculated depth fluctuation value of all pixel points as the fluctuation value threshold.

[0085] In one embodiment, the method for obtaining the standard size of the foreground target comprises:

[0086] S301, collect depth images of normal passers-by of different heights and different body shapes, thereby obtaining multiple normal pass depth images;

[0087] S302, extract foreground targets from each normal pass depth image respectively and calculate the pixel area of the foreground target;

[0088] S303, average the calculated pixel area of the foreground target, thereby obtaining the standard size of the foreground target.

[0089] In one embodiment, the method for calculating the smoothness of the motion trajectory of the geometric centroid of the foreground target comprises:

[0090] S401, continuously collect N frames of depth images of the pass region, and calculate the geometric centroid of the foreground target region for each frame of depth image; the centroid is a two-dimensional coordinate point;

[0091] S402, arrange the calculated geometric centroids in chronological order, thereby obtaining a trajectory sequence;

[0092] S403, for every two adjacent geometric centroids in the trajectory sequence, calculate the corresponding displacement vector;

[0093] S404, for every two consecutive displacement vectors, calculate the corresponding cosine of the included angle;

[0094] S405, average the calculated cosine of the included angle, and take it as the smoothness of the motion trajectory of the geometric centroid of the foreground target.

[0095] When the trajectory is smooth (close to a straight line), the direction of the continuous displacement vector is almost unchanged, and the corresponding included angle tends to 0 degrees, and its cosine of the included angle tends to 1.

[0096] When the trajectory is sharply turned or shaken, the corresponding included angle will become larger, and its cosine of the included angle will decrease, and even become negative (if it turns 180 degrees).

[0097] Therefore, the higher the average of the final cosine of the included angle (the closer to 1), the smoother the trajectory; the lower the average of the cosine of the included angle, the more "winding" or "shaking" the trajectory.

[0098] Non-susceptible pass management system embodiment

[0099] The present application also provides a non-susceptible pass management system. As shown in Figure 2 the non-susceptible pass management system comprises a processor and a memory, and the memory stores computer program instructions which, when executed by the processor, implement the non-susceptible pass management method according to the first aspect of the present application.

[0100] The non-susceptible pass management system also comprises a communication bus and a communication interface and other components well known to those skilled in the art, the settings and functions of which are known in the art, and thus will not be described here.

[0101] In the description of the present specification, the meaning of "a plurality of", "several" is at least two, for example, two, three or more, etc., unless otherwise explicitly specifically limited.

[0102] Although the present specification has shown and described several embodiments of the present application, it will be apparent to those skilled in the art that such embodiments are provided only by way of example. Those skilled in the art will think of many changes, changes and alternatives without departing from the idea and spirit of the present application. It should be understood that various alternatives to the embodiments of the present application described herein can be employed in practicing the present application.

Claims

1. A non-intrusive passage management method characterized by comprising: The method comprises: collecting a depth image of a passing area and segmenting a foreground object from the depth image; for each pixel point in the foreground object region, calculating a depth fluctuation value in a neighborhood window of the point; selecting, from the pixel points in the foreground object region, a pixel point whose depth fluctuation value in the neighborhood window is greater than a fluctuation value threshold, and taking the selected pixel point as a structural abnormal point; calculating a spatial dispersion degree of the structural abnormal point relative to a geometric centroid of the foreground object; determining a morphological abnormality factor of the foreground object, the morphological abnormality factor being used to represent a deviation degree of a size of the foreground object relative to a standard size of the foreground object; the standard size of the foreground object refers to a size of the foreground object corresponding to a case where only one person passes through the passing channel; taking the spatial dispersion degree as a basic score, and weighting the basic score by using a nonlinear amplification function positively correlated with the morphological abnormality factor, so as to obtain a tailing potential score, and the expression is: In the formula, represents the trailing potential score, represents the spatial dispersion degree of the pixel points in the structural anomaly point set relative to the geometric centroid of the foreground target, represents the morphological anomaly factor of the foreground target, is a positive weight coefficient, is a maximum value function, represents the logarithm with base 10; combining the tailing potential score with at least one auxiliary feature to form an enhanced feature vector; and inputting the enhanced feature vector into a pre-trained random forest classifier to obtain a final tailing behavior judgment result; in response to the tailing behavior judgment result being tailing, executing a management strategy, the management strategy including triggering an audible and light alarm, keeping the gate closed, and sending alarm information to a background security center.

2. The inductive pass management method of claim 1, wherein For a pixel point (i, j), the depth fluctuation value in the neighborhood window of the pixel point (i, j) is calculated according to the following expression: ； where D ldf (i,j) represents the depth fluctuation value in the neighborhood window of pixel point (i,j), W ij is the neighborhood window of size N x N centered at point (i,j), d(p,q) is the depth value of any point (p,q) in the neighborhood window, μ ij is the average depth value of all pixel points in the neighborhood window W ij .

3. The inductive pass management method of claim 1, wherein The expression for calculating the spatial dispersion degree is: In the formula, I sad represents the spatial dispersion degree of the pixel points in the structural anomaly point set relative to the geometric centroid of the foreground target, p k is the pixel coordinate of the kth structural anomaly point, and M represents the number of pixel points in the structural anomaly point set, is the pixel coordinate of the geometric centroid of the entire foreground target region.

4. The inductive pass management method of claim 1, wherein, The expression for calculating the morphological abnormality factor of the foreground object is: ； wherein a shape abnormality factor indicative of a shape of the foreground object, a pixel area indicative of a pixel area of the foreground object, a standard size indicative of a standard size of the foreground object.

5. The inductive pass management method of claim 1, wherein, The neighborhood window is a window with a size of 5x5.

6. The inductive pass management method of claim 1, wherein, The determination method of the fluctuation value threshold comprises: collecting multiple depth images of normal passing of a single person; the collected depth image data includes images of people with different heights, different body shapes and different clothes passing through the passing area; extracting the foreground object region of each depth image and calculating the depth fluctuation value in the neighborhood window of each pixel point in each foreground object region; taking the 95% quantile of the calculated depth fluctuation values of all pixel points as the fluctuation value threshold.

7. The inductive pass management method of claim 1, wherein, The acquisition method of the standard size of the foreground object comprises: collecting depth images of normal passing of people with different heights and different body shapes, thereby obtaining multiple normal passing depth images; extracting the foreground object from each normal passing depth image and calculating the pixel area of the foreground object; averaging the calculated pixel areas of the foreground objects, thereby obtaining the standard size of the foreground object.

8. The non-intrusive tolling method of any one of claims 1-7, wherein, The auxiliary features include at least one of an average passing speed of the foreground object within a preset time length and a motion trajectory smoothness of the geometric centroid of the foreground object.

9. A non-intrusive access management system comprising a processor and a memory, the memory storing computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the non-susceptible passing management method of claim 8.

Images