Real-time business circle flow analysis and boundary calibration system based on edge computing

Abstract

The application discloses a real-time commercial district people flow analysis and boundary calibration system based on edge computing and concretely relates to the technical field of edge computing; video streams and wireless probe signals collected by multiple edge computing nodes are used to obtain node observation records; the node observation records are mapped to a unified space grid and a time window to construct cross-node event sequences; the correlation confidence of cross-node moving targets is analyzed through dynamic time warping to mark repeated, fragmented and independent targets, so that people flow density, flow direction and flow conservation residual error are corrected; and finally, the active boundary of the commercial district is dynamically updated based on people flow correction data and grid connectivity. The application effectively solves the problem that the existing method cannot accurately and stably correlate cross-node targets in the edge node junction area, leading to abnormal changes of the active boundary of the commercial district, and improves the accuracy of people flow analysis and boundary calibration of the commercial district.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and more specifically, to a real-time business district pedestrian flow analysis and boundary calibration system based on edge computing. Background Technology

[0002] Currently, the pedestrian flow analysis methods widely used in the commercial field mainly adopt single-node or centralized statistical approaches. That is, pedestrian flow information is processed separately through each independent node or a unified data center. Each node usually independently collects video or wireless probe signals and separately identifies and counts pedestrian flow information within its coverage area. Then, the independent statistical results of each node are aggregated and analyzed as a whole.

[0003] Existing analysis methods lack an effective cross-node target association mechanism, which makes it easy for the same target to be counted repeatedly by different nodes when people move within the boundaries or overlapping areas of node coverage, or for the target to be fragmented into multiple isolated statistical individuals when crossing node areas. This affects the accuracy of people flow statistics and leads to deviations in the judgment of people flow density and direction. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a method and system for monitoring and evaluating the greening of urban gardens to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A real-time business district pedestrian flow analysis and boundary calibration system based on edge computing includes: Node recording module: Acquires node observation records of multiple edge computing nodes; Grid candidate module: Divide overlapping node groups according to the adjacency and overlap of node coverage areas, map node observation records to a unified spatial grid and a unified time window, and obtain boundary candidate grids; Cross-point event module: Extracts the target vanishing point, target appearing point, spatial gate crossing direction, and wireless signal continuation interval from the boundary candidate grid to construct a cross-node event sequence; The association confidence module performs dynamic time warping based on the entry and exit time phase difference, spatial gate crossing direction consistency, and wireless signal continuity interval in the cross-node event sequence to obtain the cross-node association confidence level. Pedestrian flow correction module: Based on the cross-node association confidence level, it labels repeated targets, split targets, and independent targets, and generates corrected pedestrian flow density, corrected pedestrian flow direction, and flow conservation residuals. Boundary Update Module: Updates the active boundary of the business district based on the corrected pedestrian density, corrected pedestrian flow direction, flow conservation residual, and the connectivity of the boundary candidate grid.

[0006] In a preferred embodiment, node observation records of multiple edge computing nodes are obtained, specifically as follows: Configure edge computing node identifiers for multiple edge computing nodes, and have multiple edge computing nodes collect video streams and wireless probe signals within the node's coverage area; Extract video detection data from the video stream; extract wireless probe data from the wireless probe signal. Read the node coordinates and node coverage area, and record the sampling time corresponding to the video detection data and wireless probe data; Calculate the node reporting delay based on the data transmission and reception times; Node observation records are generated by associating edge computing node identifiers with video detection data, wireless probe data, node coordinates, sampling time, node reporting delay, and node coverage.

[0007] In a preferred embodiment, overlapping node groups are divided according to the adjacency and overlap relationships of node coverage areas. Node observation records are then mapped to a unified spatial grid and a unified time window to obtain boundary candidate grids. Specifically: Based on the edge calculation node identifier, node coordinates and node coverage area in the node observation record, determine the adjacency and overlap relationships between the node coverage areas to form overlapping node groups; The sampling time is time-aligned based on the node reporting delay, a unified spatial grid is established based on the node coordinates and node coverage, and a unified time window is established based on the time-aligned sampling time. Video detection data and wireless probe data are mapped to a unified spatial grid and a unified time window. The unified spatial grid located at the boundary of the node coverage area or the intersecting area of ​​the node coverage area is filtered to obtain the boundary candidate grid.

[0008] In a preferred embodiment, the target vanishing point, target spawn point, spatial gate crossing direction, and wireless signal continuity interval are extracted from the boundary candidate grid to construct a cross-node event sequence, specifically: Within the candidate boundary grid, the grid boundary corresponding to the boundary where the node coverage area is adjacent or the area where the node coverage area intersects is used as the spatial gate line; Based on video detection data, the target vanishing point is generated at the termination position of the relative spatial gate line within a unified time window; The target occurrence point is generated based on the starting position of the relative spatial gate line within a unified time window, according to the video detection data. The spatial gate line crossing direction is generated based on the positional changes of the target vanishing point and the target appearing point relative to the spatial gate line. The wireless signal continuity interval is generated based on the consecutive occurrence of wireless probe data within adjacent unified time windows. By associating the sampling time with the target disappearance point, target appearance point, spatial gate crossing direction, and wireless signal continuity interval, a cross-node event sequence is constructed.

[0009] In a preferred embodiment, the spatial gate line crossing direction is generated based on the positional changes of the target vanishing point and the target appearing point relative to the spatial gate line, specifically: Based on the positional relationship between the vanishing point and the appearing point of the target on both sides of the spatial gate line, determine the departure side corresponding to the vanishing point and the entry side corresponding to the appearing point of the target; The direction from the departure side to the entry side is taken as the initial crossing direction. The initial crossing direction is verified according to the order of the sampling time corresponding to the target vanishing point and the sampling time corresponding to the target appearing point, and the spatial gate crossing direction is generated.

[0010] In a preferred embodiment, dynamic time warping is performed based on the entry and exit time phase difference, the consistency of the spatial gate crossing direction, and the wireless signal continuity interval in the cross-node event sequence to obtain the cross-node association confidence level, specifically: Pair the target vanishing point and target appearing point located at adjacent edge computing nodes in the cross-node event sequence to form candidate cross-node event pairs; The entry and exit time phase difference is generated based on the difference between the sampling time corresponding to the target disappearance point and the sampling time corresponding to the target appearance point in the candidate cross-node event pair. Compare the spatial gate crossing directions in candidate cross-node event pairs to generate spatial gate crossing direction consistency; Find the wireless signal duration interval that coincides with the sampling time of the candidate cross-node event pair, limit the matching range of dynamic time warping by the wireless signal duration interval, and perform dynamic time warping based on the phase difference between entry and exit time and the consistency of the spatial gate crossing direction to obtain the cross-node association confidence.

[0011] In a preferred embodiment, dynamic time warping is performed based on the phase difference between entry and exit times and the consistency of the spatial gate crossing direction to obtain the cross-node association confidence score, specifically: Arrange candidate cross-node event pairs according to sampling time to form an event sequence to be normalized; Within the matching range defined by the wireless signal continuation interval, dynamic time warping is performed on the event sequence to be warped, using the phase difference between entry and exit times as the time axis constraint and the consistency of the spatial gate crossing direction as the path constraint. Cross-node association confidence is generated based on the minimum cumulative distance obtained by dynamic time warping.

[0012] In a preferred embodiment, based on the cross-node association confidence level to label repeating targets, split targets, and independent targets, the corrected pedestrian density, corrected pedestrian flow direction, and flow conservation residuals are generated, specifically as follows: Based on the cross-node association confidence, candidate cross-node event pairs are merged, and candidate cross-node event pairs that correspond to the same target vanishing point or the same target appearing point are marked as duplicate targets. Mark the vanishing point or the appearing point of a target that has a wireless signal continuity range and has not completed cross-node pairing as a split target; Candidate cross-node event pairs that are not classified as duplicate or split targets are marked as independent targets; Based on the merging results of duplicate targets, the supplementation results of fragmented targets, and the retention results of independent targets, the corrected pedestrian density and corrected pedestrian flow direction are generated. The flow conservation residual is generated based on the change of the corrected pedestrian density within the candidate boundary grid over a unified time window, as well as the inflow and outflow of the corrected pedestrian flow direction.

[0013] In a preferred embodiment, a flow conservation residual is generated based on the change in the corrected pedestrian density within the boundary candidate grid over a uniform time window, as well as the inflow and outflow of the corrected pedestrian flow direction. Specifically: Within each boundary candidate grid, the change in corrected pedestrian density with respect to the unified time window is generated based on the difference in corrected pedestrian density between adjacent unified time windows. Based on the corrected flow direction, the inflow into the boundary candidate grid and the outflow from the boundary candidate grid are statistically analyzed; The flow conservation residual is generated based on the change in the corrected pedestrian density over a uniform time window, the difference between the inflow and outflow, and the flow rate.

[0014] In a preferred embodiment, the active boundary of the business district is updated based on the corrected pedestrian density, corrected pedestrian flow direction, flow conservation residual, and connectivity of the candidate boundary grids. Specifically: Based on the continuous distribution of the corrected pedestrian density within a unified time window, active candidate grids are extracted from the boundary candidate grids; The flow direction connection relationship between active candidate grids is determined based on the corrected flow direction of people; Based on the flow conservation residual, active candidate grids that do not match the changes in the corrected pedestrian flow direction, the corrected pedestrian flow direction, and the corrected pedestrian flow density over a uniform time window are removed from the active candidate grids. Connectivity merging is performed based on the connectivity relationships of the retained active candidate grids in the boundary candidate grids, and the active boundary of the business district is updated based on the flow direction connectivity.

[0015] The technical effects and advantages of the real-time commercial area pedestrian flow analysis and boundary calibration system based on edge computing of this invention are as follows: By acquiring node observation records from multiple edge computing nodes, pedestrian-related data collected by different nodes have a unified data source. Overlapping node groups are divided according to the adjacency and overlap relationships of node coverage areas, and node observation records are mapped to a unified spatial grid and a unified time window, enabling data from different edge computing nodes to be compared under the same spatial and temporal reference. Target vanishing points, target appearing points, spatial gate crossing directions, and wireless signal continuity intervals are extracted from the boundary candidate grid, forming a cross-node event sequence reflecting the cross-node movement process. Cross-node association confidence is obtained through dynamic time warping, which can determine the target association relationship in areas of asynchronous sampling and node boundaries. By marking duplicate targets, fragmented targets, and independent targets, corrected pedestrian density, corrected pedestrian flow direction, and flow conservation residuals are generated, reducing duplicate and fragmented statistics in boundary areas. The active boundary of the business district is updated by combining the corrected pedestrian data, flow conservation residuals, and the connectivity of the boundary candidate grid, allowing the active boundary of the business district to be calibrated according to the actual pedestrian flow distribution. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the structure of the real-time commercial area pedestrian flow analysis and boundary calibration system based on edge computing of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention. Example 1

[0018] Figure 1 This invention presents a real-time commercial area pedestrian flow analysis and boundary calibration system based on edge computing, comprising: Node recording module: Acquires node observation records of multiple edge computing nodes; Grid candidate module: Divide overlapping node groups according to the adjacency and overlap of node coverage areas, map node observation records to a unified spatial grid and a unified time window, and obtain boundary candidate grids; Cross-point event module: Extracts the target vanishing point, target appearing point, spatial gate crossing direction, and wireless signal continuation interval from the boundary candidate grid to construct a cross-node event sequence; The association confidence module performs dynamic time warping based on the entry and exit time phase difference, spatial gate crossing direction consistency, and wireless signal continuity interval in the cross-node event sequence to obtain the cross-node association confidence level. Pedestrian flow correction module: Based on the cross-node association confidence level, it labels repeated targets, split targets, and independent targets, and generates corrected pedestrian flow density, corrected pedestrian flow direction, and flow conservation residuals. Boundary Update Module: Updates the active boundary of the business district based on the corrected pedestrian density, corrected pedestrian flow direction, flow conservation residual, and the connectivity of the boundary candidate grid.

[0019] Obtain node observation records from multiple edge computing nodes, including: Multiple edge computing nodes are pre-deployed at the entrances and exits of the business district, intersections of commercial streets, entrances and exits of underground passages, entrances to indoor and outdoor corridors, and near the open boundaries of the business district. Edge computing node identifiers are used to distinguish different edge computing nodes; these identifiers can be formed by combining region codes, installation location codes, and device sequence codes. Edge computing node identifiers are only used to classify video detection data, wireless probe data, node coordinates, sampling time, node reporting latency, and node coverage area as belonging to the same edge computing node.

[0020] The node coverage area is the spatial region where the corresponding edge computing node can stably acquire video streams and wireless probe signals. The node coverage area can be determined jointly by the field-of-view boundary of the video image and the receptive boundary of the wireless probe signal acquired during the installation and commissioning phase. The field-of-view boundary of the video image is obtained by setting several calibration points on the ground in the commercial area and recording the pixel positions of these points in the video image. The receptive boundary of the wireless probe signal is obtained by moving a test terminal along the path surrounding the edge computing node and recording the wireless probe signal strength. When both the ground projection area corresponding to the field-of-view boundary of the video image and the ground area corresponding to the receptive boundary of the wireless probe signal meet the stable acquisition conditions, the overlapping area of ​​the ground projection area and the ground area is determined as the node coverage area. When the wireless probe signal is only used as auxiliary observation data, the ground projection area corresponding to the field-of-view boundary of the video image is used as the node coverage area, and the receptive boundary of the wireless probe signal is used as the auxiliary verification range of the node coverage area. The node coverage area can be represented as a polygonal region formed by connecting multiple coordinate points sequentially, or as an irregular region centered on the edge computing node's installation location and corrected for occlusion areas. The node coordinates are the coordinates of the edge computing node installation location in the business district spatial coordinates, which can be obtained through surveying and mapping, indoor positioning, or annotation on installation drawings.

[0021] When multiple edge computing nodes acquire video streams within their coverage area, the video streams are input to the edge computing nodes as consecutive image frames. After acquiring the video stream, the edge computing nodes timestamp the stream, ensuring each frame has a corresponding sampling time, and then perform image preprocessing. Image preprocessing includes image size unification, image distortion correction, image brightness normalization, and invalid region occlusion. Invalid regions include walls, sky, shop signs, fixed advertising screens, and static backgrounds not belonging to pedestrian traffic areas. After image preprocessing, the edge computing nodes identify pedestrian targets in the video stream and generate video detection data for them. The video detection data includes the edge computing node identifier, sampling time, pedestrian target number, pedestrian target's bounding rectangle, pedestrian target's center position, pedestrian target's direction of movement, and the pedestrian target's location within the node's coverage area. The pedestrian target number is used to distinguish different targets within consecutive video frames of the same edge computing node; however, the pedestrian target number is not directly used as the basis for determining the same target when crossing different edge computing nodes.

[0022] When extracting video detection data from the video stream, the edge computing node determines the bounding rectangle of the pedestrian target in a single frame image, and then determines the pedestrian target's movement direction based on the change in the pedestrian target's center position at adjacent sampling times. The pedestrian target's center position is determined by the center point of the bounding rectangle, and the position of the pedestrian target within the node's coverage area is calculated through the mapping relationship between video frame coordinates and commercial area spatial coordinates. This mapping relationship is established during the installation and debugging phase using ground calibration points. The commercial area spatial coordinates and video frame coordinates of the ground calibration points are stored in pairs, and the edge computing node calculates the coordinate transformation relationship based on these pairs. After extracting the video detection data, the edge computing node binds the video detection data with the edge computing node identifier and sampling time, enabling the video detection data to be mapped in subsequent steps according to the edge computing node identifier, a unified spatial grid, and a unified time window.

[0023] When multiple edge computing nodes collect wireless probe signals within their coverage area, these signals originate from wireless probe frames or connection management frames emitted by mobile terminals. After receiving the wireless probe signals, each edge computing node records the reception time, signal strength, the corresponding temporary terminal identifier, and the edge computing node identifier. The temporary terminal identifier can be obtained by irreversibly hashing the terminal address carried in the wireless probe signal. The input to the irreversible hash is the terminal address and sampling date carried in the wireless probe signal, and the output is a fixed-length string used to determine consecutive occurrences within the same sampling date. The wireless probe data includes the edge computing node identifier, sampling time, temporary terminal identifier, signal strength, and the number of times the wireless probe signal is received. The number of times the same temporary terminal identifier is received by the same edge computing node within the sampling interval corresponding to the same sampling time.

[0024] When extracting wireless probe data from wireless probe signals, the edge computing node discards wireless probe signals lacking reception time or signal strength, and then aggregates the remaining wireless probe signals according to the terminal temporary identifier and sampling time. The sampling time can be recorded using the edge computing node's local clock. To enable the association between video detection data and wireless probe data in the same node's observation record, the edge computing node converts the sampling times of both video detection data and wireless probe data to the same time format. The time format can be a timestamp composed of year, month, day, hour, minute, second, and millisecond. After generating wireless probe data, the edge computing node binds the wireless probe data with the edge computing node identifier and sampling time, enabling the wireless probe data to be associated with video detection data generated by the same edge computing node near the same sampling time.

[0025] When reading node coordinates and coverage area, the node coordinates and coverage area corresponding to the edge computing node identifier are read. To avoid configuration errors in node coordinates and coverage area, the edge computing node identifier, node coordinates, and node coverage area are combined to form the node's static information. Before associating the node static information with video detection data and wireless probe data, it is necessary to check whether the edge computing node identifier is consistent. If the edge computing node identifier in the node static information is consistent with the edge computing node identifier in the video detection data, and also consistent with the edge computing node identifier in the wireless probe data, then association is allowed.

[0026] Edge computing nodes retain separate video sampling timestamps used when generating video detection data and wireless sampling timestamps used when generating wireless probe data. To create a unified node observation record, edge computing nodes write the video sampling timestamps and wireless sampling timestamps into the same sampling timestamp field group. This field group includes the video sampling timestamp, the wireless sampling timestamp, and the sampling timestamp source. The sampling timestamp source is used to distinguish whether the sampling timestamps originate from the video stream, the wireless probe signal, or a common sampling interval. When both video detection data and wireless probe data exist within the same acquisition interval, the sampling timestamp can be the start, end, or center time of the corresponding acquisition interval.

[0027] When calculating node reporting latency based on data transmission and reception times, the edge computing node writes the data transmission time when sending a data packet containing video detection data and wireless probe data, and writes the data reception time when receiving the data packet. The node reporting latency is the time difference between the data reception time and the data transmission time. To reduce the impact of single network fluctuations on node reporting latency, the data transmission and reception times of multiple data packets can be collected within a statistical period, and the node reporting latency can be generated based on the median of multiple time differences. The length of the statistical period can be set according to the data reporting frequency of the edge computing nodes, for example, it can be set to 30 seconds, 60 seconds, or 120 seconds.

[0028] Using the edge computing node identifier as the first association field, video detection data, wireless probe data, node coordinates, and node coverage area corresponding to the same edge computing node identifier are grouped into the same record to be generated. Then, using the sampling time as the second association field, video detection data and wireless probe data with adjacent times under the same edge computing node identifier are merged into the same time location. Node reporting delay is written into the same record to be generated. If multiple pedestrian targets and multiple terminal temporary identifiers exist near the same edge computing node identifier at the same sampling time, multiple video detection data entries and multiple wireless probe data entries are retained in the same node observation record. These multiple video detection data entries and multiple wireless probe data entries are not bound one-to-one to avoid prematurely identifying video targets and terminal temporary identifiers as the same object. The node observation record includes at least the edge computing node identifier, video detection data, wireless probe data, node coordinates, sampling time, node reporting delay, and node coverage area.

[0029] Overlapping node groups are formed based on the adjacency and overlap relationships of node coverage areas. Node observation records are mapped to a unified spatial grid and a unified time window to obtain candidate boundary grids, including: The process reads the edge computing node identifier, node coordinates, and node coverage area from the node observation records, and assigns the node coverage area to the corresponding edge computing node according to the edge computing node identifier. Each node coverage area is converted into a planar region represented by boundary line segments and boundary vertices, and the node coordinates are used as the reference position for the corresponding node coverage area. The conversion process preserves the original boundary coordinates of the node coverage area. The conversion process generates a closed boundary according to the connection order of multiple coordinate points within the node coverage area. The closed boundary is written into the spatial record corresponding to the edge computing node identifier, enabling the determination of adjacency and overlap relationships between different node coverage areas within the same business district spatial coordinate system.

[0030] When determining the adjacency relationship between node coverage areas, the coverage areas corresponding to two edge calculation nodes are selected sequentially. The shortest distance between the boundary segments of the two node coverage areas is calculated, and the distribution of the boundary vertices of the two node coverage areas is used to determine whether there is an adjacency boundary between them. The shortest distance between boundary segments uses the coordinates of the endpoints of the boundary segments of the two node coverage areas as input. The unit of the shortest distance is consistent with the length unit of the business district spatial coordinates. When the business district spatial coordinates use meters as the length unit, the unit of the shortest distance is meters. The adjacency distance threshold is used to determine whether two node coverage areas are spatially adjacent. The adjacency distance threshold is determined based on the calibration error of the node coverage areas and the position conversion error of the video detection data. The process for determining the contact distance threshold is as follows: During the installation and commissioning phase, multiple ground calibration points are selected. The distance difference between the actual position of each ground calibration point in the business district spatial coordinates and the converted position in the video frame coordinates is calculated. These distance differences are then sorted by value, and the largest difference is selected as the position conversion error. Next, the coverage area of ​​the same node is calibrated multiple times, and the distance difference between the boundary coordinates obtained from these calibrations is calculated. The largest stable value among these distance differences is selected as the calibration error of the node's coverage area. The position conversion error is added to the calibration error of the node's coverage area to obtain the base value of the contact distance threshold. An upper limit is then imposed on this base value based on the minimum passable width of the business district's passageway. The contact distance threshold can be determined based on installation and commissioning samples; for example, it can be set to between 0.5 meters and 2 meters. If the shortest distance between the boundary segments of the coverage areas of two nodes is not greater than the adjacency distance threshold, and the boundary of the impassable obstruction area written during installation and debugging is not located between the coverage areas of the two nodes, then the coverage areas of the two nodes are determined to be adjacent, and the boundary segments between the coverage areas of the two nodes that meet the adjacency distance threshold are recorded as the adjacency boundaries of the node coverage areas. The boundary of the impassable obstruction area can be formed by the boundary coordinates of fixed walls, enclosed railings, enclosed green belts, and fixed building exterior walls on the business district ground map.

[0031] When determining the overlap between node coverage areas, two node coverage areas that have already been determined to be adjacent are selected sequentially, and the intersection area of ​​the two node coverage areas in the same business district spatial coordinates is calculated. The intersection area is formed by the intersection point of the boundary segments of the two node coverage areas, the boundary vertex falling into the middle of the other node's coverage area, and the overlapping boundary segments. When calculating the area of ​​the intersection area, the boundaries of the intersection area are sorted into closed polygons in a clockwise or counterclockwise direction, and then the area is calculated based on the coordinates of the boundary vertices of the closed polygons. The unit of the intersection area area is consistent with the unit of area in the business district spatial coordinates. When the business district spatial coordinates use meters as the unit of length, the unit of the intersection area area is square meters. The overlap area threshold is used to exclude extremely small intersection areas caused by the calibration deviation of the node coverage area boundaries. The process for determining the overlap area threshold is as follows: During the installation and commissioning phase, the coverage area of ​​the same node is calibrated multiple times, and the area difference formed between the calibrated boundaries is calculated. These area differences are then sorted by size, and the largest difference is selected as the base value for the overlap area threshold. If the unified spatial grid already has a preset grid side length, the area of ​​a single unified spatial grid can also be used as the lower limit constraint for the overlap area threshold. The overlap area threshold can be determined based on the calibration accuracy of the node coverage area; for example, it can be set to 1 to 3 times the area of ​​a single unified spatial grid. If the area of ​​the intersection of two node coverage areas is greater than the overlap area threshold, then the two node coverage areas are considered to have an overlap relationship, and the spatial area jointly covered by the two node coverage areas is recorded as the intersection area. If the area of ​​the intersection of two node coverage areas is not greater than the overlap area threshold, but the shortest distance between the boundary segments of the two node coverage areas is not greater than the adjacency distance threshold, then only the adjacent relationship is recorded, and the overlap relationship is not recorded.

[0032] Using edge computing node identifiers as the grouping object, edge computing node identifiers with adjacent or overlapping relationships are grouped into the same overlapping node group. If the node coverage area corresponding to one edge computing node is adjacent or overlapping with the node coverage areas corresponding to multiple edge computing nodes, then the multiple edge computing node identifiers are grouped into the same overlapping node group based on the spatial connectivity formed by the adjacent and overlapping relationships. The overlapping node group includes the overlapping node group identifier, the edge computing node identifiers participating in the overlapping node group, the boundary of the node coverage areas, the intersecting area of ​​the node coverage areas, and the reading position of the corresponding node observation record. The overlapping node group identifier can be formed by concatenating the edge computing node identifiers participating in the overlapping node group in lexicographical order. For example, when the edge computing node identifiers participating in the overlapping node group are A01, A02, and A03, the overlapping node group identifier can be generated as A01-A02-A03.

[0033] When aligning sampling times based on node reporting delay, the sampling time and node reporting delay are read from the node observation record, and delay compensation is applied to the sampling time. The node reporting delay is the time difference between the data reception time and the data transmission time, and the unit of the node reporting delay is seconds or milliseconds. The sampling times corresponding to the same edge computing node identifier are converted to the same time base, and the sampling time is added to the node reporting delay to obtain the time-aligned sampling time. Alternatively, when uniformly recording sampling events at the receiving location, the time recorded at the receiving location is subtracted from the node reporting delay to obtain the time-aligned sampling time. Whether to use addition or subtraction depends on the source of the sampling time field in the node observation record; when the sampling time field indicates the time when edge computing node data collection occurs, addition is used; when the sampling time field indicates the time when data arrives at the receiving location, subtraction is used. If the same edge computing node identifier generates multiple node reporting delays within the same statistical period, the median of the multiple node reporting delays is taken as the node reporting delay participating in time alignment. If a node reporting delay is negative, or exceeds the statistical period length, the corresponding node reporting delay will be marked as an abnormal node reporting delay, and time alignment will not be performed using the abnormal node reporting delay. The length of the statistical period is determined based on the data reporting frequency of the edge computing nodes. For example, when an edge computing node uploads node observation records once per second, the statistical period can be set to 30 to 120 seconds.

[0034] When establishing a unified spatial grid based on node coordinates and node coverage, the business district analysis area is determined. The business district analysis area is formed by the outer region of all node coverage, or by the combined area of ​​the business district management boundary and all node coverage. Within the business district analysis area, continuous spatial regions are divided according to fixed grid side lengths, forming multiple unified spatial grids. The grid side length of the unified spatial grid is determined jointly based on the position conversion error of the video detection data, the calibration error of the node coverage, and the spatial stability scale of the wireless probe data. The spatial stability scale of the wireless probe data is obtained by placing mobile test terminals at known locations during the installation and debugging phase, recording the signal strength of the mobile test terminals when received by edge computing nodes at different locations, and statistically analyzing the spatial distance range corresponding to the stable signal strength intervals. This spatial distance range is used as the spatial stability scale of the wireless probe data. The position conversion error, the node coverage calibration error, and the spatial stability scale of the wireless probe data are compared, and a value not less than the sum of the position conversion error and the node coverage calibration error, and not greater than the typical width of the business district passage area, is selected as the grid side length. The grid side length can be determined based on the installation and debugging sample; for example, the grid side length can be set to 1 meter, 2 meters, or 5 meters. Each unified spatial grid records the unified spatial grid number, unified spatial grid boundary coordinates, unified spatial grid center coordinates, and the business district analysis area to which the unified spatial grid belongs.

[0035] When establishing a unified time window based on time-aligned sampling times, the time-aligned sampling times are divided into continuous and non-overlapping time intervals. The length of the unified time window is determined jointly based on the sampling frequency of video detection data, the occurrence interval of wireless probe data, and the fluctuation range of node reporting latency. The process for determining the length of the unified time window is as follows: The distribution of time differences between the occurrence times of video detection data and wireless probe data within the same node's coverage area is statistically analyzed, and a time length that covers the main time difference distribution is selected as the base time length; the difference between the maximum and minimum values ​​of node reporting latency within the statistical period is statistically analyzed, and this difference is used as the fluctuation range of node reporting latency; the base time length and the fluctuation range of node reporting latency are added together to obtain the candidate length of the unified time window; when the candidate length is greater than the shortest resolvable time interval between adjacent pedestrian passage events within the business district, the shortest resolvable time interval is used as the upper limit of the unified time window length. The length of the unified time window can be determined based on sample data; for example, the length of the unified time window can be set to 1 second, 3 seconds, or 5 seconds. Each unified time window records the unified time window number, the start time of the unified time window, and the end time of the unified time window.

[0036] When mapping video detection data to a unified spatial grid and a unified time window, the following information is read from the video detection data: edge computing node identifier, sampling time, pedestrian target number, pedestrian target center position, pedestrian target movement direction, and the position of the pedestrian target within the node's coverage area. The unified spatial grid into which the position within the pedestrian target's node coverage area falls is used as the unified spatial grid corresponding to the video detection data. The unified time window into which the sampling time of the time-aligned video detection data falls is used as the unified time window corresponding to the video detection data. After mapping, a video mapping record is generated, which includes the edge computing node identifier, video detection data, unified spatial grid number, and unified time window number. If the position within the pedestrian target's node coverage area lies on the boundary line of two unified spatial grids, the distance between the pedestrian target's center position and the center coordinates of the adjacent unified spatial grids is calculated, and the unified spatial grid with the smaller distance is used as the unified spatial grid corresponding to the video detection data; if the distances are the same, the unified spatial grid in which the pedestrian target's movement direction points is used as the unified spatial grid corresponding to the video detection data. If the video detection data lacks the location of the pedestrian target within the coverage area of ​​the node, the corresponding video detection data will be marked as unmapped video detection data, and the unmapped video detection data will not be written into the boundary candidate grid.

[0037] When mapping wireless probe data to a unified spatial grid and a unified time window, the edge computing node identifier, sampling time, terminal temporary identifier, signal strength, and number of wireless probe signal receptions are read from the wireless probe data. The wireless probe data is assigned to the node coverage area of ​​the corresponding edge computing node identifier, and the spatial affiliation within the node coverage area is determined by combining the signal strength and the number of wireless probe signal receptions. If the same terminal temporary identifier is received by only one edge computing node within the same unified time window, the wireless probe data is mapped to the unified spatial grid covered by the node coverage area of ​​the corresponding edge computing node, and the occurrence status, signal strength, and number of wireless probe signal receptions are recorded in each unified spatial grid. If the same terminal temporary identifier is received by multiple edge computing nodes within the same unified time window, the signal strength and number of wireless probe signal receptions received by different edge computing nodes are compared, and the wireless probe data is preferentially mapped to the unified spatial grid covered by the node coverage area of ​​the edge computing node with higher signal strength and more wireless probe signal receptions. The signal strength difference threshold is obtained by moving test terminals to known locations and statistically analyzing signal strength differences during the installation and commissioning phase. The differences in signal strength received by multiple edge computing nodes at the same known location are sorted, and the difference that can distinguish between closer and farther edge computing nodes is selected as the signal strength difference threshold. For example, the signal strength difference threshold can be set from 3 dB to 8 dB. After mapping is completed, a wireless mapping record is generated, including the edge computing node identifier, wireless probe data, unified spatial grid number, and unified time window number. If the wireless probe data lacks signal strength or a temporary terminal identifier, the corresponding wireless probe data is marked as unmapped wireless probe data, and unmapped wireless probe data is not written to the boundary candidate grid.

[0038] When filtering candidate boundary grids, the following inputs are used: overlapping node groups, the boundary of node coverage areas, the intersecting region of node coverage areas, unified spatial grids, video mapping records, and wireless mapping records. For edge computing nodes with adjacent relationships, unified spatial grids that intersect with the boundary of the node's coverage area are searched, or unified spatial grids whose distance from the boundary of the node's coverage area is not greater than the intersection distance threshold are searched, and the found unified spatial grids are used as candidate filtering objects. For edge computing nodes with overlapping relationships, unified spatial grids with overlapping areas in the intersecting region of the node's coverage area are searched, and the found unified spatial grids are used as candidate filtering objects. It is checked whether the candidate filtering object contains video mapping records or wireless mapping records identified by at least one edge computing node from the overlapping node group within a unified time window. If the candidate filtering object does not have video mapping records or wireless mapping records in multiple consecutive unified time windows, the candidate filtering object is not used as a candidate boundary grid. The number of consecutive unified time windows is determined based on the length of the unified time window and the time required for pedestrians within the business district to cross the boundary of the node coverage area or the area where the node coverage areas intersect. For example, the number of consecutive unified time windows can be set to 2 to 5 unified time windows.

[0039] The selected unified spatial grids are marked as boundary candidate grids, and the following information is saved for each boundary candidate grid: boundary candidate grid number, unified spatial grid number, unified time window number, associated overlapping node group identifier, corresponding node coverage boundary or node coverage intersection area, video mapping record, and wireless mapping record. Boundary candidate grids represent spatial grids where cross-node event analysis is required at the adjacent or overlapping node coverage areas of multiple edge computing nodes.

[0040] Extract the target vanishing point, target spawn point, spatial gate crossing direction, and wireless signal continuity interval from the candidate boundary grid to construct a cross-node event sequence, including: The system reads the boundary candidate grid numbers, unified spatial grid numbers, unified time window numbers, overlapping node group identifiers, node coverage boundary boundaries, node coverage intersection areas, video mapping records, and wireless mapping records stored in the boundary candidate grid. Video mapping records include edge computing node identifiers, video detection data, unified spatial grid numbers, and unified time window numbers; wireless mapping records include edge computing node identifiers, wireless probe data, unified spatial grid numbers, and unified time window numbers. Based on the overlapping node group identifiers, the edge computing node identifiers participating in cross-node event analysis are determined. Then, based on the unified spatial grid numbers and unified time window numbers, the video detection data and wireless probe data within the boundary candidate grid are read, ensuring that the target vanishing point, target appearing point, spatial gate crossing direction, and wireless signal continuity interval all originate from the same spatial and temporal range corresponding to the same boundary candidate grid.

[0041] Within the candidate boundary grid, the grid boundary corresponding to the boundary where the node coverage ranges meet or the area where the node coverage ranges intersect is used as the spatial gate. The spatial gate represents the boundary line segment used to distinguish the coverage ranges of adjacent edge computing nodes. The spatial gate is a set of unified spatial grid boundary coordinates, and the unit of the spatial gate coordinates is consistent with the length unit of the business district spatial coordinates. When the business district spatial coordinates use meters as the length unit, the unit of the spatial gate coordinates is meters. For candidate boundary grids formed by the boundaries where the node coverage ranges meet, the unified spatial grid boundary that intersects with the boundary where the node coverage ranges meet, or the unified spatial grid boundary whose distance from the boundary where the node coverage ranges meet is not greater than the intersection distance threshold, is used as the spatial gate. For candidate boundary grids formed by the area where the node coverage ranges intersect, the intersection line between the area where the node coverage ranges intersect and the unified spatial grid boundary is found, and the unified spatial grid boundary where the intersection line is located is used as the spatial gate. If multiple unified spatial grid boundaries exist for the same candidate boundary, the unified spatial grid boundary that simultaneously borders the coverage areas of two edge computing nodes is selected. If multiple optional unified spatial grid boundaries simultaneously border the coverage areas of two edge computing nodes, the unified spatial grid boundary with the largest overlap length with the node coverage area or the largest overlap length of the intersecting region with the node coverage area is selected. After the spatial gate is generated, the coordinates of the unified spatial grid boundary corresponding to the spatial gate, the identifier of the edge computing node associated with the spatial gate, the unified spatial grid number associated with the spatial gate, and the unified time window number associated with the spatial gate are recorded.

[0042] When generating the vanishing point of a target based on the termination position of the spatial gate line relative to the video detection data within a unified time window, continuous video detection data is read according to the edge computing node identifier and pedestrian target number within the unified time window corresponding to the boundary candidate grid. The position of the pedestrian target within the coverage area of ​​the node in the video detection data has been converted to the business district spatial coordinates, and the position within the coverage area of ​​the node where the pedestrian target is located is used as the target position. For the same edge computing node identifier and the same pedestrian target number, the target positions are arranged according to the sampling time, and the vertical distance from each target position to the spatial gate line is calculated. The proximity range of the spatial gate line is used to determine whether the target position is close to the spatial gate line. The width of the proximity range of the spatial gate line is determined based on the position conversion error of the video detection data and the grid side length of the unified spatial grid. The width of the proximity range of the spatial gate line is a value that is not less than the position conversion error of the video detection data and not greater than the grid side length of the unified spatial grid. For example, the width of the proximity range of the spatial gate line can be set to 0.5 meters to 2 meters. If the target location corresponding to the same edge computing node identifier and the same pedestrian target number enters the vicinity of the spatial gate line within a unified time window, and the target location does not appear in the video detection data corresponding to the same edge computing node identifier after its last appearance within the vicinity of the spatial gate line, the last appearance of the target location is taken as the termination location. The termination location, the sampling time corresponding to the termination location, the edge computing node identifier, the pedestrian target number, the unified spatial grid boundary coordinates corresponding to the spatial gate line, and the unified time window number are recorded together to generate the target vanishing point.

[0043] If the same pedestrian target ID continues to appear in the video detection data corresponding to the same edge computing node after the termination position, no target vanishing point is generated. If the same pedestrian target ID is temporarily missing due to occlusion in the video detection data, the time difference between the last sampling time before the missing pedestrian target ID and the first sampling time after the missing is calculated, with the unit of the time difference consistent with the unit of the sampling time. The preset missing tolerance time is determined based on the video stream frame rate and the time required for pedestrians to cross the spatial gate in the business district. The lower the video stream frame rate, the longer the preset missing tolerance time; the wider the range near the spatial gate, the longer the preset missing tolerance time. For example, the preset missing tolerance time can be set to 0.5 seconds to 2 seconds. If the time difference is less than the preset missing tolerance time, no target vanishing point is generated; if the time difference reaches the preset missing tolerance time, and the last target position before the missing is within the range near the spatial gate, the last target position before the missing is taken as the termination position and a target vanishing point is generated. If the video detection data lacks a pedestrian target number or the location of the pedestrian target within the coverage area of ​​the node, the corresponding video detection data will be marked as video detection data that did not participate in the generation of the target vanishing point.

[0044] When generating target occurrence points based on the starting position relative to the spatial gate line within a unified time window using video detection data, continuous video detection data is read according to the edge computing node identifier and pedestrian target number within the unified time window corresponding to the boundary candidate grid, and the target positions are arranged according to the sampling time. For the same edge computing node identifier and the same pedestrian target number, if the target position first appears within the vicinity of the spatial gate line, and the target position moves from the vicinity of the spatial gate line to the center of the node coverage area at subsequent sampling times, the position where the target position first appears within the vicinity of the spatial gate line is taken as the starting position. The starting position, the sampling time corresponding to the starting position, the edge computing node identifier, the pedestrian target number, the unified spatial grid boundary coordinates corresponding to the spatial gate line, and the unified time window number are recorded together to generate the target occurrence point. If the same pedestrian target number has already appeared continuously in the video detection data corresponding to the same edge computing node identifier before the starting position, no target occurrence point is generated. If the video detection data contains a single pedestrian target bounding rectangle, it is checked whether the single pedestrian target bounding rectangle forms a continuous target position at subsequent sampling times. The criteria for determining continuous target locations are that the same pedestrian target number has adjacent target locations within at least two adjacent sampling times, and the distance between adjacent target locations is no greater than the product of the maximum walking speed of pedestrians in the business district and the interval between adjacent sampling times. The maximum walking speed of pedestrians in the business district can be calculated based on the change in the center position of pedestrian targets in sample videos collected during the installation and debugging phase and the sampling time interval. For example, the maximum walking speed of pedestrians in the business district can be set to 2 to 3 meters per second. No target appearance point is generated when no continuous target locations are formed.

[0045] When generating the spatial gate line crossing direction based on the positional changes of the target vanishing point and the target appearing point relative to the spatial gate line, the spatial side identification on both sides of the spatial gate line is first determined. The spatial side identification is the spatial region marker where the target position is located relative to the spatial gate line. For spatial gate lines corresponding to the boundaries of node coverage areas, the two sides of the spatial gate line are marked as the node coverage area identification of the corresponding edge computing node. For spatial gate lines corresponding to the intersecting areas of node coverage areas, the intersecting area of ​​node coverage areas, the non-overlapping part of the node coverage area corresponding to one edge computing node, and the non-overlapping part of the node coverage area corresponding to another edge computing node are respectively used as spatial side identification determination areas. Based on the positional relationship between the target vanishing point and the target appearing point on both sides of the spatial gate line, the departure side corresponding to the target vanishing point and the entry side corresponding to the target appearing point are determined. The departure side is the spatial side identification where the target vanishing point is located, and the entry side is the spatial side identification where the target appearing point is located. If the target vanishing point is located on the spatial gate line itself, the departure side is determined based on the target position at the sampling time one step before the target vanishing point. If the target appearing point is located on the spatial gate line itself, the entry side is determined based on the target position at the sampling time one step after the target appearing point. If the target position cannot be obtained at the sampling time before the target vanishing point, or the target position cannot be obtained at the sampling time after the target appearing point, the corresponding target vanishing point or target appearing point is marked as a spatial side undetermined event, and the spatial gate line crossing direction is not generated using the spatial side undetermined event.

[0046] The initial crossing direction is taken as the direction from the departure side to the entry side. This initial crossing direction is then validated based on the temporal relationship between the sampling times corresponding to the target vanishing point and the target appearing point, generating a spatial gate crossing direction. The input for temporal relationship validation is the sampling times corresponding to the target vanishing point, the target appearing point, and the time-aligned sampling times. If the sampling time corresponding to the target vanishing point is earlier than the sampling time corresponding to the target appearing point, and the departure side corresponding to the target vanishing point and the entry side corresponding to the target appearing point are located on opposite sides of the spatial gate, the initial crossing direction is confirmed as the spatial gate crossing direction. If the sampling time corresponding to the target appearing point is earlier than the sampling time corresponding to the target vanishing point, the initial crossing direction is reversed to form the direction to be validated, and it is checked whether the direction to be validated is consistent with the target position's movement direction at adjacent sampling times. The target position's movement direction at adjacent sampling times is provided by the pedestrian target movement direction in the video detection data; if the pedestrian target movement direction in the video detection data is missing, the target position's movement direction at adjacent sampling times is generated based on the target position difference at adjacent sampling times. If the direction to be verified is consistent with the direction of movement of the target position at adjacent sampling times, the direction to be verified is confirmed as the direction of space gate crossing; if the direction to be verified is inconsistent with the direction of movement of the target position at adjacent sampling times, the corresponding target vanishing point and target appearing point are marked as direction undetermined events, and the direction undetermined events are not used to generate the direction of space gate crossing.

[0047] When generating a wireless signal continuation interval based on the continuous occurrence of wireless probe data within adjacent unified time windows, wireless mapping records are read from the unified spatial grid number corresponding to the boundary candidate grid, the current unified time window number, the previous unified time window number, and the next unified time window number. Wireless probe data is then aggregated according to the terminal temporary identifier. Adjacent unified time windows refer to unified time windows that are sequentially connected to the current unified time window. For the same terminal temporary identifier, if wireless probe data appears within at least two adjacent unified time windows, and the edge computing node identifiers corresponding to the wireless probe data belong to the same overlapping node group, the sampling time of the first occurrence is taken as the start time of the wireless signal continuation interval, and the sampling time of the last occurrence is taken as the end time of the wireless signal continuation interval. The wireless signal continuation interval includes the terminal temporary identifier, start time, end time, the unified time window number, the corresponding edge computing node identifier, the signal strength sequence, and the wireless probe signal reception count sequence. The signal strength sequence is arranged according to the sampling time, corresponding to the signal strength of the same terminal temporary identifier, and the wireless probe signal reception count sequence is arranged according to the sampling time, corresponding to the number of wireless probe signals received by the same terminal temporary identifier. If the same terminal temporary identifier appears only in one unified time window, no wireless signal continuity interval will be generated.

[0048] If the same terminal temporary identifier is briefly missing between adjacent unified time windows, a missing tolerance window is set based on the wireless probe signal acquisition interval. The unit of the missing tolerance window is the number of unified time windows. The missing tolerance window is determined based on the time intervals of continuous wireless probe signal transmission from the same test terminal during the installation and commissioning phase. Multiple time intervals of continuous wireless probe signal transmission from the same test terminal are counted, and the number of unified time windows covering most of these time intervals is selected as the missing tolerance window. For example, the missing tolerance window can be set to one or two unified time windows. If the missing duration is no greater than the missing tolerance window, the time periods before and after the missing event are merged into the same wireless signal continuation interval; if the missing duration is greater than the missing tolerance window, the time periods before and after the missing event are generated as different wireless signal continuation intervals. If the wireless probe data lacks a terminal temporary identifier, sampling time, or edge computing node identifier, the corresponding wireless probe data is marked as wireless probe data that did not participate in the generation of the wireless signal continuation interval.

[0049] A cross-node event sequence is constructed by associating the target vanishing point, target appearing point, spatial gate crossing direction, and wireless signal continuation interval with the sampling time. The cross-node event sequence is a set of events arranged according to the sampling time. Each event in the cross-node event sequence includes the event type, sampling time, edge computing node identifier, unified spatial grid number, unified time window number, associated spatial gate, and data source. Event types include target vanishing point events, target appearing point events, spatial gate crossing direction events, and wireless signal continuation interval events. The target vanishing point is written into the cross-node event sequence according to its corresponding sampling time; the target appearing point is written into the cross-node event sequence according to its corresponding sampling time; the spatial gate crossing direction is written into the cross-node event sequence according to the time position between the sampling times corresponding to the target vanishing point and the target appearing point; and the wireless signal continuation interval is written into the cross-node event sequence according to its start and end times. If the target vanishing point, target appearing point, and wireless signal continuation interval overlap in time, the wireless signal continuation interval is written into the cross-node event sequence as a wireless continuity reference event between the target vanishing point and the target appearing point. If multiple events have the same sampling time, they are written into the cross-node event sequence in a fixed order: target vanishing point event, target appearing point event, spatial gate crossing direction event, and wireless signal continuation interval event.

[0050] Based on the entry and exit time phase difference, spatial gate crossing direction consistency, and wireless signal duration interval in the cross-node event sequence, dynamic time warping is performed to obtain the cross-node association confidence, including: Read the boundary candidate grid number, overlapping node group identifier, unified spatial grid number, unified time window number, spatial gate line, target vanishing point, target appearing point, spatial gate line crossing direction, and wireless signal continuity interval from the cross-node event sequence. Based on the overlapping node group identifier, read the edge computing node identifiers participating in the overlapping node group, and determine adjacent edge computing nodes based on the edge computing node identifiers associated with the spatial gate line. Adjacent edge computing nodes are edge computing nodes whose coverage areas are adjacent or overlapping on both sides of the same spatial gate line.

[0051] When pairing target vanishing points and target appearing points located at adjacent edge computing nodes in a cross-node event sequence, the target vanishing points and target appearing points corresponding to the same spatial gate are arranged according to the sampling time. The target vanishing point of the edge computing node on one side of the spatial gate is selected, and the target appearing point of the edge computing node on the other side of the spatial gate is searched. If the time interval between the sampling time corresponding to the target vanishing point and the sampling time corresponding to the target appearing point falls within the candidate pairing time range, and the target vanishing point and target appearing point correspond to the same spatial gate, then the target vanishing point and target appearing point are recorded as a candidate cross-node event pair. The candidate pairing time range is determined based on the width of the adjacent range of the spatial gate, the grid side length of the unified spatial grid, the maximum walking speed of pedestrians in the business district, and the interval between adjacent sampling times. The lower limit of the candidate pairing time range is the interval between adjacent sampling times, and the upper limit of the candidate pairing time range is the time length obtained by dividing the sum of the width of the adjacent range of the spatial gate and the grid side length of the unified spatial grid by the maximum walking speed of pedestrians in the business district. This time length is then compared with the length of the unified time window, and the larger value is selected as the upper limit of the candidate pairing time range. The unit of the candidate pairing time range is consistent with the time unit of the sampling time. For example, when the width of the adjacent range of the spatial gate line is 1 meter, the grid side length of the unified spatial grid is 2 meters, the maximum walking speed of pedestrians in the business district is 2 meters per second, and the length of the unified time window is 3 seconds, the candidate pairing time range can be set to an interval of 3 seconds between adjacent sampling times.

[0052] An entry / exit time phase difference is generated based on the difference between the sampling time corresponding to the target's vanishing point and the sampling time corresponding to the target's appearance point in the candidate cross-node event pair. The entry / exit time phase difference represents the time interval between the target leaving the node coverage area of ​​one edge computing node and appearing in the node coverage area of ​​another edge computing node. The entry / exit time phase difference is obtained by subtracting the sampling time corresponding to the target's vanishing point from the sampling time corresponding to the target's appearance point. If the sampling time is expressed in seconds, the unit of the entry / exit time phase difference is seconds; if the sampling time is expressed in milliseconds, the unit of the entry / exit time phase difference is milliseconds. When the entry / exit time phase difference is positive, it indicates that the sampling time corresponding to the target's appearance point is later than the sampling time corresponding to the target's vanishing point; when the entry / exit time phase difference is negative, it indicates that the sampling time corresponding to the target's appearance point is earlier than the sampling time corresponding to the target's vanishing point. The value, unit, and sign of the entry / exit time phase difference are written into the candidate cross-node event pair.

[0053] Compare the spatial gate line crossing directions in candidate cross-node event pairs to generate spatial gate line crossing direction consistency. Spatial gate line crossing direction consistency represents the correspondence between the candidate crossing directions formed from the target vanishing point to the target appearing point in the candidate cross-node event pair and the spatial gate line crossing directions already generated in the cross-node event sequence. Read the departure side corresponding to the target vanishing point and the entry side corresponding to the target appearing point, and form candidate crossing directions from the departure side to the entry side. Then read the spatial gate line crossing directions associated with the same spatial gate line, the same unified time window number, or adjacent unified time window numbers in the cross-node event sequence. If the candidate crossing direction is the same as the spatial gate line crossing direction, record the spatial gate line crossing direction consistency as consistent; if the candidate crossing direction is opposite to the spatial gate line crossing direction, record the spatial gate line crossing direction consistency as opposite; if the target vanishing point or target appearing point is marked as an undetermined spatial side event or an undetermined direction event, record the spatial gate line crossing direction consistency as undetermined. Write the spatial gate line crossing direction consistency into the candidate cross-node event pair, and use the spatial gate line crossing direction consistency as the path constraint input for subsequent dynamic time warping.

[0054] When searching for wireless signal continuation intervals that coincide with the sampling times of candidate cross-node event pairs, the sampling times corresponding to the target vanishing point and the target appearing point in the candidate cross-node event pair are read, and the time interval between these two times is taken as the event pairing time interval. Wireless signal continuation intervals are read from the same boundary candidate grid number, the same overlapping node group identifier, the same unified spatial grid number, or adjacent unified spatial grid numbers. If the start time of the wireless signal continuation interval is not later than the end time of the event pairing time interval, and the end time is not earlier than the start time of the event pairing time interval, the wireless signal continuation interval is determined to coincide with the sampling times of the candidate cross-node event pair. If multiple wireless signal continuation intervals coincide with the event pairing time interval, all coinciding wireless signal continuation intervals are retained and arranged in descending order of coincidence duration. If no coinciding wireless signal continuation intervals exist, the candidate cross-node event pair is marked as a wireless signal continuation interval missing event, and the candidate pairing time range is used as the matching range for dynamic time warping.

[0055] When defining the matching range for dynamic time warping based on the wireless signal duration interval, the matching time range is determined according to the start and end times of the wireless signal duration interval. The start time of the matching time range is the start time of the wireless signal duration interval minus the time spread, and the end time of the matching time range is the end time of the wireless signal duration interval plus the time spread. The time spread is used to absorb sampling deviations between video detection data and wireless probe data. The time spread is determined based on the length of the unified time window, the fluctuation range of node reporting delay, and the wireless probe signal acquisition interval. The length of the unified time window, the fluctuation range of node reporting delay, and the wireless probe signal acquisition interval are compared, and the maximum value is selected as the time spread. The length of the unified time window, the fluctuation range of node reporting delay, and the wireless probe signal acquisition interval are all converted to the same time unit before comparison. For example, if the length of the unified time window is 3 seconds, the fluctuation range of node reporting delay is 0.5 seconds, and the wireless probe signal acquisition interval is 1 second, the time spread can be set to 3 seconds. Only candidate cross-node event pairs whose sampling times corresponding to both the vanishing point and the appearance point of the target are within the matching time range are allowed to enter dynamic time warping.

[0056] Candidate cross-node event pairs are arranged according to their sampling times to form a sequence of events to be normalized. This sequence includes the candidate cross-node event pair identifier, the sampling time corresponding to the target vanishing point, the sampling time corresponding to the target appearing point, the phase difference between entry and exit times, the consistency of the spatial gate crossing direction, the identifier of the associated wireless signal continuation interval, and the boundary candidate grid number. Candidate cross-node event pairs are arranged from earliest to latest according to the sampling time corresponding to the target vanishing point. If two candidate cross-node event pairs have the same sampling time corresponding to the target vanishing point, they are arranged from earliest to latest according to the sampling time corresponding to the target appearing point. If both the sampling time corresponding to the target vanishing point and the sampling time corresponding to the target appearing point are the same, the candidate cross-node event pairs are arranged according to a fixed order of the edge computing node identifiers. Each candidate cross-node event pair in the sequence of events to be normalized retains its corresponding spatial gate, unified spatial grid number, and unified time window number, ensuring that dynamic time normalization only occurs between cross-node events corresponding to the same boundary candidate grid.

[0057] Within the matching range defined by the wireless signal's duration, dynamic time warping is performed on the event sequence to be warped, using the phase difference between entry and exit times as the time axis constraint and the consistency of the spatial gate crossing direction as the path constraint. Dynamic time warping means finding the path with the minimum cumulative distance between two event subsequences in chronological order, even when their sampling times are not completely consistent. The vanishing point events of the target computation nodes on one side of the spatial gate are formed into a first event subsequence according to their sampling times, and the appearing point events of the target computation nodes on the other side of the spatial gate are formed into a second event subsequence according to their sampling times. Both the first and second event subsequences originate from the same boundary candidate grid number, the same overlapping node group identifier, and the same spatial gate. The vanishing point events in the first event subsequence and the appearing point events in the second event subsequence are used as the objects to be matched, and the local distance for each object to be matched is calculated.

[0058] When calculating local distances, first calculate the absolute value of the difference between the absolute value of the phase difference between entry and exit times and the reasonable travel time. Divide this absolute value by the reasonable travel time to obtain the time difference value. The reasonable travel time is determined based on the width of the area adjacent to the spatial gate line, the grid side length of the unified spatial grid, and the maximum walking speed of pedestrians in the business district. The reasonable travel time is the time length obtained by dividing the position interval by the maximum walking speed of pedestrians in the business district. The position interval is the spatial distance between the target vanishing point and the target appearing point. When the spatial distance between the target vanishing point and the target appearing point cannot be obtained, the position interval is the sum of the width of the area adjacent to the spatial gate line and the grid side length of the unified spatial grid. The unit of the reasonable travel time is the same as the unit of the phase difference between entry and exit times. Then, a directional penalty value is generated based on the consistency of the spatial gate line crossing direction. When the consistency of the spatial gate line crossing direction is consistent, the directional penalty value is 0. When the consistency of the spatial gate line crossing direction is undetermined, the directional penalty value is the intermediate penalty value. When the consistency of the spatial gate line crossing direction is opposite, the directional penalty value is the maximum penalty value. The intermediate and maximum penalty values ​​are determined based on the directional error distribution of known cross-node movement events and known non-cross-node movement events in the installation and commissioning sample. For example, the intermediate penalty value can be set to 0.5 to 1.5, and the maximum penalty value can be set to 1 to 3. Then, a wireless continuity penalty value is generated based on whether the wireless signal continuation range covers the object to be matched. When the wireless signal continuation range covers the object to be matched, the wireless continuity penalty value is 0; when the wireless signal continuation range is missing, the wireless continuity penalty value is non-zero. The non-zero penalty value is determined based on the proportion of missing wireless probe signals in the installation and commissioning sample. For example, the non-zero penalty value can be set to 0.5 to 2. The local distance is obtained by adding the time difference value, the directional penalty value, and the wireless continuity penalty value.

[0059] When building the cumulative distance table, the vanishing point events in the first event subsequence are used as rows, and the appearing point events in the second event subsequence are used as columns. Each position in the cumulative distance table corresponds to one vanishing point event and one appearing point event as the target object to be matched. The cumulative distance at the starting position of the cumulative distance table is set as the local distance of the starting position. When calculating the cumulative distance for each position in the first row, the cumulative distance of the previous column position is added to the local distance of the current position. When calculating the cumulative distance for each position in the first column, the cumulative distance of the previous row position is added to the local distance of the current position. When calculating the cumulative distance for positions that are neither in the first row nor the first column, the cumulative distance of the position above is calculated plus the local distance of the current position, the cumulative distance of the position to the left is calculated plus the local distance of the current position, and the cumulative distance of the position to the upper left is calculated plus the local distance of the current position. The minimum value among the three candidate cumulative distances is selected as the cumulative distance of the current position. The cumulative distance table is calculated in ascending order of row and column numbers.

[0060] In dynamic time warping, path constraints and matching range constraints are set. The matching range constraint requires that the sampling times corresponding to both the target vanishing point and the target appearing point be within the matching time range defined by the wireless signal duration interval. If the target to be matched does not meet the matching range constraint, the corresponding position in the cumulative distance table is marked as an unreachable position, and the cumulative distance of the unreachable position is not calculated. Path constraints are determined by the consistency of the spatial gate line crossing direction. When the spatial gate line crossing direction consistency is opposite, diagonal movement from the upper left position to the current position is not allowed; only movement from the upper or left position is permitted. When the spatial gate line crossing direction consistency is consistent or undetermined, movement from the upper, left, or upper left position to the current position is permitted. If all reachable previous positions for a given position are unreachable, the current position is marked as unreachable. The dynamic time warping ends when the last row and last column of the cumulative distance table are calculated, or all uncalculated positions are marked as unreachable.

[0061] Cross-node association confidence is generated based on the minimum cumulative distance obtained from dynamic time warping. If the terminal position of the cumulative distance table is reachable, the cumulative distance of the terminal position is taken as the minimum cumulative distance; if the terminal position of the cumulative distance table is not reachable, the position with the smallest cumulative distance is selected from the reachable positions in the last row and last column, and the cumulative distance of the selected position is taken as the minimum cumulative distance. The dynamic time warping path is obtained by tracing backward from the position corresponding to the minimum cumulative distance to the starting position, and the number of cumulative distance table positions contained in the dynamic time warping path is counted as the length of the dynamic time warping path. The unit of the dynamic time warping path length is the number of events. The normalized cumulative distance is obtained by dividing the minimum cumulative distance by the dynamic time warping path length. Cross-node association confidence is generated based on the normalized cumulative distance, and the value of the cross-node association confidence ranges from 0 to 1. When the normalized cumulative distance is 0, the cross-node association confidence score is 1; when the normalized cumulative distance reaches or exceeds the upper limit of the confidence distance, the cross-node association confidence score is 0; when the normalized cumulative distance is between 0 and the upper limit of the confidence distance, the cross-node association confidence score is generated according to the ratio between the normalized cumulative distance and the upper limit of the confidence distance, so that the larger the normalized cumulative distance, the smaller the cross-node association confidence score. The upper limit of the confidence distance is determined based on the installation and debugging samples. The normalized cumulative distances corresponding to known cross-node movement events and known non-cross-node movement events are statistically analyzed, and the distance value that can distinguish between the two types of samples is selected as the upper limit of the confidence distance. For example, the upper limit of the confidence distance can be set to 3 to 8. The cross-node association confidence score is written into the candidate cross-node event pairs, and the following are retained: entry and exit time phase difference, consistency of spatial gate crossing direction, wireless signal continuity interval, event sequence to be normalized, dynamic time normalization path, minimum cumulative distance, and normalized cumulative distance.

[0062] Based on the cross-node association confidence level, duplicate targets, split targets, and independent targets are labeled, and corrected pedestrian density, corrected pedestrian flow direction, and flow conservation residuals are generated, including: Set an association merging threshold. The association merging threshold is used to determine whether candidate cross-node event pairs have cross-node association relationships. The value of the association merging threshold ranges from 0 to 1. The association merging threshold is determined based on the installation and debugging samples, which include known cross-node movement events and known non-cross-node movement events. Known cross-node movement events are formed when testers carry test terminals across spatial gate lines along a preset route. Known non-cross-node movement events are formed when testers carry test terminals within the coverage area of ​​a single node but do not cross spatial gate lines. The cross-node association confidence scores corresponding to known cross-node movement events and known non-cross-node movement events are calculated separately. The association merging threshold is selected when the majority of known cross-node movement events are above the association merging threshold and the majority of known non-cross-node movement events are below the association merging threshold. For example, the association merging threshold can be set to 0.6 to 0.85. Candidate cross-node event pairs with a cross-node association confidence level not less than the association merging threshold are designated as mergeable candidate cross-node event pairs, while candidate cross-node event pairs with a cross-node association confidence level less than the association merging threshold are designated as low-association candidate cross-node event pairs.

[0063] For mergeable candidate cross-node event pairs, checks are performed for the same vanishing point and the same appearance point. The same vanishing point check uses the edge computing node identifier, pedestrian target number, sampling time corresponding to the vanishing point, spatial gate line, and boundary candidate grid number as comparison fields. If multiple mergeable candidate cross-node event pairs have the exact same vanishing point, they are determined to correspond to the same vanishing point. The same appearance point check uses the edge computing node identifier, pedestrian target number, sampling time corresponding to the appearance point, spatial gate line, and boundary candidate grid number as comparison fields. If multiple mergeable candidate cross-node event pairs have the exact same appearance point, they are determined to correspond to the same appearance point. If either the vanishing point or the appearance point lacks any comparison field, the candidate cross-node event pair containing the missing field is marked as a field-missing candidate cross-node event pair, and these pairs are not included in the duplicate target marking.

[0064] Candidate cross-node event pairs corresponding to the same vanishing point or the same appearing point of a target are marked as duplicate targets. For multiple mergeable candidate cross-node event pairs corresponding to the same vanishing point of a target, the mergeable candidate cross-node event pair with the highest cross-node association confidence is retained as the representative event pair for merging, and the remaining mergeable candidate cross-node event pairs are merged into the representative event pair. For multiple mergeable candidate cross-node event pairs corresponding to the same appearing point of a target, the mergeable candidate cross-node event pair with the highest cross-node association confidence is retained as the representative event pair for merging, and the remaining mergeable candidate cross-node event pairs are merged into the representative event pair. If the cross-node association confidence of multiple mergeable candidate cross-node event pairs is the same, the mergeable candidate cross-node event pair with the smaller absolute value of the entry and exit time phase difference is selected as the representative event pair for merging; if the absolute value of the entry and exit time phase difference is still the same, the mergeable candidate cross-node event pair with consistent spatial gate line crossing direction records is selected as the representative event pair for merging; if the spatial gate line crossing direction consistency is still the same, the representative event pair for merging is determined according to the rule of the earlier sampling time corresponding to the target appearing point. The merging of representative event pairs, candidate cross-node event pairs that are merged into the merging representative event pairs, duplicate target labels, and merging criteria together form the merging result of duplicate targets.

[0065] Target vanishing points or target appearing points with wireless signal continuation intervals but without completed cross-node pairing are marked as fragmented targets. Incomplete cross-node pairing means that the target vanishing point has not formed a candidate cross-node event pair with a cross-node association confidence level not less than the association merging threshold, or the target appearing point has not formed a candidate cross-node event pair with a cross-node association confidence level not less than the association merging threshold. Within the range defined by the boundary candidate grid number and the unified time window number, target vanishing points and target appearing points without completed cross-node pairing are searched. Then, wireless signal continuation intervals that coincide with the target vanishing point or target appearing point without completed cross-node pairing at the sampling time are searched. If the start time of the wireless signal continuation interval is not later than the sampling time corresponding to the target vanishing point or the target appearing point, and the end time of the wireless signal continuation interval is not earlier than the sampling time corresponding to the target vanishing point or the target appearing point, the wireless signal continuation interval is considered to coincide with the target vanishing point or target appearing point without completed cross-node pairing at the sampling time. If there is no wireless signal continuation interval with the same sampling time at the target vanishing point or the target appearing point that has not completed cross-node pairing, the target vanishing point or the target appearing point that has not completed cross-node pairing will not be marked as a split target.

[0066] When patching a severed target, the system searches for incomplete cross-node pairing events of the opposite type to the severed target within the same overlapping node group identifier, the same spatial gate line, and adjacent unified time windows. If the severed target is the target vanishing point, the system searches for the target spawn point on the adjacent edge computing node side; if the severed target is the target spawn point, the system searches for the target vanishing point on the adjacent edge computing node side. It is required that the searched event and the severed target share the same radio signal duration interval, and the time interval between the searched event and the severed target falls within the candidate pairing time range. When these conditions are met, the severed target and the searched event form a patching event pair, and the patching event pair is written into the patching result of the severed target. If the same severed target corresponds to multiple searched events, the searched event with the longest overlapping duration of the shared radio signal duration interval with the severed target is selected first; if the overlapping durations of the radio signal duration intervals are the same, the searched event with the smaller time interval with the severed target is selected; if the time intervals are still the same, the searched event with consistent spatial gate line crossing direction is selected as the consistent searched event. If no matching event can be found for the severed target, the event record of one side of the severed target is retained, and the failure status of the repair is recorded in the repair result of the severed target.

[0067] Candidate cross-node event pairs that are not classified as duplicate or split targets are marked as independent targets. Independent targets include candidate cross-node event pairs with a cross-node association confidence level not less than the association merging threshold and that do not correspond to the same target vanishing point or the same target appearing point. They also include candidate cross-node event pairs with low association that do not have overlapping wireless signal duration intervals and have not triggered split target marking. For independent targets, the candidate cross-node event pair identifier, target vanishing point, target appearing point, cross-node association confidence level, entry / exit time phase difference, spatial gate line crossing direction consistency, and spatial gate line are saved. Independent targets are not merged or patched. Independent targets are included in the statistics based on the retained results when generating corrected pedestrian density and corrected pedestrian flow direction. The retained results for independent targets include the independent target label, the corresponding candidate cross-node event pair, and the corresponding boundary candidate grid number. If a candidate cross-node event pair with missing fields still contains the sampling time of the target vanishing point and the target appearing point, as well as the spatial gate line, the candidate cross-node event pair with missing fields will be retained as an independent target; if a candidate cross-node event pair with missing fields lacks the target vanishing point or the target appearing point, the candidate cross-node event pair with missing fields will not be written into the corrected target set.

[0068] Based on the merging results of duplicate targets, the completion results of fragmented targets, and the retention results of independent targets, the corrected pedestrian density and corrected pedestrian flow direction are generated. First, a corrected target set is formed within each boundary candidate grid and each unified time window. For duplicate targets, only one representative event pair is recorded in the corrected target set, and candidate cross-node event pairs merged into the representative event pair are not recorded repeatedly. For fragmented targets, one completion event pair is recorded when completion is successful, and one unilateral event record is recorded when completion fails. The retention results of independent targets are recorded one by one in the corrected target set according to candidate cross-node event pairs. The number of targets in the corrected target set is counted, and the corrected pedestrian density is generated based on the unified spatial grid area corresponding to the boundary candidate grid. The unified spatial grid area is calculated from the unified spatial grid boundary coordinates; when the unified spatial grid boundary coordinates use meters as the unit of length, the unit of the unified spatial grid area is square meters. The corrected pedestrian density is the value obtained by dividing the number of targets in the corrected target set by the area of ​​the uniform spatial grid. The unit of the corrected pedestrian density is the number of targets per square meter. If the area of ​​the uniform spatial grid is 0 or the area of ​​the uniform spatial grid cannot be calculated, the corresponding boundary candidate grid is marked as an area anomaly boundary candidate grid, and the corrected pedestrian density of the area anomaly boundary candidate grid is not calculated.

[0069] Read the spatial gate line crossing direction of each merged representative event pair, supplementary event pair, and independent target corresponding candidate cross-node event pair in the corrected target set, and determine the direction of the target crossing the spatial gate line based on the spatial gate line crossing direction. For single-sided event records that fail to supplement, if the single-sided event record is the target vanishing point, record the direction from the departure side to the spatial gate line corresponding to the target vanishing point as the single-sided flow direction; if the single-sided event record is the target appearing point, record the direction from the spatial gate line to the entry side corresponding to the target appearing point as the single-sided flow direction. Count the number of targets in the same flow direction according to a unified time window to generate the corrected pedestrian flow direction. The corrected pedestrian flow direction includes the starting point side, the ending point side, the corresponding spatial gate line, the unified time window number, and the number of targets. The unit for the number of targets is the number of targets, and the starting point side and the ending point side both originate from the spatial sides on both sides of the spatial gate line. If the spatial gate line crossing direction is recorded as an event with an undetermined direction, write the corresponding target into the undetermined direction record, and the undetermined direction record is not included in the inflow and outflow of the corrected pedestrian flow direction.

[0070] A flow conservation residual is generated based on the change in corrected pedestrian density within the boundary candidate grid over a unified time window, as well as the inflow and outflow of corrected pedestrian flow. Within each boundary candidate grid, the corrected pedestrian density is arranged in chronological order according to the unified time window. For adjacent unified time windows, the corrected pedestrian density of the previous unified time window is subtracted from the corrected pedestrian density of the later unified time window to obtain the difference in corrected pedestrian density. The unit of the difference in corrected pedestrian density is the target number per square meter. To ensure dimensional consistency with the inflow and outflow, the difference in corrected pedestrian density is multiplied by the area of ​​the unified spatial grid corresponding to the boundary candidate grid to generate the change in corrected pedestrian density over a unified time window. The unit of the change in corrected pedestrian density over a unified time window is the target number. When the change in corrected pedestrian density over a unified time window is positive, the number of targets in the boundary candidate grid increases; when the change in corrected pedestrian density over a unified time window is negative, the number of targets in the boundary candidate grid decreases; when the change in corrected pedestrian density over a unified time window is 0, the number of targets in the boundary candidate grid remains unchanged.

[0071] The inflow and outflow of people entering and leaving the boundary candidate grid are statistically analyzed based on the corrected pedestrian flow direction. Inflow represents the number of targets entering the current boundary candidate grid from the side covered by adjacent boundary candidate grids or adjacent nodes within a unified time window. Outflow represents the number of targets leaving the current boundary candidate grid to the side covered by adjacent boundary candidate grids or adjacent nodes within a unified time window. The number of targets whose endpoints in the spatial gate crossing direction are located inside the current boundary candidate grid is included in the inflow, and the number of targets whose starting points in the spatial gate crossing direction are located inside the current boundary candidate grid is included in the outflow. For single-sided event records of failed connections, the direction of the single-sided flow determines whether it is included in the inflow or outflow; if the single-sided flow is from the spatial gate line to the inside of the current boundary candidate grid, it is included in the inflow; if the single-sided flow is from the inside of the current boundary candidate grid to the spatial gate line, it is included in the outflow. The unit for inflow and outflow is the number of targets. If there is no corrected pedestrian flow direction in the current boundary candidate grid, both inflow and outflow are recorded as 0.

[0072] Flow conservation residuals are generated based on the changes in corrected pedestrian density over a unified time window, and the differences between inflow and outflow. First, the outflow is subtracted from the inflow to obtain the net inflow. The unit of net inflow is the target number. Then, the difference between the net inflow and the changes in corrected pedestrian density over a unified time window is calculated, and the absolute value of the difference is taken as the flow conservation residual. The unit of flow conservation residual is the target number. If the corrected pedestrian density is missing in the preceding or following unified time window, the flow conservation residual between the corresponding adjacent unified time windows is not calculated, and the corresponding boundary candidate grid and the corresponding unified time window are recorded as density missing states. Flow conservation residuals are saved for each boundary candidate grid and each unified time window, along with the changes in corrected pedestrian density over the unified time window, inflow, outflow, and net inflow used to generate the flow conservation residuals.

[0073] Based on the corrected pedestrian density, corrected pedestrian flow direction, flow conservation residual, and connectivity of the candidate boundary grids, the active boundary of the business district is updated, including: Based on the continuous distribution of corrected pedestrian density within a unified time window, active candidate grids are extracted from the boundary candidate grids. The continuous distribution of corrected pedestrian density within a unified time window indicates that the same boundary candidate grid has a corrected pedestrian density not lower than the active density threshold within consecutive unified time windows. The active density threshold is used to distinguish whether there is continuous pedestrian flow within the boundary candidate grids; the unit of the active density threshold is the number of targets per square meter. The active density threshold is determined based on installation and debugging samples. The corrected pedestrian density of the boundary candidate grids is statistically analyzed during low-traffic periods in the business district to obtain the low-traffic density distribution. Similarly, the corrected pedestrian density of the boundary candidate grids is statistically analyzed during normal traffic periods to obtain the normal traffic density distribution. The active density threshold is selected as the value at the larger end of the low-traffic density distribution that falls within the smaller end of the normal traffic density distribution. This ensures that scattered targets formed during low-traffic periods are not extracted as active candidate grids, while continuous targets formed during normal traffic periods can be extracted as active candidate grids. For example, the active density threshold can be set from 0.05 targets per square meter to 0.3 targets per square meter.

[0074] Within each boundary candidate grid, the corrected pedestrian density is arranged according to the time sequence of a unified time window, and it is checked whether the corrected pedestrian density within consecutive unified time windows is not lower than the active density threshold. The number of continuous windows is used to limit the number of unified time windows in which the corrected pedestrian density needs to continuously meet the active density threshold. The number of continuous windows is determined based on the length of the unified time window and the time required for pedestrians to cross the boundary candidate grid in the business district. When the length of the unified time window is short, the number of continuous windows increases; when the length of the unified time window is long, the number of continuous windows decreases. For example, the number of continuous windows can be set to 2 to 5 unified time windows. If the same boundary candidate grid has a corrected pedestrian density not lower than the active density threshold within consecutive unified time windows corresponding to the number of continuous windows, the same boundary candidate grid is marked as an active candidate grid. If the same boundary candidate grid has a corrected pedestrian density higher than the active density threshold only in one unified time window, and the corrected pedestrian density in adjacent unified time windows is lower than the active density threshold, the same boundary candidate grid is not marked as an active candidate grid.

[0075] The flow direction connection between active candidate grids is determined based on the corrected flow direction. The corrected flow direction includes the starting point side, the ending point side, the corresponding spatial gate line, the unified time window number, and the target quantity. The flow direction connection relationship indicates that within the same unified time window, one active candidate grid is connected to another active candidate grid through the corrected flow direction. The unified spatial grid number and spatial gate line of the active candidate grid are read, and the corrected flow direction within the same unified time window is read. If the boundary candidate grid corresponding to the starting point side of the corrected flow direction is one active candidate grid, and the boundary candidate grid corresponding to the ending point side of the corrected flow direction is another active candidate grid, a flow direction connection relationship is established between the two active candidate grids. The flow direction connection relationship stores the starting point active candidate grid number, the ending point active candidate grid number, the unified time window number, the corresponding spatial gate line, and the target quantity. If the starting point side or the ending point side of the corrected flow direction cannot be mapped to an active candidate grid, the corresponding corrected flow direction is not used to establish a flow direction connection relationship.

[0076] When determining flow direction connections, a flow direction connection number threshold is set. This threshold excludes weak connections caused by single noisy targets or isolated targets, and the unit for the flow direction connection number threshold is the number of targets. The flow direction connection number threshold is determined based on the target number distribution of real cross-grid passage events and isolated false detection events in the installation and debugging samples. The number of targets corresponding to real cross-grid passage events and isolated false detection events is statistically analyzed, and the number of targets that retains real cross-grid passage events while reducing isolated false detection events is selected as the flow direction connection number threshold. For example, the flow direction connection number threshold can be set from 1 to 3 targets. If the number of targets in the corrected flow direction is not less than the flow direction connection number threshold, the corresponding flow direction connection is retained; if the number of targets in the corrected flow direction is less than the flow direction connection number threshold, the corresponding flow direction connection is not established. If the corrected flow direction is recorded as an undetermined direction record, flow direction connection relationships are not established using undetermined direction records.

[0077] Based on the flow conservation residuals, active candidate grids are removed from the active candidate grids if the changes in the corrected pedestrian flow direction, the corrected pedestrian flow direction, and the corrected pedestrian density over a unified time window are mismatched. A residual removal threshold is set to determine whether the changes in the corrected pedestrian flow direction, the corrected pedestrian flow direction, and the corrected pedestrian density over a unified time window match within the active candidate grid. The residual removal threshold is determined based on installation and commissioning samples. The flow conservation residuals of known normal traffic boundary candidate grids are statistically analyzed over multiple unified time windows to obtain a normal residual distribution. The residual value closest to the larger end of the normal residual distribution is selected as the residual removal threshold. For example, the residual removal threshold can be set to 2 to 5 targets.

[0078] Within each active candidate grid, the flow conservation residual is read and compared to the residual elimination threshold. If the flow conservation residual of the same active candidate grid within the current unified time window is greater than the residual elimination threshold, the flow conservation residual of the same active candidate grid within adjacent unified time windows is checked. If the same active candidate grid has a flow conservation residual greater than the residual elimination threshold within consecutive unified time windows corresponding to the number of consecutive residual abnormal windows, the same active candidate grid is removed from the active candidate grid. The number of consecutive residual abnormal windows is determined based on the length of the unified time window and the fluctuation range of node reporting latency. When the length of the unified time window is short or the fluctuation range of node reporting latency is large, the number of consecutive residual abnormal windows is increased. For example, the number of consecutive residual abnormal windows can be set to 2 to 4 unified time windows. If the flow conservation residual is greater than the residual elimination threshold only within one unified time window, and the flow conservation residual in adjacent unified time windows is not greater than the residual elimination threshold, the same active candidate grid is retained.

[0079] Connectivity merging is performed based on the connectivity relationships of the retained active candidate grids within the boundary candidate grids. Connectivity relationships indicate that two retained active candidate grids are spatially connected in the unified spatial grid through boundary connections, corner connections, or connections through intersecting regions of node coverage areas. The boundary coordinates of the unified spatial grid corresponding to the retained active candidate grids are read, and the boundary coordinates of any two retained active candidate grids are compared. If two retained active candidate grids share a common boundary segment, they are considered spatially adjacent. If two retained active candidate grids only share a common corner, a flow-direction connection is checked. If a flow-direction connection exists, the two retained active candidate grids are considered spatially adjacent; otherwise, they are not. If two retained active candidate grids are located on opposite sides of the intersecting region of the same node coverage area, and a flow-direction connection exists between them, they are considered spatially adjacent.

[0080] During connectivity merging, starting with any one unmerged retained active candidate grid, other retained active candidate grids with connectivity to it are searched, and these are added to the same connected set. Starting with a newly added retained active candidate grid in the connected set, other retained active candidate grids with connectivity are searched until no new retained active candidate grids are added to the connected set. Connectivity merging is repeated for the remaining unmerged retained active candidate grids, resulting in one or more connected sets. Each connected set stores the connected set number, the numbers of the included active candidate grids, the unified time window number, the associated flow direction connectivity, and the corresponding unified spatial grid boundary coordinates. If a connected set contains only one retained active candidate grid and there is no flow direction connectivity within the corresponding unified time window, it is marked as an isolated connected set.

[0081] The active boundary of the business district is updated based on the flow direction connectivity. Isolated connected sets are removed. The determination method for the continued existence of isolated connected sets is: the unified spatial grid number corresponding to the same isolated connected set is marked as an active candidate grid within a continuous unified time window, and the number of continuous unified time windows is not less than the number of continuous windows. The outer contour boundary is extracted for the connected sets that are not removed. The outer contour boundary is composed of the line segments of the unified spatial grid boundary located on the outside of the connected set. The outer contour boundary is checked for directional breaks according to the flow direction connectivity. A directional break indicates that there is a lack of flow direction connectivity between adjacent active candidate grids on both sides of the outer contour boundary, and the flow conservation residual is greater than the residual removal threshold. If there is a directional break in the outer contour boundary, the outer contour boundary is contracted at the corresponding position of the directional break so that the active candidate grid corresponding to the directional break does not enter the active boundary of the business district. If there is a continuous flow direction connectivity between adjacent positions of the outer contour boundary, the corresponding outer contour boundary is retained. The coordinate unit of the outer contour boundary is consistent with the length unit of the coordinate of the unified spatial grid boundary.

[0082] When updating the active boundaries of a business district, the outer contour boundaries of connected sets that have not been eliminated are used as candidate active boundaries for the current unified time window. If multiple connected sets that have not been eliminated exist within the same unified time window, multiple candidate active boundaries are generated for each, and the existence of pedestrian connectivity between these candidate active boundaries is determined based on the flow direction connection relationship. If at least one flow direction connection relationship with a target number not less than the flow direction connection number threshold exists between multiple candidate active boundaries, these multiple candidate active boundaries are merged into a single active boundary. If no flow direction connection relationship with a target number not less than the flow direction connection number threshold exists between multiple candidate active boundaries, multiple independent active boundaries are retained. The active boundary of a business district stores the active boundary number, unified time window number, outer contour boundary coordinates, included active candidate grid numbers, associated flow direction connection relationships, associated corrected pedestrian density, and associated flow conservation residuals.

[0083] When updating the business district activity boundary between consecutive unified time windows, the activity boundary of the current unified time window is compared with that of the previous unified time window. If there is an overlapping area between the activity boundary of the current unified time window and that of the previous unified time window, the activity boundary of the current unified time window is retained, and the active candidate grids corresponding to the overlapping area are recorded as continuously active areas. If the activity boundary of the current unified time window expands outward relative to the activity boundary of the previous unified time window, it is checked whether the active candidate grids in the expanded area have flow direction connections and whether the flow conservation residual is not greater than the residual elimination threshold. If the active candidate grids in the expanded area have flow direction connections and the flow conservation residual is not greater than the residual elimination threshold, the expanded area is incorporated into the activity boundary of the business district. If the active candidate grids in the expanded area do not have flow direction connections, or the flow conservation residual is greater than the residual elimination threshold, the expanded area is not incorporated into the activity boundary of the business district. The output is the active boundary of the business district updated according to a unified time window. The active boundary of the business district serves as the result of real-time business district pedestrian flow analysis and boundary calibration.

[0084] 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 real-time commercial area pedestrian flow analysis and boundary calibration system based on edge computing, characterized in that, include: Node recording module: Acquires node observation records of multiple edge computing nodes; Grid candidate module: Divide overlapping node groups according to the adjacency and overlap of node coverage areas, map node observation records to a unified spatial grid and a unified time window, and obtain boundary candidate grids; Cross-point event module: Extracts the target vanishing point, target appearing point, spatial gate crossing direction, and wireless signal continuation interval from the boundary candidate grid to construct a cross-node event sequence; The association confidence module performs dynamic time warping based on the entry and exit time phase difference, spatial gate crossing direction consistency, and wireless signal continuity interval in the cross-node event sequence to obtain the cross-node association confidence level. Pedestrian flow correction module: Based on the cross-node association confidence level, it labels repeated targets, split targets, and independent targets, and generates corrected pedestrian flow density, corrected pedestrian flow direction, and flow conservation residuals. Boundary Update Module: Updates the active boundary of the business district based on the corrected pedestrian density, corrected pedestrian flow direction, flow conservation residual, and the connectivity of the boundary candidate grid.

2. The real-time commercial area pedestrian flow analysis and boundary calibration system based on edge computing according to claim 1, characterized in that, Obtain node observation records from multiple edge computing nodes, specifically: Configure edge computing node identifiers for multiple edge computing nodes, and have multiple edge computing nodes collect video streams and wireless probe signals within the node's coverage area; Extract video detection data from the video stream; extract wireless probe data from the wireless probe signal. Read the node coordinates and node coverage area, and record the sampling time corresponding to the video detection data and wireless probe data; Calculate the node reporting delay based on the data transmission and reception times; Node observation records are generated by associating edge computing node identifiers with video detection data, wireless probe data, node coordinates, sampling time, node reporting delay, and node coverage.

3. The real-time commercial area pedestrian flow analysis and boundary calibration system based on edge computing according to claim 2, characterized in that, Overlapping node groups are formed based on the adjacency and overlap relationships of node coverage areas. Node observation records are then mapped to a unified spatial grid and a unified time window to obtain candidate boundary grids, specifically: Based on the edge calculation node identifier, node coordinates and node coverage area in the node observation record, determine the adjacency and overlap relationships between the node coverage areas to form overlapping node groups; The sampling time is time-aligned based on the node reporting delay, a unified spatial grid is established based on the node coordinates and node coverage, and a unified time window is established based on the time-aligned sampling time. Video detection data and wireless probe data are mapped to a unified spatial grid and a unified time window. The unified spatial grid located at the boundary of the node coverage area or the area where the node coverage areas intersect is selected to obtain the boundary candidate grid.

4. The real-time commercial area pedestrian flow analysis and boundary calibration system based on edge computing according to claim 3, characterized in that, Extract the target vanishing point, target spawn point, spatial gate crossing direction, and wireless signal continuity interval from the boundary candidate grid to construct a cross-node event sequence, specifically: Within the candidate boundary grid, the grid boundary corresponding to the boundary where the node coverage area is adjacent or the area where the node coverage area intersects is used as the spatial gate line; Based on video detection data, the target vanishing point is generated at the termination position of the relative spatial gate line within a unified time window; Based on video detection data, the target occurrence point is generated at the starting position relative to the spatial gate line within a unified time window; The spatial gate line crossing direction is generated based on the positional changes of the target vanishing point and the target appearing point relative to the spatial gate line. The wireless signal continuity interval is generated based on the consecutive occurrence of wireless probe data within adjacent unified time windows. By associating the sampling time with the target disappearance point, target appearance point, spatial gate crossing direction, and wireless signal continuity interval, a cross-node event sequence is constructed.

5. The real-time commercial area pedestrian flow analysis and boundary calibration system based on edge computing according to claim 4, characterized in that, The spatial gate line crossing direction is generated based on the positional changes of the target vanishing point and the target spawn point relative to the spatial gate line, specifically: Based on the positional relationship between the vanishing point and the appearing point of the target on both sides of the spatial gate line, determine the departure side corresponding to the vanishing point and the entry side corresponding to the appearing point of the target; The direction from the departure side to the entry side is taken as the initial crossing direction. The initial crossing direction is verified according to the order of the sampling time corresponding to the target vanishing point and the sampling time corresponding to the target appearing point, and the spatial gate crossing direction is generated.

6. The real-time commercial area pedestrian flow analysis and boundary calibration system based on edge computing according to claim 5, characterized in that, Based on the entry and exit time phase difference, spatial gate crossing direction consistency, and wireless signal duration interval in the cross-node event sequence, dynamic time warping is performed to obtain the cross-node association confidence, specifically: Pair the target vanishing point and target appearing point located at adjacent edge computing nodes in the cross-node event sequence to form candidate cross-node event pairs; The entry and exit time phase difference is generated based on the difference between the sampling time corresponding to the target disappearance point and the sampling time corresponding to the target appearance point in the candidate cross-node event pair. Compare the spatial gate crossing directions in candidate cross-node event pairs to generate spatial gate crossing direction consistency; Find the wireless signal duration interval that coincides with the sampling time of the candidate cross-node event pair, limit the matching range of dynamic time warping by the wireless signal duration interval, and perform dynamic time warping based on the phase difference between entry and exit time and the consistency of the spatial gate crossing direction to obtain the cross-node association confidence.

7. The real-time commercial area pedestrian flow analysis and boundary calibration system based on edge computing according to claim 6, characterized in that, Dynamic time warping is performed based on the phase difference between entry and exit times and the consistency of the spatial gate crossing direction to obtain the cross-node association confidence, specifically: Arrange candidate cross-node event pairs according to sampling time to form an event sequence to be normalized; Within the matching range defined by the wireless signal continuation interval, dynamic time warping is performed on the event sequence to be warped, using the phase difference between entry and exit times as the time axis constraint and the consistency of the spatial gate crossing direction as the path constraint. Cross-node association confidence is generated based on the minimum cumulative distance obtained by dynamic time warping.

8. The real-time commercial area pedestrian flow analysis and boundary calibration system based on edge computing according to claim 7, characterized in that, Based on the cross-node association confidence level, duplicate targets, split targets, and independent targets are labeled, and corrected pedestrian density, corrected pedestrian flow direction, and flow conservation residuals are generated, specifically as follows: Based on the cross-node association confidence, candidate cross-node event pairs are merged, and candidate cross-node event pairs that correspond to the same target vanishing point or the same target appearing point are marked as duplicate targets. Mark the vanishing point or the appearing point of a target that has a wireless signal continuity range and has not completed cross-node pairing as a split target; Candidate cross-node event pairs that are not classified as duplicate or split targets are marked as independent targets; Based on the merging results of duplicate targets, the supplementation results of fragmented targets, and the retention results of independent targets, the corrected pedestrian density and corrected pedestrian flow direction are generated. The flow conservation residual is generated based on the change of the corrected pedestrian density within the candidate boundary grid over a unified time window, as well as the inflow and outflow of the corrected pedestrian flow direction.

9. The real-time commercial area pedestrian flow analysis and boundary calibration system based on edge computing according to claim 8, characterized in that, Based on the change in corrected pedestrian density within the candidate boundary grid over a unified time window, and the inflow and outflow of corrected pedestrian flow directions, a flow conservation residual is generated, specifically as follows: Within each boundary candidate grid, the change in corrected pedestrian density with respect to the unified time window is generated based on the difference in corrected pedestrian density between adjacent unified time windows. Based on the corrected flow direction, the inflow into the boundary candidate grid and the outflow from the boundary candidate grid are statistically analyzed; The flow conservation residual is generated based on the change in the corrected pedestrian density over a uniform time window, the difference between the inflow and outflow, and the flow rate.

10. The real-time commercial area pedestrian flow analysis and boundary calibration system based on edge computing according to claim 9, characterized in that, Based on the corrected pedestrian density, corrected pedestrian flow direction, flow conservation residual, and connectivity of the candidate boundary grids, the active boundary of the business district is updated, specifically as follows: Based on the continuous distribution of the corrected pedestrian density within a unified time window, active candidate grids are extracted from the boundary candidate grids; The flow direction connection relationship between active candidate grids is determined based on the corrected flow direction of people; Based on the flow conservation residual, active candidate grids that do not match the changes in the corrected pedestrian flow direction, the corrected pedestrian flow direction, and the corrected pedestrian flow density over a uniform time window are removed from the active candidate grids. Connectivity merging is performed based on the connectivity relationships of the retained active candidate grids in the boundary candidate grids, and the active boundary of the business district is updated based on the flow direction connectivity.

Images