A method and system for detecting passenger invasion behavior in an urban rail transit station

Abstract

The application discloses a kind of urban rail transit station passenger intrusion behavior detection method and system, detection method is based on urban rail transit station monitoring video data, first, the feature points of moving passenger in current frame image are extracted in the adjacent frame images of urban rail transit station monitoring video using pyramid layered LK optical flow algorithm;Then, the density index of the feature points of moving passenger is calculated using elliptical domain subtraction clustering algorithm, and the clustering of feature points is carried out accordingly, to obtain the position information of moving passenger;Afterwards, the position information of moving passenger in current frame image is input into target tracking algorithm, to predict the position of passenger in future time;Finally, according to the early warning area actually delimited in station, the intrusion situation in early warning area is detected in real time using virtual wall detection technology.The detection method of the application is accurate and reliable and has high real-time performance, and can provide a more reliable and safe environment for subsequent urban rail transit station intelligent operation and maintenance.

Description

Technical Field

[0001] This invention belongs to the field of intelligent visual recognition technology, and in particular relates to a method and system for detecting passenger intrusion behavior in urban rail transit stations. Background Technology

[0002] With the network development of urban rail transit, the passenger flow of urban rail transit is showing a continuous upward trend. The surge in passenger flow has brought many safety challenges to the operation and management of urban rail transit. Subway stations are an important part of urban transportation lines. Subway stations have large passenger flows and complex personnel, making them important nodes for ensuring the safety of urban residents' travel. In urban rail transit travel, due to the subjective and random nature of passenger movement, urban rail stations are not completely open public areas. There are designated areas for station equipment, prohibited areas, and dangerous areas, requiring monitoring of passenger distribution and behavior in key areas of the station. Currently, urban rail stations are designed and equipped with a large number of closed-circuit television (CCTV) monitoring systems, with a focus on key areas such as turnstiles, machine rooms, escalators, and platforms. However, manual video monitoring capabilities are limited, and it is not easy to observe and monitor the surveillance video in real time for extended periods. To address the aforementioned issues, computer processing of surveillance video data can pinpoint passenger locations in key areas and enable real-time passenger tracking within those areas. This allows for real-time monitoring of passenger intrusion behavior and prompt alarms, enabling the construction of an intelligent video surveillance system for urban rail stations. Combined with real-time data analysis, precise control measures can be implemented to resolve the problem of unauthorized passenger intrusion. Therefore, researching intelligent video surveillance systems for stations not only effectively enhances the safety capabilities of urban rail transit but also provides a more reliable and secure environment for intelligent operation and maintenance, contributing to the sustainable development of urban rail transit.

[0003] Existing research on passenger intrusion detection in urban rail transit stations mainly focuses on machine learning and deep learning methods. Yang Yun (Yang Yun, Wei Qiushi, Hu Shoujian, et al. Research and application practice of intelligent station video analysis system[J]. Electronic Technology and Software Engineering, 2023(05):167-170.) et al. analyzed the action parameters of station passengers based on the results of OpenPose human morphology recognition, and combined machine learning to construct a passenger tracking algorithm based on K-means clustering to achieve accurate identification of passenger intrusion behavior in stations; Zhang Yuxiang (Zhang Yuxiang, Li Xianwang, He Deqiang, et al. Passenger behavior recognition in subway stations based on improved multi-algorithm fusion[J]. Journal of Railway Science and Engineering, 2023,20(11):4096-4106.) et al. used the YOLOv5_m target detection algorithm to frame passengers, then used the SPPE single-person pose estimation algorithm to extract the skeletal key points of the framed passengers, and finally used the 2s-STEGCN algorithm to analyze the coordinates of the passenger skeletal key points to complete the identification of passenger intrusion behavior. However, the above studies require a dataset of passenger intrusion behavior in urban rail transit station scenarios to train the model, and it is difficult to explain the detection principles and decision-making processes of the model for passenger behavior, limiting the application of these algorithms in real-world scenarios. Furthermore, current research on using station CCTV system data to perceive passenger intrusion behavior is insufficient. With the maturity and development of machine vision technology, perceiving passenger intrusion behavior within urban rail transit stations based on surveillance videos has become possible. Simultaneously, as the requirements for station safety in urban rail transit continue to increase, real-time and accurate monitoring of passenger behavior within stations has become an urgent problem to be solved. Summary of the Invention

[0004] In view of the shortcomings pointed out in the background art above, the present invention provides a method and system for detecting passenger intrusion behavior in urban rail transit stations, aiming to solve the problems existing in the prior art mentioned above.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for detecting passenger intrusion behavior in urban rail transit stations, the method comprising the following steps:

[0007] S1. Using the pyramid-layered LK optical flow algorithm, optical flow tracking is performed in adjacent frames of urban rail transit station monitoring video to obtain optical flow values. The multi-layer optical flow tracking results are the locations of feature points. After preprocessing, the feature points are scattered in the area of ​​moving passengers in the current frame image, thus the feature points of moving passengers in the current frame image are extracted.

[0008] S2. Calculate the density index of feature points of moving passengers using the elliptic domain subtractive clustering algorithm, and perform feature point clustering based on the density index of the feature points to obtain the location information of moving passengers.

[0009] S3. Input the position information of the moving passenger in the current frame image into the Mean Shift target tracking algorithm, calculate the Mean Shift vector using pixel weights, update the position information of the target passenger based on the calculated Mean Shift vector, and predict the position information of the passenger in the next frame image, that is, predict the position of the passenger in the future.

[0010] S4. Based on the actual warning area defined by the urban rail transit station, a virtual wall area is set within the monitoring range. Virtual wall detection technology is used to determine the relationship between the passenger's position and the virtual wall at future times, thereby detecting intrusions within the warning area in real time.

[0011] Preferably, in step S1, the extraction process of the feature points of the moving passenger specifically includes the following steps:

[0012] S11. Use the GFTT feature detection algorithm to detect key feature points from the previous frame of the station monitoring video.

[0013] S12. Using image pyramid technology, the previous frame and the current frame of the surveillance video are scaled down from high resolution to low resolution to obtain an image pyramid.

[0014] S13. Perform optical flow tracing on the lower resolution image layer of the image pyramid to obtain the initial optical flow value, pass the optical flow value down to the higher resolution image layer, and perform a consistent optical flow tracing strategy. The result of multi-layer optical flow tracing is the position of the feature point.

[0015] S14. After preprocessing, the feature points are scattered across the area of ​​the moving passenger in the current frame image.

[0016] Preferably, the scaling factors of the image pyramid from top to bottom are 1 / 8, 1 / 4, 1 / 2, and 1.

[0017] Preferably, in step S13, the optical flow tracking steps are as follows:

[0018] S131. Initialize the displacement vector of the feature point to zero;

[0019] S132. Calculate the error and Jacobian matrix based on the pixel values ​​corresponding to adjacent frame images;

[0020] S133. Calculate the Hessian matrix and deviation vector based on the error and Jacobian matrix;

[0021] S134. Calculate the displacement vector based on the Hessian matrix and the deviation vector;

[0022] S135. Superimpose the calculated displacement vector onto the current displacement vector of the feature point to complete the tracking of the feature point.

[0023] Preferably, in step S14, the preprocessing method is as follows: the corresponding feature points between the previous frame image and the current frame image are tracked through the image pyramid, the corresponding feature points are used as the starting and ending points of the pixel displacement, the current displacement vector of the pixel is drawn, and a threshold condition for the pixel displacement length is set so that the feature points are scattered on the passenger area in the current frame image.

[0024] Preferably, in step S2, the feature point clustering includes the following steps:

[0025] S21. Define the density calculation index for elliptic domains;

[0026] S22. Calculate the density index of passenger feature points, find the feature point with the largest density index as the first cluster center, that is, obtain the position information of the first moving passenger in the current frame image;

[0027] S23. Remove the previous cluster center and use the maximum density index to correct the density index of other feature points;

[0028] S24. Find the feature point with the largest density index after correction as the cluster center, that is, obtain the location information of the next moving passenger in the current frame image;

[0029] S25. Repeat steps S23 and S24 until all valid moving passenger location information is obtained, then exit the iteration.

[0030] Preferably, in step S4, the virtual wall detection technology determines the positional relationship between the passenger's position and the virtual wall at a future time by including the following steps:

[0031] S41. Based on the actual delineation of the station's early warning area, set up virtual wall areas that are regular rectangles or irregular polygons;

[0032] S42. Obtain the vertex coordinates of the virtual wall, calculate the equation of the line on which the virtual wall is located, determine the position of the passenger and the position of the virtual wall through the geometric relationship between the point and the line, and obtain the passenger intrusion situation in the warning area.

[0033] The present invention further provides a detection system for passenger intrusion behavior in urban rail transit stations, comprising:

[0034] Optical flow feature extraction module: It has a built-in pyramid layered LK optical flow algorithm to perform optical flow tracking on the received urban rail transit station monitoring video images to extract feature points of moving passengers in the images, and send the feature point information to the moving passenger positioning module;

[0035] Moving passenger positioning module: It has an elliptic domain subtractive clustering algorithm, receives the feature point information of moving passengers sent by the optical flow feature extraction module, and obtains the location information of moving passengers by clustering the feature points of moving passengers through the elliptic domain subtractive clustering algorithm, and sends the location information of moving passengers to the station passenger tracking module.

[0036] Station passenger tracking module: It has a built-in Mean Shift target tracking algorithm, receives the location information of moving passengers sent by the moving passenger positioning module, updates the location information of moving passengers through the Mean Shift target tracking algorithm, thereby predicting the passenger's location at a future time, and sends the predicted future passenger's location to the passenger intrusion detection module;

[0037] Passenger Intrusion Detection Module: Equipped with virtual wall detection technology, it determines the positional relationship between the received future passenger location and the virtual wall, concludes whether the passenger is within the virtual wall area, thereby obtaining passenger intrusion information within the warning area and issuing corresponding warning signals.

[0038] Compared with the shortcomings and deficiencies of existing technologies, the present invention has the following beneficial effects:

[0039] This invention uses sparse optical flow algorithm, subtractive clustering algorithm and target tracking algorithm to obtain the location information of moving passengers in urban rail transit stations in real time. The information is accurate, reliable and has high real-time performance. Based on the location information of moving passengers, virtual wall detection technology can be used to detect passenger intrusion in the warning area in real time, thereby effectively improving the security capabilities of smart urban rail stations. Attached Figure Description

[0040] Figure 1 This is a flowchart of a method for detecting passenger intrusion behavior in urban rail transit stations provided in an embodiment of the present invention.

[0041] Figure 2 This is a flowchart of the LK optical flow algorithm provided in an embodiment of the present invention.

[0042] Figure 3 This is a flowchart of the subtractive clustering algorithm provided in an embodiment of the present invention.

[0043] Figure 4 This is a flowchart of the Mean Shift target tracking algorithm provided in an embodiment of the present invention.

[0044] Figure 5 This is a schematic diagram illustrating the positional relationship between the passenger's location and the virtual wall area provided in an embodiment of the present invention.

[0045] Figure 6 This is a schematic diagram of the detection system provided in an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0047] A method for detecting passenger intrusion behavior in urban rail transit stations, flowchart as follows: Figure 1 As shown, the steps are as follows:

[0048] S1. Using the pyramid-layered LK (Lucas-Kanade) optical flow algorithm, optical flow tracking is performed in adjacent frames of urban rail transit station monitoring video to obtain optical flow values. The results of multi-layer optical flow tracking are the locations of feature points. After preprocessing, the feature points are scattered in the area of ​​moving passengers in the current frame image, thus extracting the feature points of moving passengers in the current frame image.

[0049] In this embodiment of the invention, the LK sparse optical flow algorithm, based on the assumptions of constant brightness, small motion, and spatial consistency, estimates pixel motion, i.e., optical flow, by utilizing the grayscale differences between adjacent frames. Unlike dense optical flow algorithms, this method is less prone to getting trapped in local extrema and solves the problem of optical flow calculation for large displacement motions. The pyramid-layered LK optical flow algorithm uses image pyramid technology to scale adjacent frames to a lower resolution, transforming large displacement motions into small displacement motions within the lower-resolution image. It then uses the Gauss-Newton iterative method to optimize the error function, achieving accurate pixel motion estimation in the image.

[0050] In this embodiment of the invention, the process of extracting feature points of a moving passenger is as follows:

[0051] First, the GFTT feature point detection algorithm is used to extract corner points from the image at time t, resulting in P feature corner points p. i (i = 1, 2, ..., P). Key feature points are detected from the previous frame of the station surveillance video using the GFTT (Good Features to Track) feature detection algorithm. The GFTT algorithm is a feature corner extraction method based on the autocorrelation matrix response value, which is faster than ORB and SIFT feature detection algorithms and more suitable for scenarios with high real-time requirements. After obtaining the feature points, the LK optical flow algorithm is used to track these feature points in conjunction with the current frame image. The implementation process of the LK optical flow algorithm is as follows: Figure 2 As shown, the image pyramid technique is used to scale the images at time t and time t+1 of the surveillance video to a lower resolution, constructing a four-layer image pyramid with scaling factors of 1 / 8, 1 / 4, 1 / 2, and 1 from top to bottom. Optical flow tracking is performed on the lower resolution image layer of the image pyramid to obtain the initial optical flow value. The optical flow value is then passed down to the higher resolution image layer, and a consistent optical flow tracking strategy is executed. The result of multi-layer optical flow tracking is the location of the feature point.

[0052] The steps of optical flow tracing are as follows:

[0053] Reference Figure 2 First, initialize the displacement vectors dx and dy of the feature points to zero; then, solve the optical flow using the Gauss-Newton iterative method. Observe one of the feature corner points, denoted as (x, y), and denote the gray level of the corresponding pixel as I. t (x,y), using Gauss-Newton's method, the solution of optical flow can be viewed as the following optimization problem:

[0054]

[0055] In the formula: Δx and Δy are the horizontal and vertical increments of the optical flow field, respectively; I t+1 (x+Δx,y+Δy) represents the gray level of the pixel corresponding to the corner point at time t+1.

[0056] In the Gauss-Newton method, the second derivative information is used to accelerate convergence to a local optimum of the error function. In each iteration, the difference between corresponding pixel values ​​in adjacent frames is calculated, also known as the error. In optical flow estimation, the variables of the error function are Δx and Δy, and the formula for the error function is:

[0057] E(Δx,Δy)=I t (x,y)-I t+1 (x+Δx,y+Δy) (2)

[0058] First, calculate the first derivative of the error, which is the Jacobian matrix of the error function. The formula is:

[0059]

[0060] The Jacobian matrix is ​​solved by the difference approximation, and its elements are the partial derivatives of the image pixel values ​​with respect to the displacement vector. In the LK optical flow algorithm, it is often assumed that within a feature window of size w×w (w=9 in this paper) centered on the feature corner point, w... 2 The motion of each pixel is similar. The feature window provides a better understanding of grayscale changes in the layout area, improving the accuracy of optical flow estimation. Therefore, the error gradient of each pixel within the feature window is calculated according to equations (2) and (3), denoted as J. i(i = 1, 2, ..., w) 2 ).

[0061] Then, calculate the Hessian matrix and deviation vector based on the error and Jacobian matrix, and calculate the second derivative of the error, i.e., the Hessian matrix of the error function, using the following formula:

[0062]

[0063] Where the Hessian matrix is ​​the sum of the products of the Jacobian matrices, and the deviation vector is the negative gradient of the error; by inversely solving the displacement vector using the relation, there exists a relationship: the product of the Hessian matrix and the displacement vector equals the deviation vector; under the premise of minimizing the error function, the increment of the error function variable is calculated using the Hessian matrix, and the increment equation is:

[0064] Δ=H -1 b (5)

[0065] In the formula: b is the negative gradient of the error function, expressed as: Δ contains the increments of variables Δx and Δy.

[0066] The above process only performs optical flow tracking on a single feature corner point in the original image at time t. Traditional optical flow algorithms have certain limitations when dealing with large pixel displacements. The LK optical flow method uses a multi-scale strategy to construct an image pyramid, transforming large pixel displacements into small displacements in a low-resolution image. This paper constructs a four-layer image pyramid with image scaling factors from top to bottom of 1 / 8, 1 / 4, 1 / 2, and 1. In each image layer, the Gauss-Newton method is used to calculate the feature corner point p. i The displacement increment is calculated. In this process, the output increment of the low-resolution image layer is always used as the input increment of the high-resolution image layer, processed layer by layer from coarse to fine, fully utilizing the feature corner information at different scales to improve the estimation accuracy of the feature corner motion. Based on the final increment output by the image pyramid, the positions of the feature corners in the image at time t are updated, thus obtaining the estimated positions of the feature corners in the image at time t+1.

[0067] By tracking the corresponding feature points between the previous and current frames using an image pyramid, and using these corresponding feature points as the starting and ending points of the pixel displacement, the current displacement vector of the pixel is drawn. A threshold condition for the pixel displacement length is set so that the feature points are scattered in the passenger area of ​​the current frame image.

[0068] Then, a threshold filtering method is used to remove feature corner points in the station background. The filtering formula for moving passenger feature points is as follows:

[0069]

[0070] In the formula: ζ is a constant, and the parameter value is set according to the actual situation.

[0071] Finally, the calculated displacement vector is superimposed on the current displacement vector of the feature point, and the displacement vectors dx and dy are updated to obtain the feature point x of the moving passenger in the adjacent frame image. i (i = 1, 2, ..., N), these feature points contain passenger location information and have good clustering properties, which can be used to locate moving passengers in the image.

[0072] S2. Use elliptic domain subtractive clustering algorithm to locate moving passengers in the station. The implementation process of subtractive clustering algorithm is as follows: Figure 3 As shown, this algorithm belongs to density-based clustering methods. Compared with distance-based clustering methods, it can find non-convex clusters and has certain noise resistance characteristics. Elliptic domain subtraction clustering algorithm is used to cluster the feature points of moving passengers to obtain their location information. The specific process of clustering feature points is as follows: the moving passenger feature points x are obtained through optical flow algorithm... i All are two-dimensional data points, denoted as (x ix ,x iy First, define the density calculation index of the elliptical domain; calculate the density index of each feature point. Considering the ratio of passenger height to width, it is more reasonable to represent the target area to be located using an ellipse. Therefore, define feature point x. i The formula for calculating the density index is:

[0073]

[0074] In the formula: γ is a constant adjusted according to the actual problem, which affects the decay trend of the exponential function curve, and is usually set to 4; r ax and r ay The constant represents the characteristic point x. i The semi-minor axis and semi-major axis in the neighborhood of the ellipse, and r ax =η·r ay Here, η is taken as 0.4, r ay The calculation formula is:

[0075]

[0076] In the formula: ||x i -x j || 2 It represents the Euclidean distance between any two feature points.

[0077] Matrix operations are used to obtain the density values ​​of all feature points; the maximum element in the density matrix is ​​then found, and the feature point with the current maximum density value is selected as the first cluster center. Maximum density is marked as This involves locating the first moving passenger in the current frame image; then, using the previous maximum density value to correct the density values ​​of other feature points, determining whether the corrected density values ​​meet the minimum density value condition, and repeating this process until all valid moving passengers are located. The calculation formula is as follows:

[0078]

[0079] In the formula: k+1 is the iteration number; M k+1 (x i M is the density index of feature point xi in the (k+1)th iteration; k (x i ), These are the density indices of feature points and cluster centers in the k-th iteration, respectively. r represents the x and y coordinates of the cluster centers in the k-th iteration; bx ,r by Since is a constant, the two define an elliptical domain where the density index significantly decays, and its calculation formula is:

[0080]

[0081] In the formula: λ is a constant, with a value between 1.25 and 1.5.

[0082] The maximum element in the density matrix is ​​queried again (the maximum element here is the corrected one, i.e., the second largest element), and the corresponding feature point is selected as the cluster center, i.e., the next moving passenger in the current frame image; Equation (9) is repeated for iteration, and the iteration ends under the following condition:

[0083]

[0084] In the formula: ε is a constant, and its value is between 0 and 1.

[0085] After the above iterative process, K cluster centers are obtained. The passenger's location area is calculated based on the semi-major axis and semi-minor axis of each cluster center and its elliptical neighborhood. The formula for calculating the radius of the location area is as follows:

[0086]

[0087] In the formula: r Dx ,r Dy To locate the semi-minor and semi-major axes of the elliptic domain for passenger D, the formula for locating the elliptic domain is:

[0088]

[0089] Finally, the K elliptic domain equations are used to mark the image coordinate system, and the final passenger positioning result is output to obtain the location information of the moving passengers at the station.

[0090] S3. Input the position information of the moving passengers in the current frame image into the Mean Shift target tracking algorithm to predict the passenger's position at future moments. Step S3 uses the classic Mean Shift algorithm in the field of target tracking to track moving passengers in the station surveillance video. The implementation process of the Mean Shift algorithm is as follows: Figure 4 As shown, the location regions of passengers at the station are first obtained. The target model is obtained by calculating the color histogram of the target region in the current frame. The probability density q of the passenger location ellipse region in the current frame image is then calculated. u (u = 1, 2, ..., m), where m is the number of target feature values, calculated using the following formula:

[0091]

[0092] In the formula: c is the normalization coefficient; K is the Epinechikov kernel function; x i For passenger feature points; the δ function is a Boolean function, when x i Within the target region, the function value is 1; otherwise, it is 0. Then, an image patch centered around the previous passenger region is extracted from the next frame, and its color histogram is calculated as a candidate model. In subsequent frames, based on the target region in the current frame, a prediction region is progressively searched. The probability density of the prediction region is calculated using the following formula:

[0093]

[0094] In the formula: y is the center coordinate of the prediction region; h is the width of the prediction region.

[0095] The Bhattacharyya coefficient is then used as a similarity function to measure the similarity between the target model and the candidate model. The larger the coefficient value, the higher the similarity. The formula for calculating the similarity function is:

[0096]

[0097] And calculate the weight value of each pixel in the current prediction region, using the following formula:

[0098]

[0099] Weights are calculated for each pixel in the candidate region based on similarity. A Mean Shift vector is then computed using these pixel weights, indicating the candidate model's next iteration position in the color space. Procedural matching is performed to find the maximum value of the similarity function, and the formula for calculating the new target position is:

[0100]

[0101] In the formula: the g function is the negative Epanechikov kernel function, i.e., g(x) = -K(x).

[0102] The neighborhood of the target region's center in the current frame image is selected as the predicted region's center. The process iteratively searches for the predicted region that maximizes the similarity function. The iteration terminates when:

[0103] ||y * -y||<ξ (19)

[0104] In the formula: ξ is a constant, and its value is between 0.1 and 1.

[0105] Finally, the target passenger's location information is updated based on the calculated Mean Shift vector to predict the passenger's location information in the next frame image, i.e., the passenger's location in the future.

[0106] S4. Based on the actual warning area defined by the urban rail transit station, a virtual wall area is set within the monitoring range. Virtual wall detection technology is used to determine the relationship between the passenger's position and the virtual wall at future times, thereby detecting intrusions within the warning area in real time.

[0107] In step S4, regarding the application of virtual wall detection technology, the main steps are setting the virtual wall area and determining the target location. For example... Figure 5 As shown, firstly, a virtual wall area DEFG is set in the monitoring field OABC according to the actual situation of the station's early warning area; a rectangular coordinate system XOY is established with the upper left corner of the monitoring video as the origin; the equations of the four virtual lines L1, L2, L3, and L4 of the virtual wall are calculated based on the coordinates of the four vertices of the virtual wall area and the two-point equation (y2-y1)(x-x1)-(x2-x1)(y-y1)=0; the midpoint P of the lower border of the passenger positioning area is selected as the target point, with coordinates (x0, y0). The positional relationship between the target point P and the four sides of the virtual wall is represented as LP1, LP2, LP3, and LP4, respectively, as follows:

[0108]

[0109] In the formula: LP = -1 indicates that the target point is to the left or above the virtual line, LP = 0 indicates that the target point is on the virtual line, and LP = 1 indicates that the target point is to the right or below the virtual line. Based on geometric relationships, the positional relationship between the target point P and the virtual wall can be determined by the following formula:

[0110]

[0111] In the formula: Status = 0 indicates that the target is to the left of virtual line L1, below L2, to the right of L3, or above L4, and the target is outside the virtual wall; Status = 1 indicates that the target is to the right of virtual line L1, above L2, to the left of L3, or below L4, and the target is inside the virtual wall, thus determining the passenger intrusion situation within the warning area.

[0112] Based on the above detection method, this invention provides a detection system for passenger intrusion behavior in urban rail transit stations, including an optical flow feature extraction module, a moving passenger localization module, a station passenger tracking module, and a passenger intrusion detection module, such as... Figure 6 As shown, the optical flow feature extraction module incorporates a pyramid-layered LK optical flow algorithm to perform optical flow tracking on received urban rail transit station monitoring video images to extract feature points of moving passengers in the images, and sends the feature point information to the moving passenger positioning module. The moving passenger positioning module incorporates an elliptic domain subtraction clustering algorithm to receive the feature point information of moving passengers sent by the optical flow feature extraction module, clusters the feature points of the moving passengers using the elliptic domain subtraction clustering algorithm to obtain the moving passenger's location information, and sends the moving passenger's location information to the station passenger tracking module. The station passenger tracking module incorporates a Mean Shift target tracking algorithm to receive the moving passenger's location information sent by the moving passenger positioning module, updates the moving passenger's location information using the Mean Shift target tracking algorithm to predict the passenger's location at a future time, and sends the predicted future passenger's location to the passenger intrusion detection module. The passenger intrusion detection module incorporates virtual wall detection technology to determine the positional relationship between the received future passenger's location and the virtual wall, concluding whether the passenger is within the virtual wall area, thereby obtaining the passenger intrusion situation within the warning area and issuing a corresponding warning signal.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting passenger intrusion behavior in urban rail transit stations, characterized in that, Includes the following steps: S1. Using the pyramid-layered LK optical flow algorithm, optical flow tracking is performed in adjacent frames of urban rail transit station monitoring video to obtain optical flow values. The multi-layer optical flow tracking results are the locations of feature points. After preprocessing, the feature points are scattered in the area of ​​moving passengers in the current frame image, thus the feature points of moving passengers in the current frame image are extracted. S2. Calculate the density index of feature points of moving passengers using the elliptic domain subtractive clustering algorithm, and perform feature point clustering based on the density index of the feature points to obtain the location information of moving passengers. The feature point clustering includes the following steps: S21. Define the density calculation index for elliptic domains; S22. Calculate the density index of passenger feature points, find the feature point with the largest density index as the first cluster center, that is, obtain the position information of the first moving passenger in the current frame image; S23. Remove the previous cluster center and use the maximum density index to correct the density index of other feature points; S24. Find the feature point with the largest density index after correction as the cluster center, that is, obtain the location information of the next moving passenger in the current frame image; S25. Repeat steps S23 and S24 until all valid moving passenger location information is obtained, then exit the iteration; S3. Input the position information of the moving passenger in the current frame image into the Mean Shift target tracking algorithm, calculate the Mean Shift vector using pixel weights, update the position information of the target passenger based on the calculated Mean Shift vector, and predict the position information of the passenger in the next frame image, that is, predict the position of the passenger in the future. S4. Based on the actual warning area defined by the urban rail transit station, a virtual wall area is set within the monitoring range. Virtual wall detection technology is used to determine the relationship between the passenger's position and the virtual wall at future times, thereby detecting intrusions within the warning area in real time.

2. The method for detecting passenger intrusion behavior in urban rail transit stations as described in claim 1, characterized in that, In step S1, the process of extracting the feature points of the moving passenger includes the following steps: S11. Use the GFTT feature detection algorithm to detect key feature points from the previous frame of the station monitoring video. S12. Using image pyramid technology, the previous frame and the current frame of the surveillance video are scaled down from high resolution to low resolution to obtain an image pyramid. S13. Perform optical flow tracing on the lower resolution image layer of the image pyramid to obtain the initial optical flow value, pass the optical flow value down to the higher resolution image layer, and perform a consistent optical flow tracing strategy. The result of multi-layer optical flow tracing is the position of the feature point. S14. After preprocessing, the feature points are scattered across the area of ​​the moving passenger in the current frame image.

3. The method for detecting passenger intrusion behavior in urban rail transit stations as described in claim 2, characterized in that, The scaling factors of the image pyramid from top to bottom are 1 / 8, 1 / 4, 1 / 2, and 1, respectively.

4. The method for detecting passenger intrusion behavior in urban rail transit stations as described in claim 2, characterized in that, In step S13, the optical flow tracking steps are as follows: S131. Initialize the displacement vector of the feature point to zero; S132. Calculate the error and Jacobian matrix based on the pixel values ​​corresponding to adjacent frame images; S133. Calculate the Hessian matrix and deviation vector based on the error and Jacobian matrix; S134. Calculate the displacement vector based on the Hessian matrix and the deviation vector; S135. Superimpose the calculated displacement vector onto the current displacement vector of the feature point to complete the tracking of the feature point.

5. The method for detecting passenger intrusion behavior in urban rail transit stations as described in claim 2, characterized in that, In step S14, the preprocessing method is as follows: the corresponding feature points between the previous frame image and the current frame image are tracked through the image pyramid, the corresponding feature points are used as the starting and ending points of the pixel displacement, the current displacement vector of the pixel is drawn, and the threshold condition of the pixel displacement length is set so that the feature points are scattered on the passenger area in the current frame image.

6. The method for detecting passenger intrusion behavior in urban rail transit stations as described in claim 1, characterized in that, In step S4, the virtual wall detection technology determines the positional relationship between the passenger's position and the virtual wall at a future time by including the following steps: S41. Based on the actual delineation of the station's early warning area, set up virtual wall areas that are regular rectangles or irregular polygons; S42. Obtain the vertex coordinates of the virtual wall, calculate the equation of the line on which the virtual wall is located, determine the position of the passenger and the position of the virtual wall through the geometric relationship between the point and the line, and obtain the passenger intrusion situation in the warning area.

7. A detection system for passenger intrusion behavior in urban rail transit stations that implements the detection method according to any one of claims 1-6, characterized in that, include: Optical flow feature extraction module: It has a built-in pyramid layered LK optical flow algorithm to perform optical flow tracking on the received urban rail transit station monitoring video images to extract feature points of moving passengers in the images, and send the feature point information to the moving passenger positioning module; Moving passenger positioning module: It has an elliptic domain subtractive clustering algorithm, receives the feature point information of moving passengers sent by the optical flow feature extraction module, and obtains the location information of moving passengers by clustering the feature points of moving passengers through the elliptic domain subtractive clustering algorithm, and sends the location information of moving passengers to the station passenger tracking module. Station passenger tracking module: It has a built-in Mean Shift target tracking algorithm, receives the location information of moving passengers sent by the moving passenger positioning module, updates the location information of moving passengers through the Mean Shift target tracking algorithm, thereby predicting the passenger's location at a future time, and sends the predicted future passenger's location to the passenger intrusion detection module; Passenger Intrusion Detection Module: Equipped with virtual wall detection technology, it determines the positional relationship between the received future passenger location and the virtual wall, concludes whether the passenger is within the virtual wall area, thereby obtaining passenger intrusion information within the warning area and issuing corresponding warning signals.

Images