Method for monitoring regional passenger flow density based on video image recognition

Abstract

This invention discloses a method for monitoring regional passenger flow density based on video image recognition, belonging to the field of video passenger flow density monitoring technology. The method includes structuring a video stream to segment independent moving entities and calculating their trajectory overlap density distribution, identifying high-frequency interactive nodes and low-frequency silent regions. Based on this, a spatial pressure field model is constructed, and its gradient characteristics are used to dynamically correct the estimated entity motion velocity. The corrected velocity and density distribution are integrated to form a dynamic density field, and temporal slicing analysis is performed to extract field strength fluctuation patterns. Historical data of this pattern drives the adaptive updating of the warning threshold. This method achieves deep perception of dynamic passenger flow behavior, improving the accuracy of density monitoring and the system's adaptive warning capability.

Description

Technical Field

[0001] This invention belongs to the field of video passenger flow density monitoring technology, specifically a method for monitoring regional passenger flow density based on video image recognition. Background Technology

[0002] Currently, regional passenger flow monitoring mainly relies on image-based target statistics and heatmap generation technologies. These methods assess the degree of congestion in an area by detecting and counting the number of people in a single frame or by generating static density distribution maps using image pixel features. Their early warning mechanisms are generally triggered based on preset fixed density thresholds, simplifying passenger flow status into an instantaneous, static quantitative indicator.

[0003] These technical solutions have shortcomings. Static analysis cannot capture the movement and dynamic interaction of passenger flow within the screen, ignoring the risk differences caused by different speeds at the same density. Furthermore, relying on manually preset and unchanging thresholds makes the system unable to adapt to temporal changes in the location and the long-term evolution of passenger behavior patterns, limiting the accuracy and scenario adaptability of early warnings.

[0004] This invention aims to address how to move beyond static counting by analyzing the interactive characteristics of passenger flow trajectories to construct a dynamic analysis model that integrates density and speed information. The system needs to be able to autonomously learn from the continuous evolution of this dynamic model, achieving intelligent iteration of early warning judgment criteria, thereby enabling a more accurate and adaptive risk assessment of the regional passenger flow safety status. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art;

[0006] Therefore, this invention proposes a method for monitoring regional passenger flow density based on video image recognition, including:

[0007] Acquire the raw video image stream provided by the monitoring equipment;

[0008] The original video image stream is decomposed into a structured form to generate a multi-level image segment set. Independent moving entity contours are identified and segmented from the multi-level image segment set.

[0009] Calculate the trajectory overlap density distribution map based on the set of displacement trajectories of the independent moving entity contours, and identify high-frequency interactive nodes and low-frequency silent regions within the trajectory overlap density distribution map;

[0010] A spatial pressure field model is generated based on the geometric relationship between the high-frequency interactive nodes and the low-frequency silent region. The estimated value of the entity's motion velocity is dynamically corrected based on the gradient change characteristics of the spatial pressure field model.

[0011] The corrected entity motion velocity estimate and trajectory overlap density distribution map are integrated to form a dynamic density field. Time-domain slice analysis is performed on the dynamic density field to extract the field intensity fluctuation pattern of each time slice.

[0012] The monitoring and early warning thresholds are adaptively updated based on historical evolution data of the field strength fluctuation pattern.

[0013] Preferably, the structural decomposition of the original video image stream further includes:

[0014] The original video image stream is divided into time block sequences according to a fixed time length;

[0015] Each of the time block sequences is spatially gridded to form a spatial grid sequence;

[0016] Temporal downsampling is performed on consecutive frames in each of the spatial grid sequences to generate a reduced frame rate grid segment;

[0017] Perform multi-scale convolutional filtering on the reduced frame rate grid segment to obtain an image pyramid hierarchy containing different levels of detail;

[0018] Separate the background stabilization layer and the foreground dynamic layer from the image pyramid hierarchy;

[0019] The foreground dynamic layer is divided into a significant motion layer and a micro-motion layer according to the motion amplitude threshold.

[0020] The significant motion layer, the micro-motion layer, and the background stabilization layer are combined to form the multi-level image segment set.

[0021] Preferably, the identification and segmentation of the contours of independent moving entities further includes:

[0022] A connected component labeling algorithm is applied to the salient motion layer of the multi-level image segment set;

[0023] Based on the results of the connected component labeling algorithm, the morphological feature vector of each labeled region is calculated;

[0024] The morphological feature vectors are used to perform motion trajectory association matching on the marked regions in adjacent frames;

[0025] The marked regions that have completed motion trajectory association matching are precisely extracted to obtain an initial entity contour set;

[0026] Optical flow accumulation calculation is performed in the micro-motion layer of the multi-level image segment set to form a motion energy map;

[0027] Adaptive threshold segmentation is performed on the motion energy map to identify micro-moving entity regions;

[0028] The spatial positions of the micro-moving entity region and the initial entity contour set are compared and fused.

[0029] All merged entity regions are smoothed and holes are filled to finally output the independent moving entity contours.

[0030] Preferably, the calculated trajectory overlap density distribution map further includes:

[0031] Establish a spatiotemporal coordinate chain for each of the independent moving entity contours in its continuously occurring frame sequence;

[0032] Project the spatiotemporal coordinate chains corresponding to the contours of all independent moving entities onto the same two-dimensional plane to form a set of trajectory lines;

[0033] The two-dimensional plane is divided into a uniform density calculation cell grid;

[0034] The number of trajectory lines passing through each density calculation unit grid is counted and used as the basic overlap density value;

[0035] Calculate the average angle between different trajectory lines passing through the same density calculation cell grid, and use it as the direction conflict coefficient;

[0036] The basic overlap density value is weighted and fused with the corresponding directional conflict coefficient to obtain the weighted overlap density value.

[0037] Gaussian smoothing filtering is applied to the weighted overlapping density values ​​of all density calculation cell grids to generate the trajectory overlapping density distribution map.

[0038] Preferably, the generation of the spatial pressure field model further includes:

[0039] In the trajectory overlap density distribution map, cell grids with weighted overlap density values ​​greater than a preset high threshold are marked as candidate high-frequency interaction nodes;

[0040] In the trajectory overlap density distribution map, continuous regions with weighted overlap density values ​​less than a preset low threshold are marked as candidate low-frequency silent regions;

[0041] Calculate the Euclidean distance between each candidate high-frequency interaction node and its nearest candidate low-frequency silent region to generate the pressure attenuation distance;

[0042] Based on the pressure attenuation distance and the weighted overlap density value of the candidate high-frequency interaction nodes, the theoretical pressure impact value of the candidate high-frequency interaction nodes on the surrounding space is calculated.

[0043] Based on the theoretical pressure influence values ​​of all candidate high-frequency interaction nodes, interpolation calculations are performed using radial basis functions on a two-dimensional plane to generate a continuous pressure field surface.

[0044] The pressure field surface is normalized so that its numerical range falls between zero and one, thus obtaining the spatial pressure field model.

[0045] Preferably, the dynamically corrected entity motion velocity estimate further includes:

[0046] Obtain the spatiotemporal coordinate chain of the independent moving entity contour, and calculate the displacement between adjacent frames as instantaneous velocity;

[0047] In the spatial pressure field model, find the pressure gradient value at the location of the center point of the contour of each independent moving entity;

[0048] Establish a mapping table between pressure gradient values ​​and velocity correction factors. The mapping table is configured such that the larger the pressure gradient value, the smaller the corresponding velocity attenuation coefficient value.

[0049] According to the mapping table, the instantaneous velocity of each independent moving entity contour is matched with a corresponding velocity correction factor;

[0050] Multiply the instantaneous velocity by the corresponding velocity correction factor to obtain the pressure-corrected instantaneous velocity;

[0051] The average value of all pressure-corrected instantaneous velocities of each independent moving entity profile during the observation period is taken as the final motion velocity estimate of the independent moving entity profile. The final motion velocity estimates of all entities constitute the set of corrected entity motion velocity estimates.

[0052] Preferably, the formation of the dynamic density field further includes:

[0053] The weighted overlap density value of the aforementioned trajectory overlap density distribution map is used as the basis for static spatial density;

[0054] Extract the average motion velocity of entities within each density calculation cell grid from the corrected set of entity motion velocity estimates;

[0055] The average motion speed is mapped into a speed influence factor through a nonlinear transformation function;

[0056] The dynamic density factor is obtained by convolving the static spatial density basis of each density calculation unit grid with the corresponding velocity influence factor.

[0057] The two-dimensional matrix containing the dynamic density factors of all density calculation cell grids is defined as the initial dynamic density field at the current moment.

[0058] Align the initial dynamic density field at the current moment with the historical dynamic density fields at multiple consecutive past moments in the time domain.

[0059] A Kalman filter is applied to the time-domain aligned continuous dynamic density field sequence to output the smoothed dynamic density field.

[0060] Preferably, the extraction of field strength fluctuation patterns for each time slice further includes:

[0061] The dynamic density field is sampled at equal intervals along the time axis to generate a series of equidistant time slices;

[0062] Calculate the global average field strength for the dynamic density field corresponding to each time slice, and use it as the base field strength of the time slice;

[0063] Calculate the difference between all local maxima and local minima in the dynamic density field of each time slice to form a sequence of field strength fluctuation ranges;

[0064] Analyze the field strength fluctuation range sequence of multiple consecutive time slices, and use an autoregressive model to predict the fluctuation range of the next time slice;

[0065] The basic field strength of each time slice is combined with its corresponding predicted fluctuation range to form the field strength fluctuation pattern feature vector of the time slice.

[0066] The field strength fluctuation pattern feature vectors of all time slices are arranged in chronological order to form the field strength fluctuation pattern evolution sequence.

[0067] Preferably, the adaptive update of the driving monitoring and early warning threshold further includes:

[0068] The initial value for the monitoring and early warning threshold is set to an empirical constant;

[0069] Extract the feature vector of the field strength fluctuation pattern within the most recent complete cycle from the evolution sequence of the field strength fluctuation pattern;

[0070] Calculate the slope of the trend of the basic field strength in the feature vector of the field strength fluctuation mode;

[0071] Calculate the moving average of the predicted fluctuation range values ​​in the feature vector of the field strength fluctuation mode;

[0072] The slope of the changing trend and the moving average are input into a pre-trained neural network model;

[0073] The neural network model outputs a threshold adjustment coefficient between zero and one.

[0074] Multiply the threshold adjustment coefficient by the current monitoring and early warning threshold to obtain the updated candidate value of the monitoring and early warning threshold;

[0075] The updated monitoring and early warning threshold candidate values ​​are subject to a change range constraint to ensure that the adjustment does not exceed a preset percentage, and finally a new monitoring and early warning threshold is generated.

[0076] Preferably, after generating the new monitoring and early warning threshold, the method further includes:

[0077] Calculate the instantaneous global field strength of the current dynamic density field in real time;

[0078] The instantaneous value of the global field strength is compared with the new monitoring and early warning threshold;

[0079] If the instantaneous value of the global field strength continues to exceed the new monitoring and early warning threshold for a preset duration, a tiered early warning protocol is triggered.

[0080] Based on the magnitude by which the instantaneous global field strength exceeds the threshold, a corresponding warning level is selected from the hierarchical warning protocol;

[0081] Execute the preset response command that is bound to the selected warning level.

[0082] Compared with the prior art, the beneficial effects of the present invention are:

[0083] By analyzing the displacement trajectories of independently moving entities, an overlap density distribution map is generated, and high-frequency interaction nodes and low-frequency silent regions within these trajectories are identified. A spatial pressure field model is constructed based on the geometric relationship between these two types of regions, transforming the static distribution into a pressure field with a spatial gradient. This gradient change is used to dynamically correct the initially observed entity velocity estimates. This process incorporates constraint information generated by local density and spatial structure into the velocity estimates. In densely populated areas or path intersections, the corrected velocity estimates more accurately characterize the actual movement capability of entities in complex environments, providing a dynamic parameter basis reflecting real physical interactions for subsequent analysis.

[0084] The corrected entity velocity estimates are integrated with the trajectory overlap density distribution map to form a dynamic density field that simultaneously encodes spatial density and velocity information. Temporal slicing analysis is performed on this dynamic density field to extract field strength fluctuation patterns and evolution sequences within different time segments. These time-series patterns extracted from continuous monitoring data are directly used to drive the adaptive update mechanism of the monitoring and early warning thresholds. The system's judgment criteria can be dynamically adjusted based on historical and real-time field strength fluctuation patterns to adapt to changes in passenger flow behavior characteristics under different time periods and scenarios, thereby improving the accuracy and scenario robustness of state discrimination and risk warning. Attached Figure Description

[0085] Figure 1 This is a flowchart illustrating the steps of the regional passenger flow density monitoring method based on video image recognition described in this invention.

[0086] Figure 2 A flowchart for identifying and segmenting the contours of independent moving entities;

[0087] Figure 3 A correlation analysis diagram of pressure gradient and speed correction in passenger flow density monitoring;

[0088] Figure 4 A heat map showing the density distribution of overlapping passenger flow trajectories in the region;

[0089] Figure 5 This is a graph showing the fluctuation range of passenger flow density field strength. Detailed Implementation

[0090] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0091] Please see Figure 1 The overall implementation scheme of this invention is as follows: Acquire the original video image stream captured by the monitoring equipment. Perform a structured decomposition operation on the original video image stream to generate a multi-level image segment set, and identify and segment independent moving entity contours within this set. Based on the set of displacement trajectories of these independent moving entity contours, calculate a trajectory overlap density distribution map, and identify high-frequency interactive nodes and low-frequency silent regions in this distribution map. Based on the geometric relationship between the high-frequency interactive nodes and the low-frequency silent regions, generate a spatial pressure field model, and dynamically correct the estimated motion velocity of the entities based on the gradient change characteristics of this model. Integrate the corrected estimated motion velocity of the entities with the trajectory overlap density distribution map to form a dynamic density field. Perform time-domain slice analysis on this dynamic density field to extract the field strength fluctuation patterns of each time slice. Finally, based on the historical evolution data of the field strength fluctuation patterns, drive the adaptive update of the monitoring and early warning threshold.

[0092] See Figure 2In one embodiment of the present invention, the original video image stream is divided into time block sequences according to a fixed time length. Each time block sequence is spatially gridded to form a spatial grid sequence. Continuous frames in each spatial grid sequence are temporally downsampled to generate reduced frame rate grid segments. Multi-scale convolutional filtering is performed on the reduced frame rate grid segments to obtain an image pyramid hierarchy containing different levels of detail. A background stabilization layer and a foreground dynamic layer are separated from the image pyramid hierarchy. The foreground dynamic layer is divided into a salient motion layer and a micro-motion layer according to a motion amplitude threshold. The salient motion layer, micro-motion layer, and background stabilization layer are combined to form a multi-level image segment set. A connected component labeling algorithm is applied to the salient motion layer of the multi-level image segment set. Based on the results of the connected component labeling algorithm, a morphological feature vector is calculated for each labeled region. Motion trajectory association matching is performed on the labeled regions in adjacent frames using the morphological feature vectors. Precise contour extraction is performed on the labeled regions that have completed motion trajectory association matching to obtain an initial entity contour set. Optical flow accumulation calculation is performed in the micro-motion layer of the multi-level image segment set to form a motion energy map. Adaptive threshold segmentation is applied to the motion energy map to identify micro-moving entity regions. These regions are then spatially compared and fused with the initial entity contour set. Contour smoothing and hole filling are performed on all fused entity regions to ultimately output independent moving entity contours.

[0093] In a specific implementation, an embodiment of a method for monitoring regional passenger flow density based on video image recognition involves the structured decomposition of the original video image stream and the identification and segmentation of the contours of independent moving entities. The implementation process can be described in conjunction with a monitoring scenario of an indoor square, where the original video image stream originates from a fixed camera deployed on the top of the square.

[0094] In practice, the original video stream is divided into time block sequences with a fixed duration of 30 seconds, thus transforming a continuous video stream into a series of independent processing time units. Each time block sequence is spatially gridded, for example, dividing each frame into a 32-pixel by 32-pixel grid, forming a spatial grid sequence. This partitioning discretizes the image into regular units in space. Consecutive frames in each spatial grid sequence are temporally downsampled at half the original frame rate, generating downsampled grid segments. This process reduces the number of frames that need to be processed. Multi-scale convolutional filtering is performed on the downsampled grid segments, using convolutional kernels of sizes 3x3, 5x5, and 7x7 to obtain image pyramid levels containing different levels of detail. Larger convolutional kernels focus on extracting macroscopic motion features, while smaller kernels retain finer details. The image pyramid hierarchy separates a background stabilization layer and a foreground dynamic layer. This separation is achieved by calculating the statistical variance of pixel values ​​within each grid across multiple consecutive frames. Regions with variance below a threshold are classified as background stabilization layers, while those above the threshold are classified as foreground dynamic layers. The foreground dynamic layer is further divided into salient motion layers and micro-motion layers based on motion amplitude thresholds. Motion amplitude is calculated by determining the average magnitude of the optical flow vector within each grid of the foreground dynamic layer. Grids with a magnitude greater than a set threshold are classified as salient motion layers, while those with a magnitude less than a threshold are classified as micro-motion layers. Combining the salient motion layer, micro-motion layer, and background stabilization layer creates a multi-level image fragment set. This set provides structured data containing different motion characteristics and spatial details for subsequent processing.

[0095] In some embodiments, a connected component labeling algorithm is applied to the salient motion layer of a multi-level image segment set. This algorithm labels regions with similar pixel values ​​and adjacent positions in the image as the same connected component, thereby initially identifying potential moving entity blocks. Based on the results of the connected component labeling algorithm, a morphological feature vector is calculated for each labeled region. The morphological feature vector includes the region's area, perimeter, aspect ratio of the bounding rectangle, and Hu moments of the contour. Motion trajectory association matching is performed on the labeled regions in adjacent frames using the morphological feature vectors. The matching process is achieved by calculating the Euclidean distance between the feature vectors and establishing associations for regions with the smallest distance. The contours of the labeled regions that have completed motion trajectory association matching are accurately extracted, and the active contour model algorithm is used to iteratively converge the boundaries of the regions to obtain an initial entity contour set.

[0096] In some embodiments, optical flow accumulation calculation is performed in the micro-motion layer of a multi-level image segment set. A dense optical flow field is calculated over five consecutive frames, and the optical flow amplitude of each frame is accumulated to form a motion energy map. Adaptive threshold segmentation is performed on the motion energy map. The adaptive threshold is automatically determined using the Otsu algorithm based on the overall grayscale distribution of the motion energy map to identify micro-motion entity regions. The micro-motion entity regions are spatially compared and fused with the initial entity contour set. The comparison process calculates the intersection-union ratio (IUR) between the micro-motion entity regions and the initial entity contour set. When the IUR is greater than 0.5, the micro-motion entity region is merged into a nearby initial entity contour; otherwise, it is added as a new entity contour. Contour smoothing and hole filling are performed on all fused entity regions. The smoothing operation uses Gaussian filtering to filter the contour point coordinates, and hole filling is performed using morphological closing operations. Finally, independent moving entity contours are output.

[0097] Optionally, for the division of the foreground dynamic layer in structured decomposition, a formulaic approach can be used to determine the salient motion layer and the micro-motion layer. One implementation defines a saliency discrimination factor, the calculation of which is expressed as follows: ;

[0098] Where: symbol Represents grid cells The significance value of the motion; symbol Represents the number of consecutive frames used for optical flow calculation; symbol It is in the Mesh cells calculated on the frame The optical flow vector at the location. By comparing this saliency value with a preset amplitude threshold, the significant motion layer and the micro-motion layer are distinguished.

[0099] It is understandable that through the above-mentioned structured decomposition and contour segmentation processing, the original video stream is transformed into a series of independent moving entity contour objects with clear spatiotemporal attributes. These objects directly represent the individual existence and motion state of pedestrians in the monitoring scene, providing accurate input data for subsequent trajectory analysis and density calculation.

[0100] Optionally, the size of the spatial grid and the frequency of temporal downsampling are not fixed and can be adjusted according to the actual resolution of the monitoring scene and the computing power of the processing system. For example, a larger grid size can be used in high-resolution video streams to balance processing accuracy and computing load.

[0101] It is understandable that the method of extracting entity contours from the salient motion layer and the micro-motion layer and then fusing them can effectively deal with pedestrian targets with different movement speeds. This ensures the detection rate of fast-moving individuals and avoids missing detection of slow-moving or briefly stationary individuals, thereby improving the completeness and accuracy of passenger flow statistics.

[0102] In one embodiment of the present invention, a spatiotemporal coordinate chain is established for each independent moving entity contour in its continuously occurring frame sequence. The spatiotemporal coordinate chains corresponding to all independent moving entity contours are projected onto the same two-dimensional plane to form a set of trajectory lines. The two-dimensional plane is divided into a uniform density calculation cell grid. The number of trajectory lines passing through each density calculation cell grid is counted as the basic overlap density value. The average angle between different trajectory lines passing through the same density calculation cell grid is calculated as the directional conflict coefficient. The basic overlap density value and the corresponding directional conflict coefficient are weighted and fused to obtain a weighted overlap density value. Gaussian smoothing filtering is applied to the weighted overlap density values ​​of all density calculation cell grids to generate a trajectory overlap density distribution map.

[0103] In a specific implementation, an embodiment of a method for monitoring regional passenger flow density based on video image recognition involves the calculation of trajectory overlap density distribution map. The implementation process can be described in conjunction with a monitoring scenario of a subway station, in which the outlines of independent moving entities have been pre-identified and segmented.

[0104] In practice, a spatiotemporal coordinate chain is established for each independent moving entity contour within its continuously appearing frame sequence. This chain consists of a series of coordinate points ordered by timestamps, each containing horizontal and vertical coordinates on the image plane and its corresponding frame number. All spatiotemporal coordinate chains corresponding to the independent moving entity contours are projected onto the same two-dimensional plane, forming a set of trajectory lines. The projection process ignores the time dimension, mapping all coordinate points to the same image coordinate system representing physical space. The two-dimensional plane is divided into a uniform density calculation cell grid, with the grid size set to match the human shoulder width scale, for example, the pixel size corresponding to 0.5 meters multiplied by 0.5 meters in the image. The number of trajectory lines crossing within each density calculation cell grid is counted as the basic overlap density value; a trajectory line is counted if any segment crosses the grid. The average angle between different trajectory lines crossing the same density calculation cell grid is calculated as the direction conflict coefficient. For cases with more than two trajectory lines within a grid, the average angle between the direction vectors of all pairwise trajectory line segments is calculated.

[0105] In some embodiments, the basic overlap density value and the corresponding directional conflict coefficient are weighted and fused to obtain a weighted overlap density value. The weighting and fusion is performed using a linear weighting method. Gaussian smoothing is applied to the weighted overlap density values ​​of all density calculation cell grids to generate a trajectory overlap density distribution map. The standard deviation of the Gaussian smoothing kernel function is set to the width of one grid cell to eliminate abrupt changes in density values ​​that may be caused by grid boundary division.

[0106] In some embodiments, the calculation of the directional conflict coefficient considers the local orientation of the trajectory. For each trajectory segment crossing the same grid, a small segment near the grid center point is taken to estimate the direction vector. The formula for weighted fusion of the basic overlap density value and the directional conflict coefficient can be expressed as: ;

[0107] Where: symbol Represents the grid coordinates Weighted overlap density value at; symbol Represents the basic overlap density value at the same location; symbol Represents the calculated directional conflict coefficient; symbol It is a preset weighting factor used to adjust the contribution of directional conflict to the final density value.

[0108] Optionally, the grid division of the two-dimensional plane can be non-uniform, with finer grid division used in known channels or bottleneck areas in the monitoring scenario, and coarser grid division used in open areas, to adapt to the different needs of density monitoring accuracy in different areas.

[0109] It is understandable that by introducing a directional conflict coefficient to correct the basic overlap density value, the trajectory overlap density distribution map not only reflects the density of trajectories in spatial location, but also characterizes the consistency of movement direction. Overlapping areas with chaotic directions usually mean higher personnel interaction and potential passage resistance, and their density weight should be enhanced.

[0110] Optionally, before performing Gaussian smoothing filtering, the weighted overlap density values ​​can be normalized to ensure that the values ​​of all grids are within the same dimension range, thus ensuring that the smoothing effect is consistent across regions with different density levels.

[0111] It is understandable that the trajectory overlap density distribution map visually presents the spatial hotspots and cold zones accumulated by historical trajectories within the monitoring area. High-frequency interactive nodes are represented by high-value continuous areas in the distribution map, while low-frequency silent areas are represented by low-value continuous areas. This provides a direct input basis for subsequent spatial pressure analysis.

[0112] In one embodiment of the present invention, in the trajectory overlap density distribution map, cell grids with weighted overlap density values ​​greater than a preset high threshold are marked as candidate high-frequency interaction nodes. In the trajectory overlap density distribution map, continuous regions with weighted overlap density values ​​less than a preset low threshold are marked as candidate low-frequency silent regions. The Euclidean distance between each candidate high-frequency interaction node and its nearest candidate low-frequency silent region is calculated to generate a pressure attenuation distance. Based on the pressure attenuation distance and the weighted overlap density value of the candidate high-frequency interaction node, the theoretical pressure influence value of the candidate high-frequency interaction node on the surrounding space is calculated. Based on the theoretical pressure influence values ​​of all candidate high-frequency interaction nodes, interpolation calculation is performed on a two-dimensional plane using radial basis functions to generate a continuous pressure field surface. The pressure field surface is normalized so that its value range falls between zero and one, resulting in a spatial pressure field model. The spatiotemporal coordinate chain of the contours of independent moving entities is obtained, and the displacement between adjacent frames is calculated as instantaneous velocity. In the spatial pressure field model, the pressure gradient value at the location of the center point of each independent moving entity contour is found. A mapping table is established between pressure gradient values ​​and velocity correction factors. The table is configured such that the larger the pressure gradient value, the smaller the corresponding velocity attenuation coefficient. Based on the mapping table, a corresponding velocity correction factor is matched to the instantaneous velocity of each independent moving entity contour. The instantaneous velocity is multiplied by the corresponding velocity correction factor to obtain the pressure-corrected instantaneous velocity. The average of all pressure-corrected instantaneous velocities for each independent moving entity contour within the observation period is taken as the final velocity estimate for that contour. The final velocity estimates of all entities constitute the corrected entity velocity estimate set.

[0113] In a specific implementation, an embodiment of a regional passenger flow density monitoring method based on video image recognition involves the generation of a spatial pressure field model and the dynamic correction of the estimated value of entity movement speed. The implementation process can be described in conjunction with a monitoring scenario of a railway station entrance hall, in which the trajectory overlap density distribution map has been pre-calculated.

[0114] In practical implementation, in the trajectory overlap density distribution map, cell grids with weighted overlap density values ​​greater than a preset high threshold are marked as candidate high-frequency interaction nodes. The preset high threshold can be determined by analyzing the statistical quantiles of historical distribution maps; for example, the top 15% of cell grids with weighted overlap density values ​​can be marked. In the trajectory overlap density distribution map, continuous regions with weighted overlap density values ​​less than a preset low threshold are marked as candidate low-frequency silent regions. The preset low threshold can be set to half the overall mean of the distribution map, and the area of ​​the continuous region must be greater than a minimum area threshold to exclude noise. The Euclidean distance between each candidate high-frequency interaction node and its nearest candidate low-frequency silent region is calculated to generate a pressure attenuation distance. The Euclidean distance is calculated based on the center coordinates of the cell grid. Based on the pressure attenuation distance and the weighted overlap density values ​​of the candidate high-frequency interaction nodes, the theoretical pressure impact value of the candidate high-frequency interaction nodes on the surrounding space is calculated.

[0115] In some embodiments, the theoretical pressure influence value can be calculated using a distance-attenuation-based model. This model treats candidate high-frequency interaction nodes as pressure sources, whose influence weakens with increasing spatial distance. Based on the theoretical pressure influence values ​​of all candidate high-frequency interaction nodes, interpolation is performed on a two-dimensional plane using radial basis functions to generate a continuous pressure field surface. The radial basis function used is a Gaussian kernel function. The pressure field surface is then normalized so that its values ​​fall between zero and one, resulting in a spatial pressure field model. The normalization process maps the maximum value on the surface to one and the minimum value to zero.

[0116] In some embodiments, the spatiotemporal coordinate chain of the contours of independent moving entities is obtained, and the displacement between adjacent frames is calculated as the instantaneous velocity. The bit removal is then divided by the inter-frame time interval to obtain the instantaneous velocity value. In the spatial pressure field model, the pressure gradient value at the center point of each independent moving entity contour is found. The pressure gradient value is obtained by performing a two-dimensional gradient operation on the spatial pressure field model, reflecting the severity of pressure change at that location. A mapping table between pressure gradient values ​​and velocity correction factors is established. The mapping table is configured such that the larger the pressure gradient value, the smaller the mapped velocity attenuation coefficient, i.e., a stronger attenuation effect is applied to the motion velocity in the high pressure gradient region.

[0117] Optionally, the theoretical pressure impact value of candidate high-frequency interaction nodes can be expressed by the following formula: ;

[0118] Where: symbol Represents the theoretical pressure influence value generated by the k-th candidate high-frequency interaction node at a distance r; symbol Represents the weighted overlap density value of the candidate high-frequency interaction node; symbol Represents the Euclidean distance to candidate high-frequency interaction nodes; symbol It is a small positive constant used to prevent the denominator from being zero. The total pressure influence value of all candidate high-frequency interaction nodes at a certain point in space is determined by their respective... The summation of contributions yields the total influence value, which, after subsequent radial basis function interpolation and normalization, ultimately forms the spatial pressure field model.

[0119] It is understandable that the spatial pressure field model transforms discrete trajectory density information into a continuous spatial pressure representation. The pressure value depends not only on the local density but also on the spatial layout relationship between high-density nodes and low-density areas, thus better reflecting the crowding and traffic pressure actually felt by people in physical space.

[0120] Optionally, the mapping relationship between the velocity correction factor and the pressure gradient value can be designed as a piecewise linear function or a nonlinear lookup table function. The specific form can be calibrated according to the response characteristics of crowd movement to the pressure gradient in the actual scenario.

[0121] Understandably, based on the mapping table, a corresponding velocity correction factor is matched to the instantaneous velocity of each independent moving entity contour. Multiplying the instantaneous velocity by the corresponding velocity correction factor yields the pressure-corrected instantaneous velocity. The average of all pressure-corrected instantaneous velocities for each independent moving entity contour within the observation period is taken as the final velocity estimate for that entity. The final velocity estimates for all entities constitute the corrected set of entity velocity estimates. This process ensures that the velocity estimation is not only based on apparent displacement but also incorporates corrections from the spatial pressure field, thus more accurately reflecting the true state of individual movement in a congested environment.

[0122] See Figure 3 This is a correlation analysis chart of pressure gradient and speed correction in passenger flow density monitoring, clearly showing the changing trends of speed correction factor, original instantaneous speed, and corrected speed under different pressure gradients. It intuitively presents the complete transmission chain from pressure gradient to speed correction to actual speed, verifying the effectiveness of the "dynamic correction of physical movement speed" in the technical solution. The high correlation of the three curves proves that the mapping model between pressure gradient and speed correction factor is reasonable and can effectively reflect the movement patterns of crowds in real-world scenarios. It provides quantitative evidence for passenger flow density monitoring systems, allowing for real-time estimation of actual pedestrian speeds through pressure gradients, thereby triggering more accurate early warning and crowd control strategies. By analyzing the pressure gradients in different areas, guidance signs or personnel deployment can be dynamically adjusted to alleviate traffic pressure in high-pressure gradient areas.

[0123] In one embodiment of the present invention, the weighted overlap density value of the trajectory overlap density distribution map is used as the static spatial density basis. The average motion velocity of entities within each density calculation unit grid is extracted from the corrected set of entity motion velocity estimates. The average motion velocity is mapped to a velocity influence factor through a nonlinear transformation function. The static spatial density basis of each density calculation unit grid is convolved with the corresponding velocity influence factor to obtain the dynamic density factor. A two-dimensional matrix containing the dynamic density factors of all density calculation unit grids is defined as the initial dynamic density field at the current moment. The initial dynamic density field at the current moment is time-domain aligned with the historical dynamic density fields of multiple consecutive past moments. A Kalman filter is applied to the time-domain aligned continuous dynamic density field sequence to output a smoothed dynamic density field. The dynamic density field is sampled at equal intervals along the time axis to generate a series of equidistant time slices. The global field strength mean is calculated for the dynamic density field corresponding to each time slice, which is used as the base field strength of the time slice. The difference between all local maxima and local minima in the dynamic density field of each time slice is calculated to form a field strength fluctuation range sequence. The field strength fluctuation range sequence of multiple consecutive time slices is analyzed, and an autoregressive model is used to predict the fluctuation range of the next time slice. The basic field strength of each time slice is combined with its corresponding predicted fluctuation range to form the field strength fluctuation pattern feature vector of the time slice. The field strength fluctuation pattern feature vectors of all time slices are arranged in chronological order to form the field strength fluctuation pattern evolution sequence.

[0124] In a specific implementation, an embodiment of a method for monitoring regional passenger flow density based on video image recognition involves the formation of a dynamic density field and the extraction of field strength fluctuation patterns for each time slice. The implementation process can be described in conjunction with a monitoring scenario of an indoor atrium in a large shopping mall, where the trajectory overlap density distribution map and the set of corrected entity motion speed estimates have been obtained in advance.

[0125] In the specific implementation, the weighted overlap density value of the trajectory overlap density distribution map is used as the static spatial density basis. The static spatial density basis is a two-dimensional matrix, where each element corresponds to the weighted overlap density value of a density calculation unit grid. The average motion velocity of entities within each density calculation unit grid is extracted from the corrected set of entity motion velocity estimates. The average motion velocity is calculated by statistically analyzing the final motion velocity estimates of all independent moving entity contours located in the same grid within a specified time window and taking the arithmetic mean. The average motion velocity is mapped to a velocity influence factor through a nonlinear transformation function. The nonlinear transformation function is designed as a monotonically increasing S-shaped curve to limit the influence of motion velocity on density within a certain range and avoid extreme velocity values ​​causing the dynamic density factor to run away from control. The static spatial density basis of each density calculation unit grid is convolved with the corresponding velocity influence factor to obtain the dynamic density factor. The convolution operation uses a 3x3 local sliding window, and the static spatial density basis value of the central grid is weighted and summed with the velocity influence factors of the surrounding grids. The two-dimensional matrix containing the dynamic density factors of all density calculation unit grids is defined as the initial dynamic density field at the current moment. The initial dynamic density field at the current moment is temporally aligned with the historical dynamic density fields from multiple consecutive past moments. This temporal alignment ensures that all dynamic density fields are completely consistent in terms of spatial grid partitioning. A Kalman filter is then applied to the temporally aligned continuous dynamic density field sequence to output a smoothed dynamic density field. The Kalman filter is used to estimate the true state of the dynamic density field and reduce the influence of observation noise.

[0126] In some embodiments, the specific form of the nonlinear transformation function can be expressed by the following formula: ;

[0127] Where: symbol Represents the average speed of motion The velocity influence factor obtained from the mapping; symbol It is a positive scaling parameter; used to control the steepness of the S-curve; symbol It is a natural constant. Through this formula, the average velocity is converted into a velocity influence factor between -1 and +1, where a positive value indicates that velocity enhances the effect of density, and a negative value indicates that it weakens the effect.

[0128] In some embodiments, the smoothed dynamic density field is sampled at equal intervals along the time axis to generate a series of equidistant time slices. The sampling period is set to five seconds, thereby discretizing the continuous dynamic density field into a sequence of slices arranged in chronological order. The global mean field strength is calculated for the dynamic density field corresponding to each time slice, serving as the base field strength for that time slice. The global mean field strength is obtained by averaging all dynamic density factors in the dynamic density field matrix. The differences between all local maxima and local minima in the dynamic density field of each time slice are calculated to form a sequence of field strength fluctuation ranges. Local extrema are determined by comparing the values ​​of each dynamic density factor with the values ​​of other dynamic density factors within its eight neighborhoods.

[0129] Optionally, the analysis and prediction of the field strength fluctuation range sequence can use a first-order autoregressive model. This model predicts the fluctuation range of the next time slice based on the field strength fluctuation range values ​​of multiple consecutive past time slices. The base field strength of each time slice is combined with its corresponding predicted fluctuation range value to form a field strength fluctuation pattern feature vector for that time slice. This feature vector is a two-dimensional vector, with the first dimension being the base field strength and the second dimension being the predicted fluctuation range value. Arranging the field strength fluctuation pattern feature vectors of all time slices in chronological order forms a field strength fluctuation pattern evolution sequence, which can be used for subsequent monitoring and early warning threshold updates. See Table 1 for an illustration of the composition of the field strength fluctuation pattern feature vector.

[0130] Table 1: Feature Vector Data of Field Intensity Fluctuation Pattern in Time Slices Time slice number Baseline electric field strength (mean) Fluctuation range forecast value T1 0.45 0.12 T2 0.52 0.15 T3 0.61 0.18 T4 0.58 0.16 T5 0.49 0.13

[0131] Table 1 shows the characteristic vectors of field strength fluctuation patterns for five consecutive time slices. The basic field strength reflects the overall intensity level of the dynamic density field, while the predicted fluctuation range characterizes the uniformity of the density distribution within the field or the intensity of the fluctuations.

[0132] Optionally, when calculating local maxima and local minima, a minimum significance threshold can be set. Only when the difference between the dynamic density factor and its neighborhood value is greater than this threshold is it identified as an extreme value, in order to avoid misjudgment caused by noise.

[0133] It is understandable that the formation of the dynamic density field integrates static spatial density and dynamic motion speed information, so that density estimation not only reflects the spatial distribution concentration of people, but also incorporates the influence of people's motion state. Meanwhile, time-domain slice analysis discretizes the continuous spatiotemporal density evolution process into a series of pattern points with characteristic vectors, thus providing a structured data foundation for quantitative analysis of the temporal evolution law of passenger flow density.

[0134] See Figure 4This is a heatmap showing the density distribution of overlapping passenger flow trajectories in a region. It visually displays the varying passenger flow density at different grid locations using color depth. This map serves as direct input for generating a spatial pressure field model. By identifying the geometric relationship between high-frequency interactive nodes and low-frequency quiet areas, the traffic pressure value at each point in the space can be calculated. The spatiotemporal variation of density distribution is the core data driving the adaptive update of the warning threshold, reflecting the periodic fluctuations in passenger flow patterns. The location and extent of high-density nodes provide management personnel with precise guidance, such as deploying more guides or adjusting traffic routes in yellow areas. The weighted overlapping density values ​​in this map form the static density basis. Combined with a physical movement speed correction factor, a dynamic density field can be constructed, supporting more accurate passenger flow monitoring.

[0135] In one embodiment of the present invention, the initial value of the monitoring and early warning threshold is set to an empirical constant. The feature vector of the field strength fluctuation pattern within the most recent complete cycle is extracted from the field strength fluctuation pattern evolution sequence. The slope of the change trend of the basic field strength in the feature vector of the field strength fluctuation pattern is calculated. The moving average of the predicted fluctuation range value in the feature vector of the field strength fluctuation pattern is calculated. The slope of the change trend and the moving average are input into a pre-trained neural network model. The neural network model outputs a threshold adjustment coefficient between zero and one. The threshold adjustment coefficient is multiplied by the current monitoring and early warning threshold to obtain an updated monitoring and early warning threshold candidate value. A change amplitude constraint is applied to the updated monitoring and early warning threshold candidate value to ensure that its adjustment does not exceed a preset percentage, ultimately generating a new monitoring and early warning threshold. The instantaneous global field strength value of the current dynamic density field is calculated in real time. The instantaneous global field strength value is compared with the new monitoring and early warning threshold. If the instantaneous global field strength value continuously exceeds the new monitoring and early warning threshold for a preset duration, a tiered early warning protocol is triggered. Based on the magnitude by which the instantaneous global field strength value exceeds the threshold, the corresponding early warning level is selected from the tiered early warning protocol. The preset response command bound to the selected early warning level is executed.

[0136] In a specific implementation, an embodiment of a regional passenger flow density monitoring method based on video image recognition involves adaptive updating of monitoring and early warning thresholds and triggering of early warning mechanisms. The implementation process can be described in conjunction with a monitoring scenario of the main entrance and exit of a city park, in which the field strength fluctuation pattern evolution sequence has been pre-extracted.

[0137] In practical implementation, the initial value for setting the monitoring and early warning threshold is an empirical constant, which can be set based on the average global field strength of the maximum dynamic density field when no congestion accidents occur in historical safe operation data. From the field strength fluctuation pattern evolution sequence, the feature vector of the field strength fluctuation pattern within the most recent complete cycle is extracted. A complete cycle can be the time span of repeated passenger flow patterns, such as 24 hours a day or 7 days a week. The slope of the change trend of the basic field strength in the field strength fluctuation pattern feature vector is calculated. The slope is obtained by linearly fitting all basic field strength values ​​within the extracted cycle, and the slope reflects the overall direction of passenger flow density change in the recent period. The moving average of the predicted fluctuation range values ​​in the field strength fluctuation pattern feature vector is calculated using a simple moving average with a window length of 10.

[0138] In some embodiments, the slope of the trend and the moving average are input into a pre-trained neural network model. The neural network model employs a multilayer perceptron structure with a single hidden layer. Its input layer has two neurons corresponding to the slope of the trend and the moving average, respectively, and its output layer has one neuron. The neural network model outputs a threshold adjustment coefficient between zero and one, representing the adjustment magnitude of the current monitoring and warning threshold recommendation based on recent pattern evolution. The threshold adjustment coefficient is multiplied by the current monitoring and warning threshold to obtain an updated monitoring and warning threshold candidate value. A change magnitude constraint is applied to the updated monitoring and warning threshold candidate value to ensure that its adjustment does not exceed a preset percentage, ultimately generating a new monitoring and warning threshold. The preset percentage can be set to 20% to avoid overly drastic adjustments in a single instance.

[0139] In some embodiments, the training process of the neural network model is completed using historical data, including historical field strength fluctuation pattern feature vectors and the optimal monitoring and early warning thresholds determined by post-analysis within the corresponding time periods. The training objective is to minimize the mean square error between the threshold adjustment coefficients predicted by the neural network model and the optimal adjustment coefficients. The instantaneous global field strength value of the current dynamic density field is calculated in real time. The calculation method for the instantaneous global field strength value is consistent with the calculation method for the time-slice base field strength, i.e., averaging all dynamic density factors in the smoothed dynamic density field at the current moment. The instantaneous global field strength value is then compared with the new monitoring and early warning threshold.

[0140] Optionally, the generation process of the threshold adjustment coefficient can be schematically described by a mathematical expression, specifically in the form of: ;

[0141] Where: symbol The threshold adjustment coefficient representing the output of the neural network model; symbol Represents the Sigmoid activation function; symbol Represents the slope of the calculated trend; symbol Represents the calculated moving average; symbol and These are the weight parameters connecting the input layer and the hidden layer in a neural network model; (symbol) This is the bias term in a neural network model. The formula describes the fundamental process within the neural network model of weighted summation of two input features and mapping them to the final output coefficients via an activation function.

[0142] It is understandable that if the instantaneous global field strength value continuously exceeds the new monitoring and warning threshold for a preset duration, a tiered warning protocol will be triggered. The preset duration can be set according to the emergency response requirements of the scenario, for example, for thirty seconds. Based on the magnitude by which the instantaneous global field strength value exceeds the threshold, the corresponding warning level is selected from the tiered warning protocol. Multiple levels can be predefined in the tiered warning protocol, such as yellow, orange, and red warnings, each corresponding to a different range of exceedance. The preset response command bound to the selected warning level is executed. The preset response command may include displaying a visual warning in the monitoring system, sending notification information to management personnel, or linking the broadcast system to issue evacuation notices.

[0143] Optionally, the warning levels in the tiered warning protocol can be divided based on the percentage exceeding a threshold. For example, exceeding the threshold by less than 10% is a yellow warning, exceeding the threshold by 10% to 30% is an orange warning, and exceeding the threshold by more than 30% is a red warning.

[0144] It is understandable that by driving the adaptive update of the monitoring and early warning threshold based on the historical evolution data of the field strength fluctuation pattern, the early warning mechanism can adapt to the long-term changes and periodic fluctuations of passenger flow density patterns, avoiding the high false alarm or missed alarm problem that may occur when the fixed threshold changes the scene. The triggering and execution of the hierarchical early warning protocol provides managers with a hierarchical and operable response basis.

[0145] See Figure 5 This is a chart analyzing the fluctuation range of passenger flow density, demonstrating the complete process of original fluctuations, moving average filtering, and threshold triggering. Its core value lies in providing a quantitative basis for the dynamic updating of warning thresholds. The moving average effectively separates instantaneous fluctuations from long-term trends, avoiding misjudgments caused by accidental fluctuations. By comparing thresholds, the vague concept of "drastic fluctuations" is transformed into a quantifiable event exceeding limits, providing clear input features for subsequent neural network models. The moving average of the fluctuation range is one of the two core features driving the adaptive updating of the warning threshold, and together with the trend slope, it determines the magnitude of the threshold adjustment coefficient. By comparing the differences between the original fluctuations and the moving average, the effectiveness of the filtering algorithm can be verified, providing feedback for model iteration.

[0146] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for monitoring regional passenger flow density based on video image recognition, characterized in that, include: Acquire the raw video image stream provided by the monitoring equipment; The original video image stream is decomposed into a structured form to generate a multi-level image segment set. Independent moving entity contours are identified and segmented from the multi-level image segment set. Calculate the trajectory overlap density distribution map based on the set of displacement trajectories of the independent moving entity contours, and identify high-frequency interactive nodes and low-frequency silent regions within the trajectory overlap density distribution map; A spatial pressure field model is generated based on the geometric relationship between the high-frequency interactive nodes and the low-frequency silent region. The estimated value of the entity's motion velocity is dynamically corrected based on the gradient change characteristics of the spatial pressure field model. The corrected entity motion velocity estimate and trajectory overlap density distribution map are integrated to form a dynamic density field. Time-domain slice analysis is performed on the dynamic density field to extract the field intensity fluctuation pattern of each time slice. The monitoring and early warning thresholds are adaptively updated based on historical evolution data of the field strength fluctuation pattern.

2. The method for monitoring regional passenger flow density based on video image recognition as described in claim 1, characterized in that, The structural decomposition of the original video image stream further includes: The original video image stream is divided into time block sequences according to a fixed time length; Each of the time block sequences is spatially gridded to form a spatial grid sequence; Temporal downsampling is performed on consecutive frames in each of the spatial grid sequences to generate a reduced frame rate grid segment; Perform multi-scale convolutional filtering on the reduced frame rate grid segment to obtain an image pyramid hierarchy containing different levels of detail; Separate the background stabilization layer and the foreground dynamic layer from the image pyramid hierarchy; The foreground dynamic layer is divided into a significant motion layer and a micro-motion layer according to the motion amplitude threshold. The significant motion layer, the micro-motion layer, and the background stabilization layer are combined to form the multi-level image segment set.

3. The method for monitoring regional passenger flow density based on video image recognition as described in claim 2, characterized in that, The process of identifying and segmenting the contours of independent moving entities further includes: A connected component labeling algorithm is applied to the salient motion layer of the multi-level image segment set; Based on the results of the connected component labeling algorithm, the morphological feature vector of each labeled region is calculated; The morphological feature vectors are used to perform motion trajectory association matching on the marked regions in adjacent frames; The marked regions that have completed motion trajectory association matching are precisely extracted to obtain an initial entity contour set; Optical flow accumulation calculation is performed in the micro-motion layer of the multi-level image segment set to form a motion energy map; Adaptive threshold segmentation is performed on the motion energy map to identify micro-moving entity regions; The spatial positions of the micro-moving entity region and the initial entity contour set are compared and fused. All merged entity regions are smoothed and holes are filled to finally output the independent moving entity contours.

4. The method for monitoring regional passenger flow density based on video image recognition as described in claim 3, characterized in that, The calculated trajectory overlap density distribution map further includes: Establish a spatiotemporal coordinate chain for each of the independent moving entity contours in its continuously occurring frame sequence; Project the spatiotemporal coordinate chains corresponding to the contours of all independent moving entities onto the same two-dimensional plane to form a set of trajectory lines; The two-dimensional plane is divided into a uniform density calculation cell grid; The number of trajectory lines passing through each density calculation unit grid is counted and used as the basic overlap density value; Calculate the average angle between different trajectory lines passing through the same density calculation cell grid, and use it as the direction conflict coefficient; The basic overlap density value is weighted and fused with the corresponding directional conflict coefficient to obtain the weighted overlap density value. Gaussian smoothing filtering is applied to the weighted overlapping density values ​​of all density calculation cell grids to generate the trajectory overlapping density distribution map.

5. The method for monitoring regional passenger flow density based on video image recognition as described in claim 4, characterized in that, The generated spatial pressure field model further includes: In the trajectory overlap density distribution map, cell grids with weighted overlap density values ​​greater than a preset high threshold are marked as candidate high-frequency interaction nodes; In the trajectory overlap density distribution map, continuous regions with weighted overlap density values ​​less than a preset low threshold are marked as candidate low-frequency silent regions; Calculate the Euclidean distance between each candidate high-frequency interaction node and its nearest candidate low-frequency silent region to generate the pressure attenuation distance; Based on the pressure attenuation distance and the weighted overlap density value of the candidate high-frequency interaction nodes, the theoretical pressure impact value of the candidate high-frequency interaction nodes on the surrounding space is calculated. Based on the theoretical pressure influence values ​​of all candidate high-frequency interaction nodes, interpolation calculations are performed using radial basis functions on a two-dimensional plane to generate a continuous pressure field surface. The pressure field surface is normalized so that its numerical range falls between zero and one, thus obtaining the spatial pressure field model.

6. The method for monitoring regional passenger flow density based on video image recognition as described in claim 5, characterized in that, The dynamically corrected entity motion velocity estimate further includes: Obtain the spatiotemporal coordinate chain of the independent moving entity contour, and calculate the displacement between adjacent frames as instantaneous velocity; In the spatial pressure field model, find the pressure gradient value at the location of the center point of the contour of each independent moving entity; Establish a mapping table between pressure gradient values ​​and velocity correction factors. The mapping table is configured such that the larger the pressure gradient value, the smaller the corresponding velocity attenuation coefficient value. According to the mapping table, the instantaneous velocity of each independent moving entity contour is matched with a corresponding velocity correction factor; Multiply the instantaneous velocity by the corresponding velocity correction factor to obtain the pressure-corrected instantaneous velocity; The average value of all pressure-corrected instantaneous velocities of each independent moving entity profile during the observation period is taken as the final motion velocity estimate of the independent moving entity profile. The final motion velocity estimates of all entities constitute the set of corrected entity motion velocity estimates.

7. The method for monitoring regional passenger flow density based on video image recognition as described in claim 6, characterized in that, The formation of the dynamic density field further includes: The weighted overlap density value of the aforementioned trajectory overlap density distribution map is used as the basis for static spatial density; Extract the average motion velocity of entities within each density calculation cell grid from the corrected set of entity motion velocity estimates; The average motion speed is mapped into a speed influence factor through a nonlinear transformation function; The dynamic density factor is obtained by convolving the static spatial density basis of each density calculation unit grid with the corresponding velocity influence factor. The two-dimensional matrix containing the dynamic density factors of all density calculation cell grids is defined as the initial dynamic density field at the current moment. Align the initial dynamic density field at the current moment with the historical dynamic density fields at multiple consecutive past moments in the time domain. A Kalman filter is applied to the time-domain aligned continuous dynamic density field sequence to output the smoothed dynamic density field.

8. The method for monitoring regional passenger flow density based on video image recognition as described in claim 7, characterized in that, The extraction of field strength fluctuation patterns for each time slice further includes: The dynamic density field is sampled at equal intervals along the time axis to generate a series of equidistant time slices; Calculate the global average field strength for the dynamic density field corresponding to each time slice, and use it as the base field strength of the time slice; Calculate the difference between all local maxima and local minima in the dynamic density field of each time slice to form a sequence of field strength fluctuation ranges; Analyze the field strength fluctuation range sequence of multiple consecutive time slices, and use an autoregressive model to predict the fluctuation range of the next time slice; The basic field strength of each time slice is combined with its corresponding predicted fluctuation range to form the field strength fluctuation pattern feature vector of the time slice. The field strength fluctuation pattern feature vectors of all time slices are arranged in chronological order to form the field strength fluctuation pattern evolution sequence.

9. The method for monitoring regional passenger flow density based on video image recognition as described in claim 8, characterized in that, The adaptive update of the driving monitoring and early warning threshold further includes: The initial value for the monitoring and early warning threshold is set to an empirical constant; Extract the feature vector of the field strength fluctuation pattern within the most recent complete cycle from the evolution sequence of the field strength fluctuation pattern; Calculate the slope of the trend of the basic field strength in the feature vector of the field strength fluctuation mode; Calculate the moving average of the predicted fluctuation range values ​​in the feature vector of the field strength fluctuation mode; The slope of the changing trend and the moving average are input into a pre-trained neural network model; The neural network model outputs a threshold adjustment coefficient between zero and one. Multiply the threshold adjustment coefficient by the current monitoring and early warning threshold to obtain the updated candidate value of the monitoring and early warning threshold; The updated monitoring and early warning threshold candidate values ​​are subject to a change range constraint to ensure that the adjustment does not exceed a preset percentage, and finally a new monitoring and early warning threshold is generated.

10. The method for monitoring regional passenger flow density based on video image recognition as described in claim 9, characterized in that, After generating the new monitoring and early warning threshold, the method further includes: Calculate the instantaneous global field strength of the current dynamic density field in real time; The instantaneous value of the global field strength is compared with the new monitoring and early warning threshold; If the instantaneous value of the global field strength continues to exceed the new monitoring and early warning threshold for a preset duration, a tiered early warning protocol is triggered. Based on the magnitude by which the instantaneous global field strength exceeds the threshold, a corresponding warning level is selected from the hierarchical warning protocol; Execute the preset response command that is bound to the selected warning level.

Images