A safety state monitoring method and system for urban slow traffic

Abstract

The application discloses a kind of safety state monitoring method and system of urban slow traffic, it is related to traffic space-time data processing technical field, including to the electronic data of urban slow traffic, constructs slow scene topology graph and generates time-varying visibility matrix.Slowness scene topology graph and time-varying visibility matrix are used to establish double memory state and generate candidate continuous trajectory.Based on candidate continuous trajectory, identity correlation score is calculated, identity corresponding relationship is determined, and subject level safety state value is output.The application constructs continuous data processing chain of synchronous modeling, double memory maintenance, associated output around urban slow traffic electronic data, realizes the unified organization and unified update of topological constraint, visibility constraint, negative evidence constraint and state value calculation.Compared with the distributed processing mode, the application can convert multi-source space-time objects into indexable, associated, accumulative digital data objects.

Description

Technical Field

[0001] This invention relates to the field of traffic spatiotemporal data processing, specifically a method and system for monitoring the safety status of urban slow-moving traffic. Background Technology

[0002] With the development of roadside sensing devices, target detection and tracking algorithms, spatiotemporal data modeling, and graph computation methods, electronic data processing technology for traffic objects has gradually expanded from single target identification to multi-source data fusion, continuous trajectory modeling, and state-level output. Existing solutions typically perform time alignment, coordinate unification, and object extraction on observations from video, radar, or lidar, and then combine this with topology, trajectory prediction, or rule constraints to achieve target association and state analysis. Technical approaches centered on object-level spatiotemporal data organization, matrix encoding, graph structure reasoning, and continuous state computation have become important development directions for traffic electronic data processing.

[0003] However, existing technologies for continuous state monitoring of slow-moving subjects still primarily rely on instantaneous observations, short-term motion continuity, or single association rules. This leads to discrete data object representation, weak association criteria, and unstable state updates under conditions of occlusion, intermittent visibility, and dense mixed traffic. Existing methods often treat unobserved data as missing noise, lacking a data organization mechanism to transform information such as expected absences, blind zone constraints, and prohibited boundary crossings into computable negative evidence, making it difficult to support continuous identity maintenance during occlusion. Existing methods typically process topological constraints, visibility constraints, and association scores in a fragmented manner, failing to form a unified data processing chain, resulting in a disconnect between trajectory candidate generation and identity correspondence. Existing methods lack sufficient connection between identity association results and state value calculation, making it difficult to uniformly map object-level continuous trajectories, occlusion information, topological deviations, and mixed traffic relationships into continuously updated subject-level state values. Therefore, existing technologies are still unable to achieve a unified calculation method for continuous identity correspondence and state values ​​of slow-moving subjects based on topological constraints, visibility constraints, and negative evidence constraints. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing spatiotemporal data processing methods suffer from discrete data object representation, insufficient utilization of occlusion constraints, poor stability of continuous association, and the problem of how to achieve continuous identity correspondence and unified calculation of subject-level security state values ​​for slow-moving subjects under topological and visibility constraints.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for monitoring the safety status of urban slow-moving traffic, which involves synchronizing electronic data of urban slow-moving traffic, constructing a topology map of the slow-moving scene, and generating a time-varying visibility matrix.

[0007] Based on the topology map of the slow-moving scene and the time-varying visibility matrix, a dual-memory state is established and candidate continuous trajectories are generated.

[0008] The identity association score is calculated based on the candidate continuous trajectory, the identity correspondence is determined, and the subject-level security status value is output.

[0009] As a preferred embodiment of the urban slow-traffic safety status monitoring method described in this invention, the construction of the slow-traffic scenario topology map includes mapping zebra crossing entrances, zebra crossing mid-sections, waiting areas, non-motorized vehicle lane entrances, mixed-traffic areas, curb boundary connection points, and pedestrian crossing exit areas from the synchronized urban slow-traffic electronic data as nodes, and mapping the passable relationships between nodes as directed edges. For each directed edge, according to the slow-traffic subject category, the allowed passage direction, length range, signal phase correlation, boundary crossing restrictions, and historical transfer frequency are written, and a node index table, edge index table, adjacency table, and topology snapshot sequence are established. When temporary barriers, road closures, or changes in traffic organization occur in the urban slow-traffic electronic data, the affected nodes and directed edges are reconstructed, and the corresponding topology snapshots are updated.

[0010] As a preferred embodiment of the urban slow-moving traffic safety status monitoring method described in this invention, the generation of the time-varying visibility matrix includes: based on the node positions and directed edge ranges in the slow-moving scene topology map, performing gridded projection on the sensing device installation pose, field of view boundary, outline of static occlusion facilities, and area occupied by dynamic occlusion entities at each sampling time to obtain a visibility code corresponding to the topology unit. The visibility code distinguishes between fully visible, semi-visible, and completely blind zones, and writes the duration, occlusion source category, belonging topology unit, and pointers to adjacent visible units into each coded unit. When a dynamic occlusion entity enters or leaves the directed edge coverage area, the visibility recalculation of the corresponding grid unit and topology unit is triggered according to the same timestamp, and the time-varying visibility matrix sequence is updated.

[0011] As a preferred embodiment of the urban slow-moving traffic safety status monitoring method described in this invention, the establishment of a dual-memory state includes: establishing positive evidence memory, negative evidence memory, current state vector, and identity survival marker for each slow-moving entity. Positive evidence memory continuously records observed location, speed, orientation, category label, and associated directed edge. Negative evidence memory continuously records expected but unoccurred events, completely blind zone units where reasonable stopping is permissible, and prohibited crossing boundary constraints, calculated based on the slow-moving scene topology and time-varying visibility matrix. When the same slow-moving entity transitions from fully visible to semi-visible or completely blind zone within a continuous sampling period, the start time of the negative evidence memory is written according to the last observed state, and the expected but unoccurred event count, completely blind zone stopping interval, and identity survival marker are updated in units of sampling periods.

[0012] As a preferred embodiment of the urban slow-moving traffic safety status monitoring method described in this invention, the generation of candidate continuous trajectories includes: reading the last observed position, speed, orientation, and directed edge of the slow-moving subject in a dual-memory state; calculating the motion boundary according to the slow-moving trajectory propagation time window; and unfolding reachable paths along adjacent directed edges in the slow-moving scene topology map. The reachable paths are intersected with the time-varying visibility matrix at each time step, and path segments that cross prohibited boundaries, enter prohibited directed edges in reverse, exceed the tolerance time of the complete blind zone, or are continuously missing from fully visible or semi-visible coding units are deleted. The remaining path segments are sorted according to positional continuity, directional continuity, and group interaction direction to form a candidate continuous trajectory sequence bound to the dual-memory state.

[0013] As a preferred embodiment of the urban slow-moving traffic safety status monitoring method described in this invention, the step of calculating the identity association score based on candidate continuous trajectories includes: constructing an association matrix between each newly observed entity and each candidate continuous trajectory; and calculating the degree of appearance consistency, motion continuity, topological reachability, occlusion consistency, and negative evidence violation. The degree of negative evidence violation is determined based on whether the newly observed entity can only be located in the corresponding complete blind zone during its disappearance period, whether there is a jump across the prohibited boundary, and whether it abnormally reappears after a continuous absence of a fully visible or semi-visible coding unit. After normalizing each degree to the same dimension, a weighted combination is performed, and the matching result with the highest score that satisfies the mutual exclusion assignment constraint is taken as the identity correspondence. The unmatched observed entities and unmatched candidate continuous trajectories are written into the new subject processing branch and the continue-maintain branch, respectively.

[0014] As a preferred embodiment of the urban slow-moving traffic safety status monitoring method described in this invention, the step of determining the identity correspondence and outputting the subject-level safety status value includes, after determining the identity correspondence, calculating the conflict approximation quantity, occlusion uncertainty quantity, topology deviation quantity, and mixed-traffic occupancy quantity according to the continuous trajectory, directed edge, current visibility code, relative position relationship of adjacent traffic participants, and mixed-traffic occupancy relationship of the corresponding slow-moving subject determined by the identity correspondence, and accumulating and updating the conflict approximation quantity, occlusion uncertainty quantity, topology deviation quantity, and mixed-traffic compression quantity according to a unified time base. When any quantity undergoes a cross-level change, continuous missing recovery, or directed edge switching, the subject-level safety status value is recalculated, and the current status record bound to the slow-moving subject is output according to the slow-moving safety status mapping rules.

[0015] As a preferred embodiment of the urban slow-moving traffic safety status monitoring system of the present invention, it includes a scene topology visible modeling module, a dual-memory continuous trajectory module, and an identity-associated status output module.

[0016] The scene topology visibility modeling module is used to synchronize urban slow-moving traffic electronic data, construct a slow-moving scene topology map, and generate a time-varying visibility matrix.

[0017] The dual-memory continuous trajectory module is used to establish a dual-memory state and generate candidate continuous trajectories based on the slow-moving scene topology map and the time-varying visibility matrix.

[0018] The identity association status output module is used to calculate the identity association score based on the candidate continuous trajectory, determine the identity correspondence, and output the subject-level security status value.

[0019] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a method for monitoring the safety status of urban slow-moving traffic.

[0020] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of a method for monitoring the safety status of urban slow-moving traffic.

[0021] The beneficial effects of this invention are: By synchronizing time, unifying coordinates, constructing topology graphs, and generating time-varying visibility matrices for urban slow-moving traffic electronic data, a structured reorganization and computable representation of discrete observation data were achieved. The original detection, tracking, pose, and semantic data were transformed into standardized data objects such as nodes, edges, matrices, and indexes, serving as unified inputs for subsequent state maintenance and correlation calculations. This provides a consistent spatiotemporal reference, a unified data organization format, and reusable constraints for subsequent steps, enabling subsequent calculations to rely on continuous digital models instead of single-frame observations.

[0022] By establishing a dual-memory state based on the slow-moving scene topology map and time-varying visibility matrix, and generating candidate continuous trajectories on this basis, the synchronous maintenance of observation information, missing information, and constraint information of the slow-moving subject is achieved. Observed data and negative evidence such as expected absences, reasonable stops, and prohibited crossings are all incorporated into the state vector and trajectory propagation process, transforming candidate trajectory generation from simple motion extrapolation to a continuous search constrained by topology, visibility, and memory. This achieves the beneficial effect of forming an interpretable candidate trajectory set even under conditions of occlusion, intermittent visibility, and path bifurcation, providing a more stable data foundation for subsequent identity correspondence calculations.

[0023] By constructing an association matrix between newly observed entities and candidate continuous trajectories, calculating multidimensional association scores, and outputting subject-level security state values, a unified digital determination is achieved, from candidate trajectory sets to identity matching results and state value records. Data constraints from different sources, such as appearance, motion, topology, occlusion, and negative evidence, are compressed into the same computational framework, and the identity matching results are further mapped to continuously updated subject-level state variables. This achieves the beneficial effect of integrating identity determination and state calculation into a unified processing chain, avoiding subsequent state outputs from relying solely on instantaneous matching results, thereby improving the continuity, traceability, and data consistency of state records. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 The above is an overall flowchart of a method for monitoring the safety status of urban slow-moving traffic provided in Embodiment 1 of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for monitoring the safety status of urban slow-moving traffic is provided, comprising: S1: Synchronize the electronic data of urban slow traffic, construct a topology map of slow traffic scenarios, and generate a time-varying visibility matrix.

[0028] The synchronized urban slow-traffic electronic data maps zebra crossing entrances, zebra crossing mid-sections, waiting areas, non-motorized vehicle lane entrances, mixed-traffic areas, curb boundary connection points, and pedestrian crossing exits to nodes. The passability relationships between nodes are mapped to directed edges. For each directed edge, based on the slow-traffic subject category, the permitted travel direction, length range, signal phase correlation, boundary crossing restrictions, and historical transfer frequency are written. A node index table, edge index table, adjacency table, and topology snapshot sequence are established. For directed edges not controlled by signals, the signal phase correlation field is filled with a null value code or a fixed no-control marker. When temporary barriers, road closures, or changes in traffic organization occur in the urban slow-traffic electronic data, the affected nodes and directed edges are reconstructed, and the corresponding topology snapshots are updated.

[0029] Furthermore, the "slow-moving" mentioned in this invention refers to slow-moving groups, including pedestrians, cyclists, e-bike riders, wheelchair users, walking aid users, scooter users, and other vulnerable traffic participants who participate in road traffic at low speeds. Pedestrians include walkers, those stopping, and those crossing the street. Cyclists and e-bike riders are identified according to their riding status, while wheelchair users, walking aid users, and scooter users are identified according to their corresponding body and vehicle combination. In urban slow-moving traffic electronic data, the aforementioned slow-moving groups are uniformly assigned a slow-moving subject category label.

[0030] Furthermore, the electronic data for urban slow-moving traffic includes at least the target detection result stream, target tracking result stream, device pose parameter stream, road semantic annotation stream, and signal phase state stream. All data streams are uniformly mapped to the same Cartesian coordinate system and timestamps are aligned with a fixed sampling period of 100 milliseconds. When the time difference between two adjacent frames exceeds 150 milliseconds, the state of the previous valid frame is used to maintain the data once; when the time difference exceeds 300 milliseconds, it is marked as a broken frame and does not participate in the current cycle's topology update.

[0031] Furthermore, nodes are generated using a semantic region center point + boundary constraints approach. For zebra crossing entrances, waiting areas, and pedestrian exit areas, nodes are generated based on the geometric center of the region. Edge boundary connection points are generated based on boundary vertices. Node merging is performed when the Euclidean distance between two candidate nodes is less than 1.2 meters and their semantic labels are identical. This 1.2-meter threshold is set based on the sum of the maximum single-step displacement of the slow-moving subject and a 0.5-meter map labeling error. Nodes are generated at equal division points along the zebra crossing centerline in the middle section, at the intersection of the lane boundary and the centerline of the entry direction at the entrance to the non-motorized vehicle lane, and at the intersection of the geometric center of the mixed traffic area and the main traffic direction in the mixed traffic zone.

[0032] Furthermore, directed edges are determined jointly by node adjacency relationships, subject type access rules, and boundary crossing restrictions. The length range of the edges is 0.85 to 1.15 times the length of the centerline between nodes. The historical transfer frequency is obtained by dividing the number of trajectories that transfer from the same starting node to the corresponding directed edge within the same time period over the past 7 days by the total number of all directed edge trajectories that leave the starting node. When the coverage overlap ratio between temporary fencing or road closure areas and existing directed edges reaches more than 30% and persists for 5 consecutive seconds, the corresponding directed edge is triggered to fail and a topology snapshot is rebuilt. The 30% overlap and 5 seconds are used to filter out short-term disturbances and local minor occlusions, respectively.

[0033] Based on the node positions and directed edge ranges in the slow-moving scene topology map, the installation pose of the sensing device, the field of view boundary, the outline of static occlusion facilities, and the area occupied by the dynamic occlusion subject at each sampling time are meshed and projected to obtain the visibility code corresponding to the topology unit. The visibility code distinguishes between fully visible, partially visible, and completely blind zones, and writes the duration, occlusion source category, belonging topology unit, and pointers to adjacent visible units into each coded unit. When the dynamic occlusion subject enters or leaves the directed edge coverage area, the visibility recalculation of the corresponding mesh unit and topology unit is triggered according to the same timestamp, and the time-varying visibility matrix sequence is updated.

[0034] Furthermore, the gridded projection uses a fixed 0.5m × 0.5m grid. First, the field of view boundary is mapped to planar coordinates based on the installation pose of the sensing device. Then, ray clipping is performed on the outline of static occlusion facilities and the area occupied by dynamic occlusion subjects to obtain the visible coverage of each grid. The 0.5m grid side length is set according to the minimum lateral width occupied by the slow-moving subject and the single-cycle position resolution to ensure that visibility coding and topology units are indexed one-to-one. The area occupied by the dynamic occlusion subject is taken from the circumscribed contour projection area of ​​the corresponding subject in the target tracking result stream in the current sampling cycle. Adjacent visible unit pointers are used to record the indices of one or more coding units that are connected to the boundary of the current coding unit and have the highest visible coverage.

[0035] Furthermore, visibility coding is determined jointly based on visibility coverage and continuous observation integrity: A visibility coverage of at least 0.8 with no occlusion for three consecutive sampling periods is coded as fully visible. A visibility coverage greater than 0.2 and less than 0.8 is coded as partially visible. A visibility coverage of no more than 0.2 or with occlusion for three consecutive sampling periods is coded as a complete blind zone. The 0.8 and 0.2 thresholds correspond to the lower limit of high-confidence observations and the upper limit of effective observation failures, respectively.

[0036] Furthermore, for a dynamically occluded subject to enter a directed edge coverage area, the overlap ratio between the projected area of ​​the dynamically occluded subject and the 0.5-meter buffer zone extending outward from the directed edge of the target is not less than 40% for three consecutive frames. Leaving the directed edge coverage area means the overlap ratio is less than 20% for three consecutive frames. After triggering a recalculation, a time-varying visibility matrix is ​​generated according to the raster encoding changes, and each raster, along with its corresponding topological unit, adjacent visible unit pointers, and duration field, is written into the matrix sequence for subsequent dual-memory state retrieval.

[0037] It should be noted that this step takes urban slow-moving traffic electronic data as input and sequentially performs time synchronization, topology graph construction, and time-varying visibility matrix generation. This transforms discrete observation results into indexable, computable, and updatable graph structure data and matrix data, providing a unified data foundation for subsequent dual-memory state maintenance, candidate continuous trajectory inference, and identity association scoring. This reflects the electronic digital data processing approach of object-oriented state organization, spatiotemporal constraint encoding, and matrix-based updating.

[0038] S2: Based on the slow-moving scene topology map and time-varying visibility matrix, establish a dual-memory state and generate candidate continuous trajectories.

[0039] For each slow-moving subject, positive evidence memory, negative evidence memory, current state vector, and identity survival marker are established. Positive evidence memory continuously records the observed position, velocity, orientation, category label, and associated directed edge. Negative evidence memory continuously records expected but unoccurred events, reasonably permissible blind zone units, and prohibited crossing boundary constraints calculated based on the slow-moving scene topology and time-varying visibility matrix. When the same slow-moving subject transitions from fully visible to semi-visible or completely blind zone within a continuous sampling period, the start time of the negative evidence memory is written according to the last observed state, and the expected but unoccurred events count, completely blind zone dwell interval, and identity survival marker are updated on a sampling period basis.

[0040] Furthermore, positive evidence memory is stored using a time-ordered array, including at least the sampling time, planar coordinates, velocity scalar, velocity direction angle, subject category label, directed edge number, and observation confidence value. Negative evidence memory is stored using a constraint event table, including at least the negative evidence memory start time, the event number that should have occurred but did not, the complete blind zone unit number, the prohibited crossing boundary number, the predicted reachable area number, and the cumulative duration period. The current state vector includes at least the current position, current velocity, current orientation, velocity change rate, and the state of the directed edge. The identity survival flag has three levels: alive, pending verification, and inactive, used to drive subsequent candidate continuous trajectory generation and identity association score calculation. The state of the directed edge includes the current directed edge number, the normalized progress value on that directed edge, and the edge state flag indicating whether the edge endpoint has been reached. The normalized progress value is the ratio of the projected distance from the current position to the starting point of the directed edge to the length of the centerline of the directed edge. The observation confidence value is taken from the target existence probability output by the target detection result stream or target tracking result stream in S1. Observation records with an observation confidence value of not less than 0.60 are written into the positive evidence memory. Observation records with an observation confidence value of less than 0.60 are only used to set the identity survival mark of the corresponding slow-moving subject to be verified and do not participate in the current position and directed edge state update.

[0041] Furthermore, the generation rule for events that should have occurred but did not occur is as follows: Starting from the current state vector of the previous sampling period, a predicted reachable region is formed within the range of adjacent directed edges allowed in the slow-moving scene topology map. Then, the predicted reachable region is intersected with the fully visible and semi-visible coding units in the time-varying visibility matrix. When the area of ​​the intersection result is not less than 1.0 square meters and there is no new observed entity of the same category as the slow-moving subject with a positional deviation not exceeding 1.5 meters in the current sampling period, an event that should have occurred but did not occur is generated. The 1.0 square meter threshold is set based on the minimum effective visible area of ​​four consecutive coding units under a 0.5m × 0.5m grid, and the 1.5m threshold is set based on the maximum single-step displacement of the slow-moving subject and the redundancy of the positioning error within a 100-millisecond sampling period.

[0042] Furthermore, the dwell time in the complete blind zone is calculated based on both the subject category and the blind zone topology length. First, the cumulative length of the centerlines of all complete blind zone units in the current directed edge and its adjacent directed edges is read, then divided by the minimum reliable operating speed corresponding to that subject category to obtain the maximum reasonable dwell time. The minimum reliable operating speed for pedestrians is 0.8 m / s, for bicycles and electric bicycles it is 1.5 m / s, and for wheelchairs it is 0.6 m / s, with a uniform 0.5-second observation redundancy time added. When the duration of negative evidence memory exceeds the maximum reasonable dwell time, the identity survival marker is changed from alive to pending verification; if no valid observation is observed for another 0.5 seconds, it is changed to inactive.

[0043] Furthermore, the current state vector is calculated using a limited discrete update method: based on the last observed position, velocity, and orientation in the previous sampling period, the current position and velocity are extrapolated within a 100-millisecond sampling period according to a rule of constant velocity dominance and limited acceleration correction. Specifically, the acceleration limit for pedestrians is ±1.5 m / s², for bicycles and electric bicycles it is ±2.0 m / s², and for wheelchairs it is ±1.0 m / s². The maximum change angle of orientation per period is 25 degrees. These limit values ​​are set based on common microscopic operational characteristics of slow-moving subjects in urban slow-moving scenarios to prevent unreasonable jumps in the current state vector during occlusion periods.

[0044] In the dual-memory state, the last observed position, velocity, orientation, and associated directed edge of the slow-moving subject are read. The motion boundary is calculated according to the slow-moving trajectory propagation time window, and reachable paths are expanded along adjacent directed edges in the slow-moving scene topology map. The reachable paths are intersected with the time-varying visibility matrix at each time step, and path segments that cross prohibited boundaries, enter prohibited directed edges in reverse, exceed the tolerance time of the complete blind zone, or are continuously missing from fully visible or semi-visible coding units are deleted. The remaining path segments are sorted according to positional continuity, directional continuity, and group interaction direction to form a candidate continuous trajectory sequence bound to the dual-memory state.

[0045] Furthermore, the propagation time window for the slow-moving trajectory is fixed at 2.0 seconds and discretely expanded into 20 propagation sub-cycles with a step size of 100 milliseconds. Within each propagation sub-cycle, the forward fan-shaped boundary and the lateral offset boundary are calculated based on the current state vector to form a motion boundary band. The length of the forward fan-shaped boundary is taken as the current velocity multiplied by the propagation sub-cycle and superimposed with the maximum acceleration limit of the corresponding subject category. The lateral offset boundary is fixed at 0.6 meters to cover pedestrian swaying, slight avoidance by bicycles, and wheelchair directional control errors. When the propagation sub-cycle exceeds 10, only the top 3 adjacent directed edges with the highest historical transfer frequency are retained to reduce the number of candidate continuous trajectory branches.

[0046] Furthermore, reachable path unfolding employs a combination of topology-first search and constraint pruning: starting with the target directed edge, it searches forward using the adjacency list for adjacent directed edges. Upon entering a new directed edge, it simultaneously verifies the allowed travel direction, boundary crossing restrictions, and signal phase correlation. If the angle between the path orientation and the allowed direction of the target directed edge is greater than 60 degrees, it is determined to be a reverse entry into a prohibited directed edge and is deleted. If the path crosses any prohibited boundary, the corresponding path segment is immediately deleted. If a path fails to obtain a matching observation for three consecutive sampling periods after entering a fully visible coding unit, or fails to obtain a matching observation for five consecutive sampling periods after entering a semi-visible coding unit, the corresponding path segment is deleted. 60 degrees, three periods, and five periods are used to distinguish between normal turning and reverse movement, effective missing data in fully visible conditions, and fault-tolerant missing data in semi-visible conditions, respectively.

[0047] Furthermore, positional continuity is scored in descending order of the planar distance between the candidate path's end position and the reference position. The reference position is the latest observed position within the last three sampling periods when valid observations exist; otherwise, it is the predicted position extrapolated from the current state vector. A distance of 0.8 meters or less is considered high continuity, and a distance exceeding 2.0 meters is considered low continuity. Orientational continuity is scored in descending order of the angle between the candidate path's end orientation and the current state vector's orientation. An angle of 20 degrees or less is considered high continuity, and an angle exceeding 50 degrees is considered low continuity. Group interaction direction is statistically determined based on the main movement directions of other slow-moving entities within the same topological unit over the last five sampling periods. When the angle between the candidate path direction and this main movement direction is no greater than 30 degrees, the ranking priority is increased. 0.8 meters, 2.0 meters, 20 degrees, 50 degrees, and 30 degrees correspond to the high-confidence matching boundary, low-confidence elimination boundary, stable orientation boundary, abnormal turning boundary, and group co-orientation determination boundary, respectively. When the number of other slow-moving entities available for statistical analysis within the same topological unit is less than two, the orientation in the current state vector is used as the substitute value for the group interaction direction.

[0048] Furthermore, candidate continuous trajectory sequences are written into the dual-memory state using a three-level indexing method: subject identifier, trajectory segment number, and propagation sub-cycle number. Each candidate continuous trajectory records at least the path edge sequence of the trajectory segment, the position of each sub-cycle, the orientation of each sub-cycle, the current visibility encoding status, and the cumulative number of negative evidence. When the number of candidate continuous trajectories corresponding to the same slow-moving subject exceeds 8, only the top 4 in the comprehensive ranking are retained and written into the trajectory cache, while the remaining path segments are deleted to ensure that the computational complexity of subsequent identity association scoring remains stable within a controllable range. The comprehensive ranking is performed by weighted summation of scores for positional continuity, directional continuity, and group interaction direction, and then by a secondary ranking based on the cumulative number of negative evidence from smallest to largest, with weights of 0.45, 0.35, and 0.20 for the three scores, respectively.

[0049] It should be noted that this step uses the slow-moving scene topology map and time-varying visibility matrix as a unified data basis, representing the slow-moving subject as four types of digital objects: positive evidence memory, negative evidence memory, current state vector, and identity survival marker. Then, it generates candidate continuous trajectories through constraint-driven reachable path expansion and periodic pruning. Compared to existing methods that rely solely on single-frame observations or short-term motion continuity, this step can transform expected-but-not-present, completely blind-zone dwellable areas, and prohibited-crossing boundaries into computable constraint data. This allows for the interpretable updating of the continuous trajectory candidate set under conditions of occlusion, intermittent visibility, and dense mixed traffic, which is difficult to achieve stably using conventional tracking links.

[0050] S3: Calculate identity association scores based on candidate continuous trajectories, determine identity correspondence, and output subject-level security status values.

[0051] For each newly observed entity, an association matrix is ​​constructed with each candidate continuous trajectory. The degree of appearance consistency, motion continuity, topological reachability, occlusion consistency, and negative evidence violation are calculated. The degree of negative evidence violation is determined based on whether the newly observed entity can only be located in the corresponding complete blind zone during its disappearance, whether there is a jump across a prohibited boundary, and whether it abnormally reappears after a period of continuous absence of fully visible or semi-visible coding units. After normalizing each degree to the same dimension, a weighted combination is performed. The matching result with the highest score and satisfying the mutual exclusion assignment constraint is used as the identity correspondence. Unmatched observed entities and unmatched candidate continuous trajectories are written into the new entity processing branch and the continue-maintain branch, respectively.

[0052] Furthermore, the association matrix is ​​stored using a two-dimensional matrix structure: the number of newly observed entities multiplied by the number of candidate continuous trajectories. Each element in the matrix corresponds to a set of association score records between observed entities and candidate continuous trajectories. Newly observed entities include at least the planar coordinates, velocity, orientation, category label, appearance feature vector, and observation confidence value for the current sampling period. Candidate continuous trajectories include at least the trajectory end position, trajectory end orientation, the directed edge to which the end belongs, the end visibility encoding status, and the cumulative number of negative evidence. When the observation confidence value of a newly observed entity is below 0.60, it only participates in low-priority matching and does not participate in the determination of high-priority identity correspondence.

[0053] Furthermore, the appearance consistency is calculated by taking the cosine similarity between the appearance feature vector of the newly observed entity and the appearance feature vector of the last valid observation of the candidate continuous trajectory. Motion continuity is calculated by taking the joint score of the distance deviation and velocity deviation between the current position of the newly observed entity and the predicted position of the end of the candidate continuous trajectory. Topological reachability is calculated by taking the existence of an allowed path and its path length deviation between the directed edge to the end of the candidate continuous trajectory and the current directed edge to the newly observed entity. Occlusion consistency is calculated by taking the degree of matching between the visibility encoding state during the disappearance of the candidate continuous trajectory and the visibility encoding state around the location of the newly observed entity. All these degrees are uniformly mapped to a range of 0 to 1, where 0 represents complete inconsistency and 1 represents complete consistency. The appearance feature vector is generated by inputting the size-normalized outer region of the corresponding slow-moving subject in the target detection result stream into the re-identification feature extraction network; the feature dimension is preferably 128 or 256.

[0054] Furthermore, the degree of violation of negative evidence is calculated as follows:

[0055] in, Indicates the degree to which negative evidence contradicts the evidence. This represents the blind zone violation, used to characterize whether a newly observed entity has a trajectory interpretation outside the corresponding complete blind zone during its disappearance. It is set to 1 when there is at least one non-blind zone reachable path with a length of not less than 1.0 meter; otherwise, it is taken as the ratio of the non-blind zone path length to the total reachable path length. This represents the boundary jump variable, used to characterize whether there is a jump that crosses a prohibited boundary. It is set to 1 when the path must cross at least one prohibited boundary to reach the current observation position, and 0 otherwise. The visible missing violation is used to characterize the degree to which the candidate continuous trajectory reappears abnormally after being continuously missing in fully visible or semi-visible coding units. It is taken as the ratio of the number of missing duration periods to the upper limit of the allowed missing periods, with the upper limit being 3 sampling periods in fully visible coding units and 5 sampling periods in semi-visible coding units. The weighting coefficients for blind zone violation, boundary jump, and visible missing are respectively, with preferred values ​​of 0.35, 0.40, and 0.25.

[0056] The formula for calculating the degree of violation of negative evidence is used to transform the three types of negative evidence constraints into a single violation quantity, which is then directly written into the corresponding element of the association matrix. When When the value is greater than 0.65, the observed entities and candidate continuous trajectories in this group will not participate in the formal matching, but will only be retained as candidates to be verified, in order to avoid unreasonable identity jumps after occlusion recovery.

[0057] Furthermore, after normalizing each degree using the same dimension, the final association score is formed by subtracting the negative evidence violation penalty value from a weighted positive combination of appearance consistency, motion continuity, topological reachability, and occlusion consistency. Only when the final association score is not lower than 0.55 does the mutual exclusion assignment constraint solution begin. The mutual exclusion assignment constraint is implemented using a block-based Hungarian assignment method: first, local candidate blocks are divided according to topological units, and then a one-to-one optimal assignment is performed within each candidate block. If the same candidate continuous trajectory is simultaneously competed for by multiple newly appearing observed entities, the group with the highest final association score and the lowest negative evidence violation is prioritized. The 0.55 threshold is set based on the empirical median after unified normalization of the five degree categories and the upper limit of the negative evidence penalty, used to isolate random nearest neighbor matching.

[0058] Furthermore, the directed edge to which a newly observed entity currently belongs is determined by projecting its current planar coordinates onto the centerline of the nearest directed edge that allows passage for that entity category. When multiple candidate directed edges exist, the directed edge with the smallest projection distance and the smallest angle between the passage directions is selected first.

[0059] After the identity correspondence is determined, the conflict approximation quantity, occlusion uncertainty quantity, topology deviation quantity, and mixed traffic compression quantity are calculated based on the continuous trajectory, directed edge, current visibility code, relative position relationship of adjacent traffic participants, and mixed traffic occupancy relationship of the corresponding slow-moving subject, as determined by the identity correspondence. These quantities are then cumulatively updated according to a unified time base. When any quantity undergoes a cross-level change, continuous missing data recovery, or directed edge switching, the subject-level safety state value is recalculated, and the current state record bound to the slow-moving subject is output according to the slow-moving safety state mapping rules.

[0060] Furthermore, the conflict approach quantity is calculated based on the predicted minimum distance and relative approach speed between the corresponding slow-moving subject and adjacent traffic participants. A high conflict approach is defined as when the predicted minimum distance is no higher than 1.5 meters and the relative approach speed is no lower than 1.0 meter per second. The predicted minimum distance is obtained by extrapolating the positions of the current subject and adjacent traffic participants over the next second in 100-millisecond steps and taking the minimum Euclidean distance. The relative approach speed is the relative speed component along the connecting line between the two subjects. The occlusion uncertainty is calculated based on the current visibility coding state, the duration of continuous complete blind spots, and the cumulative number of negative evidence. The topology deviation is calculated based on the lateral offset between the centerline of the continuous trajectory and the centerline of its associated directed edge, and whether the directed edge is occupied in reverse. The mixed-traffic compression is calculated based on the number of subjects per unit area and the relative traffic direction dispersion within the mixed-traffic area. The above thresholds correspond to the effective safety boundary, the upper limit of occlusion duration, and the upper limit of path deviation under close-range interaction between slow-moving subjects, respectively.

[0061] Furthermore, adjacent traffic participants are limited to the directed edge to which the current subject belongs and other traffic participants within 3 meters of the current subject on the adjacent directed edge, and whose timestamps are the same. The number of subjects per unit area in the mixed traffic area is the ratio of the number of subjects within the range to the effective mixed traffic area, and the relative traffic direction dispersion is the ratio of the standard deviation of the direction angle of each subject within the range to 180 degrees.

[0062] Furthermore, the calculation of the main-level safety state value is expressed as follows:

[0063] in, This indicates the security status value at the subject level. This represents the normalized result of the conflict approximation quantity. This represents the normalized result of the occlusion uncertainty. This represents the normalized result of the topological deviation. This represents the normalized result of the mixed-line compression. The memory penalty is determined by combining the cumulative number of negative evidence against the current subject with the number of consecutive missing recovery attempts, and is used to reflect the stability of identity continuity during the occlusion period. The memory penalty is determined by the weighted sum of the current subject's cumulative number of negative evidence divided by 8 and the number of consecutive missing recovery attempts divided by 3, and is truncated to the range of 0 to 1, with the preferred weights for the two items being 0.7 and 0.3. These are the weighting coefficients for the five items mentioned above, with preferred values ​​of 0.30, 0.20, 0.20, 0.15, and 0.15.

[0064] This formula is used to convert identity mapping results into continuously updatable subject-level security state values, which serve as the core numerical field in the output state record. When the same subject experiences occlusion recovery, topology switching, or sudden changes in mixed-row density over multiple consecutive sampling periods, it can maintain continuous state updates under a unified dimension.

[0065] Furthermore, the subject-level security status value is cumulatively updated based on a unified time base of 100 milliseconds, and a hysteresis mapping rule is set: when When it is less than 0.30, it is mapped to a safe state. A value not less than 0.30 and less than 0.55 is mapped to a state of interest. A value not less than 0.55 and less than 0.80 is mapped to a warning state. A value of 0.80 or higher is mapped to a high-risk state. When the state value rises across levels, the state record is updated immediately. When the state value falls across levels, the lower-level condition must be met for three consecutive sampling periods before the state record is updated, in order to suppress frequent state jitter caused by single-cycle noise.

[0066] Furthermore, continuous loss recovery refers to the process by which candidate continuous trajectories that are pending verification or continue to maintain branches regain a valid identity correspondence within the time limit of the complete blind zone. Directed edge switching refers to the process by which the normalized progress value at the end of the continuous trajectory corresponding to the slow-moving subject reaches 1 and enters the next allowed directed edge. When triggering the recalculation of the subject-level security state value, in addition to recalculating each component of the current sampling period, the conflict approximation amount and occlusion uncertainty amount of the previous two sampling periods are also backtracked simultaneously to generate a smoothed current state record. This state record, along with the subject identifier, the current directed edge number, the current visibility encoding state, and the timestamp, is written into the state output table for subsequent time-series queries and state tracking.

[0067] It should be noted that this step takes candidate continuous trajectories as input, first constructs an association matrix between observed entities and candidate trajectories, and simultaneously encodes five types of constraints—appearance, motion, topology, occlusion, and negative evidence—in the matrix elements. Then, identity correspondence is obtained through mutual exclusion assignment. Subsequently, continuous trajectories, visibility encoding, and adjacent entity relationships are further mapped into subject-level security state values ​​with unified dimensions. Essentially, this step performs matrix-based calculations, normalization fusion, and temporal updates of multi-source spatiotemporal data, object states, and constraints. Compared to existing methods that rely solely on instantaneous appearance matching or short-term trajectory extension, this step can maintain consistent updates of identity correspondence and state output under occlusion recovery, topology switching, and mixed-trajectory disturbances, which is difficult to achieve stably in conventional tracking links.

[0068] Example 2, an embodiment of the present invention, provides a safety status monitoring system for urban slow-moving traffic, including a scene topology visible modeling module, a dual-memory continuous trajectory module, and an identity-associated status output module.

[0069] The scene topology visibility modeling module is used to synchronize electronic data of urban slow traffic, construct a slow traffic scene topology map and generate a time-varying visibility matrix.

[0070] The dual-memory continuous trajectory module is used to establish dual-memory states and generate candidate continuous trajectories based on the topology map of the slow-moving scene and the time-varying visibility matrix.

[0071] The identity association status output module is used to calculate the identity association score based on the candidate continuous trajectory, determine the identity correspondence, and output the subject-level security status value.

Claims

1. A method for monitoring the safety status of urban slow-moving traffic, characterized in that, include: Synchronize electronic data of urban slow-moving traffic, construct a topology map of slow-moving scenarios, and generate a time-varying visibility matrix; Based on the topology map of the slow-moving scene and the time-varying visibility matrix, a dual-memory state is established and candidate continuous trajectories are generated. The identity association score is calculated based on the candidate continuous trajectory, the identity correspondence is determined, and the subject-level security status value is output.

2. The method for monitoring the safety status of urban slow-moving traffic as described in claim 1, characterized in that: The construction of the slow-moving scene topology map includes... The synchronized urban slow traffic electronic data includes zebra crossing entrances, zebra crossing middle sections, waiting areas, non-motorized vehicle lane entrances, mixed motorized and non-motorized traffic areas, road edge boundary connection points, and pedestrian crossing exit areas, which are mapped as nodes. The passable relationships between nodes are mapped as directed edges. According to the slow-moving subject category, write the allowed passage direction, length range, signal phase correlation, boundary crossing restriction and historical transfer frequency for each directed edge, and build a node index table, edge index table, adjacency table and topology snapshot sequence; When temporary barriers, road closures, or changes in traffic organization appear in the electronic data of urban slow-moving traffic, reconstruction is performed on the affected nodes and directed edges, and the corresponding topology snapshot is updated.

3. The method for monitoring the safety status of urban slow-moving traffic as described in claim 2, characterized in that: The generation of the time-varying visibility matrix includes, Based on the node positions and directed edge ranges in the slow-moving scene topology map, the installation pose of the sensing device, the field of view boundary, the outline of the static occlusion facility, and the area occupied by the dynamic occlusion subject at each sampling time are meshed and projected to obtain the visibility code corresponding to the topology unit. Visibility coding distinguishes between fully visible, partially visible, and completely blind zones, and writes the duration, occlusion source category, topological unit to which it belongs, and pointers to adjacent visible units to each coding unit; When a dynamically occluding subject enters or leaves the directed edge coverage area, the visibility of the corresponding grid cell and topology cell is recalculated according to the same timestamp, and the time-varying visibility matrix sequence is updated.

4. The method for monitoring the safety status of urban slow-moving traffic as described in claim 3, characterized in that: The establishment of the dual-memory state includes, For each slow-moving subject, establish positive evidence memory, negative evidence memory, current state vector and identity survival marker. Positive evidence memory continuously records the observed position, speed, orientation, category label and the directed edge to which it belongs. Negative evidence memory continuously records events that should have occurred but did not occur, complete blind zone units where reasonable stopping is possible, and boundary constraints that prohibit crossing, which are calculated based on the slow-moving scene topology map and time-varying visibility matrix. When the same slow-moving subject changes from fully visible to semi-visible or completely blind zone within a continuous sampling period, the negative evidence memory start time is written according to the last observation state, and the count of subjects that should have appeared but did not appear, the dwell interval in the completely blind zone, and the identity survival marker are updated in units of sampling period.

5. The method for monitoring the safety status of urban slow-moving traffic as described in claim 4, characterized in that: The generation of candidate continuous trajectories includes, In the dual-memory state, read the last observed position, velocity, orientation and directed edge of the slow-moving subject, calculate the motion boundary according to the propagation time window of the slow-moving trajectory, and expand the reachable path along the adjacent directed edge in the topology map of the slow-moving scene; The reachable path is intersected with the time-varying visibility matrix at each time step, and path segments that cross forbidden boundaries, enter into prohibited directed edges in reverse, exceed the duration of the complete blind zone, or are continuously missing from the fully visible or semi-visible coding unit are deleted. The remaining path segments are sorted according to positional continuity, directional continuity, and group interaction direction to form a candidate continuous trajectory sequence bound to the dual-memory state.

6. The method for monitoring the safety status of urban slow-moving traffic as described in claim 5, characterized in that: The calculation of identity association score based on candidate continuous trajectory includes... For each newly observed entity, an association matrix is ​​constructed between it and each candidate continuous trajectory. The degree of consistency in appearance, the degree of continuity in motion, the degree of topological reachability, the degree of consistency in occlusion, and the degree of violation of negative evidence are calculated respectively. The degree of violation of negative evidence is determined based on whether the newly observed entity can only be located in the corresponding complete blind zone during the disappearance period, whether there is a jump across the prohibited boundary, and whether it appears abnormally after the complete or semi-visible coding unit has been missing for a long time. After normalizing each degree to the same dimension, a weighted combination is performed. The matching result with the highest score and satisfying the mutual exclusion assignment constraint is used as the identity correspondence. The unmatched observed entities and unmatched candidate continuous trajectories are written into the new subject processing branch and the continue to maintain branch, respectively.

7. The method for monitoring the safety status of urban slow-moving traffic as described in claim 6, characterized in that: The process of determining the identity correspondence and outputting the subject-level security status value includes, After the identity correspondence is determined, the conflict approach quantity, occlusion uncertainty quantity, topology deviation quantity and mixed traffic compression quantity are calculated according to the continuous trajectory, directed edge, current visibility code, relative position relationship of adjacent traffic participants and mixed traffic occupation relationship of the corresponding slow-moving subject determined by the identity correspondence, and the conflict approach quantity, occlusion uncertainty quantity, topology deviation quantity and mixed traffic compression quantity are cumulatively updated according to a unified time base. When any quantity undergoes a cross-level change, continuous missing recovery, or directed edge switching, the subject-level security state value is recalculated, and the current state record bound to the slow-moving subject is output according to the slow-moving security state mapping rules.

8. A safety status monitoring system for urban slow-moving traffic, employing the safety status monitoring method for urban slow-moving traffic as described in any one of claims 1 to 7, characterized in that: Includes a scene topology visible modeling module, a dual-memory continuous trajectory module, and an identity-associated state output module; The scene topology visibility modeling module is used to synchronize urban slow traffic electronic data, construct a slow traffic scene topology map and generate a time-varying visibility matrix; The dual-memory continuous trajectory module is used to establish a dual-memory state and generate candidate continuous trajectories based on the slow-moving scene topology map and the time-varying visibility matrix. The identity association status output module is used to calculate the identity association score based on the candidate continuous trajectory, determine the identity correspondence, and output the subject-level security status value.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the urban slow-moving traffic safety status monitoring method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the urban slow-moving traffic safety status monitoring method according to any one of claims 1 to 7.

Images