A subway limit invasion early warning method and system based on trajectory prediction and risk mapping

Abstract

The application relates to the technical field of subway construction safety monitoring, and discloses a subway limit invasion early warning method and system based on track prediction and risk mapping. The method comprises the following steps: a three-dimensional model of an early warning area is constructed and divided into multiple early warning sub-areas, and a visual scene is built; laser radar collection point cloud data is laid out, and denoising, registration and standardization preprocessing are carried out; based on a detection-tracking framework, a 3D target detection algorithm and an adaptive volume Kalman filtering algorithm are used to identify and track limit invasion targets, and future motion tracks and limit invasion ranges of the targets are predicted; a dynamic adaptive weighted fusion model is constructed, foreign body static risks, motion dynamic risks and line environment related risks are fused, target comprehensive risk values are calculated, and early warning grades are divided; early warning data is superimposed and displayed with a three-dimensional scene, and interactive operation is supported. The application realizes accurate identification, advanced early warning and dynamic risk assessment of limit invasion targets, and improves the safety of subway construction.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and early warning technology for subway safety, specifically relating to a subway encroachment warning method and system based on trajectory prediction and risk mapping. Background Technology

[0002] With the continuous expansion of urban rail transit networks, the construction environment for subways is becoming increasingly complex. Subway construction typically takes place near existing lines or in densely populated urban areas. Construction machinery, temporary facilities, and external operations can easily encroach on the subway's safety clearance, posing a serious threat to the operational safety of existing lines and the structural safety of projects under construction. Therefore, monitoring and providing early warnings of encroachments to subway construction areas are crucial means to ensure both construction and operational safety.

[0003] In existing technologies, monitoring and early warning for subway construction encroachments largely rely on video surveillance, manual inspections, or single-sensor threshold alarms. However, these technologies have many shortcomings: video surveillance is easily affected by factors such as changes in lighting, inclement weather, and obstructions, and its monitoring effectiveness drops significantly at night and in rainy or foggy conditions, making it difficult to achieve uninterrupted monitoring around the clock; manual inspections have limited frequency, making it difficult to achieve uninterrupted monitoring around the clock, and they rely on personnel experience, resulting in a high rate of missed detections and failing to meet the needs of real-time early warnings; traditional single sensors can only achieve linear or point detection, resulting in a large number of monitoring blind spots, and they cannot obtain three-dimensional spatial information of the encroaching target, making it difficult to accurately perceive the target's size, position, and movement status.

[0004] Furthermore, the aforementioned technologies generally lack the ability to predict the future movement trends of intruding targets, failing to achieve true advance warning. This forces relevant personnel to react passively only after an intrusion occurs, making it difficult to take preventative measures in advance, resulting in poor early warning effectiveness. Simultaneously, existing risk assessment methods mostly employ static weights, which cannot adapt to dynamic changes in on-site conditions. Factors such as weather changes, construction phase transitions, and changes in train operation status are not effectively incorporated into the risk level assessment system, leading to discrepancies between assessment results and actual conditions. Traditional early warning level classifications are also relatively coarse, lacking detailed visual indicators and tiered response mechanisms, which hinders rapid decision-making in emergency response.

[0005] Therefore, there is an urgent need for an intrusion warning method that can achieve all-weather, high-precision, and full-coverage monitoring, and has the functions of trajectory prediction and dynamic risk assessment. Summary of the Invention

[0006] This invention aims to solve the problems of poor environmental adaptability, insufficient spatial perception and lack of advanced prediction capability in existing subway construction encroachment monitoring and early warning technologies, and provides a subway encroachment early warning method and system based on trajectory prediction and risk mapping.

[0007] A subway encroachment warning method based on trajectory prediction and risk mapping, characterized by the following steps: S1. Create a 3D model of the subway construction early warning area, divide the subway early warning area into multi-level early warning sub-areas, and build a visualization scene based on the BIM model of subway stations and sections. S2. Deploy lidar within the subway early warning area to collect point cloud data and perform preprocessing; S3. The preprocessed point cloud data is analyzed and processed using an adaptive capacitive Kalman filter algorithm to identify and track intruding targets within the subway warning area; historical motion parameters of the intruding targets are extracted, a kinematic model of the intruding targets is established, and the target's trajectory is predicted based on the historical motion parameters to determine the possible intrusion range of the target. S4. Based on the possible intrusion range of the target, calculate the comprehensive risk value of the intruding target, and automatically classify the intruding target according to the comprehensive risk value of the target to determine the corresponding early warning level; S5. Push the warning data, including the warning level, target information, movement trajectory and risk assessment results, to the user display interface in real time, and overlay the warning information with the scene. The target comprehensive risk value mentioned in step S4 is calculated using a dynamic adaptive weighted fusion model. The sub-risk items of this model are divided into three categories: static risk value of the foreign object itself, dynamic risk value of the foreign object's movement, and risk value associated with the line environment. The dynamic adaptive weighted fusion model is implemented through a dynamic adjustment factor, and the calculation method for this dynamic adjustment factor is as follows: in, Dynamic adjustment factor For sub-risk items, the risk value is... , For indicator functions; For instantaneous high-risk incentive coefficient, This represents the incentive coefficient for trend abrupt change. To accumulate the continuous incentive coefficient, The high-risk threshold The change in sub-risk value, For high rate of change threshold, To maintain a high level for a prolonged period, The threshold for inertial activation. This is the threshold for attention.

[0008] Further, in step S1, high-precision 3D modeling technology is used to model the subway construction early warning area, and the hierarchical analysis method is used to divide the early warning sub-regions. The on-site geological conditions (C1), structural risk level (C2), and external operation impact level (C3) are the first-level evaluation indicators. The on-site geological conditions (C1) include soil and rock stability (C11), groundwater distribution (C12), and adverse geological impact (C13). The structural risk level (C2) includes structural type (C21), construction stage (C22), and stress state (C23). The external operation impact level (C3) includes construction machinery operation intensity (C31), vehicle traffic density (C32), and personnel activity frequency (C33). The early warning area includes three-level early warning sub-regions: core area, warning area, and monitoring area.

[0009] Furthermore, in step S2, lidar equipment is deployed along the warning area and the deployment location covers the entire subway warning area; point cloud data is collected in real time and the collected raw point cloud data is uploaded to the platform system in real time; the raw point cloud data uploaded to the platform system is preprocessed, including removing invalid noise points, registering point cloud coordinates, unifying the coordinate reference of the point cloud data, and standardizing the point cloud data.

[0010] Furthermore, in step S2, the preprocessing includes: removing invalid noise points through a combination of statistical filtering and radius filtering algorithms, and achieving point cloud coordinate registration through an iterative nearest point algorithm.

[0011] Further, in step S3, the PointRCNN point cloud 3D target detection algorithm is selected to extract the target depth features and contour information, and to obtain candidate detection states including the target's three-dimensional coordinates, contour size, appearance features, and confidence level; a three-dimensional constant velocity angular rate motion model is used to estimate the motion state of the intruding target, and the prediction results are smoothed and optimized by combining the capacitive Kalman filter algorithm; an affinity model that integrates the target's appearance features, geometric parameters, and spatial distance correlation is constructed, and a greedy algorithm is used to solve for the optimal matching pair; an adaptive capacitive Kalman filter algorithm is used to dynamically update the tracking state of the intruding target by real-time correction of the innovation sequence and dynamic adjustment of the filter parameters, and to record the target's motion speed, motion direction, and acceleration.

[0012] Furthermore, in step S4, the static risk value of the foreign object body includes volume, material type, and intrusion location; the dynamic risk value of the foreign object movement includes movement speed and intrusion duration; and the risk value associated with the subway environment is set in combination with the characteristics of the subway scene's subway environment.

[0013] Furthermore, in step S5, the user display interface supports a variety of real-time interactive operations, including rapid location of warning points, target trajectory tracing, warning details query, and multi-view review.

[0014] The present invention also provides a subway encroachment warning system based on trajectory prediction and risk mapping, which uses the above method to issue warnings for subway encroachments.

[0015] The embodiments of the present invention have the following technical effects: 1) This invention constructs a multi-layered three-dimensional model of the early warning area, integrates three-dimensional laser scanning technology and BIM technology to achieve millimeter-level modeling of the early warning area, divides the core area, warning area and monitoring area, and combines immersive visualization simulation scene and differentiated color markings to realize intuitive presentation and efficient interaction of risk areas, which greatly improves the accuracy and intuitiveness of subway construction early warning area management and solves the problems of vague division and low visualization of traditional early warning areas.

[0016] 2) This invention adopts a combination of optimized deployment of high-resolution lidar sensors and multi-step data preprocessing. Through joint noise reduction of statistical filtering and radius filtering, coordinate registration of ICP algorithm, and standardized format conversion, invalid noise points are effectively removed, coordinate reference and data format are unified, ensuring the integrity, accuracy and standardization of point cloud data, providing high-quality data support for subsequent identification and tracking of intruding targets, and significantly improving data utilization efficiency and the accuracy of subsequent algorithm processing.

[0017] 3) Based on the detection-tracking framework, this invention selects the PointRCNN point cloud 3D target detection algorithm and the adaptive capacitive Kalman filter algorithm, combined with a three-dimensional constant velocity angular rate motion model, a fusion affinity model and bipartite graph matching technology, to achieve accurate identification, continuous and stable tracking and motion trajectory prediction of intruding targets. It effectively resists interference from complex working conditions such as dust, occlusion and long distance, and solves the defects of traditional tracking algorithms such as easy drift, missed detection and low recognition accuracy, thus improving the reliability and real-time performance of intruding target monitoring.

[0018] 4) This invention constructs a multi-level early warning classification system based on a dynamic adaptive weighted fusion model, introduces a dynamic weight adjustment mechanism, and combines three sub-risk items—static risk of foreign objects, dynamic risk, and line environment-related risk—to achieve accurate quantification of comprehensive risk. It automatically classifies four early warning levels and performs visual labeling in red, orange, yellow, and green colors, solving the problem that traditional static weights cannot adapt to changes in on-site working conditions, improving the accuracy and adaptability of risk assessment, and providing a scientific basis for emergency response decisions.

[0019] 5) This invention is guided by intelligence, realizing real-time push of early warning data to multiple terminals, accurate overlay of early warning information with three-dimensional scenes, and providing convenient interactive functions such as rapid location of early warning points, target trajectory backtracking, early warning details query and multi-view review. It replaces the traditional manual investigation mode, greatly improves the efficiency of early warning information transmission and emergency response speed, reduces the intensity of manual labor, and ensures the safety and reliability of subway construction. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the subway encroachment warning method based on trajectory prediction and risk mapping provided in this embodiment of the invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are part of this invention. Specific Implementation Example 1 This invention provides a subway encroachment warning method based on trajectory prediction and risk mapping. This method integrates lidar three-dimensional point cloud perception, target trajectory tracking prediction and dynamic risk mapping to achieve all-weather monitoring and graded early warning of subway construction encroachment.

[0024] like Figure 1 As shown, the method of the present invention includes the following steps: S1. Create a 3D model of the subway construction early warning area, divide the subway early warning area into multi-level early warning sub-areas, and build a visualization scene based on the BIM model of subway stations and sections. S2. Deploy lidar within the subway early warning area to collect point cloud data and perform preprocessing; S3. The preprocessed point cloud data is analyzed and processed using an adaptive capacitive Kalman filter algorithm to identify and track intruding targets within the subway warning area; historical motion parameters of the intruding targets are extracted, a kinematic model of the intruding targets is established, and the target's trajectory is predicted based on the historical motion parameters to determine the possible intrusion range of the target. S4. Based on the possible intrusion range of the target, calculate the comprehensive risk value of the intruding target, and automatically classify the intruding target according to the comprehensive risk value of the target to determine the corresponding early warning level; S5. Push the warning data, including the warning level, target information, movement trajectory and risk assessment results, to the user display interface in real time, and overlay the warning information with the scene. The target comprehensive risk value mentioned in step S4 is calculated using a dynamic adaptive weighted fusion model. The sub-risk items of this model are divided into three categories: static risk value of the foreign object itself, dynamic risk value of the foreign object's movement, and risk value associated with the line environment. The dynamic adaptive weighted fusion model is implemented through a dynamic adjustment factor, and the calculation method for this dynamic adjustment factor is as follows: in, Dynamic adjustment factor For sub-risk items, the risk value is... , For indicator functions; For instantaneous high-risk incentive coefficient, This represents the incentive coefficient for trend abrupt change. To accumulate the continuous incentive coefficient, The high-risk threshold The change in sub-risk value, For high rate of change threshold, To maintain a high level for a prolonged period, The threshold for inertial activation. This is the threshold for attention.

[0025] Further, in step S1, high-precision 3D modeling technology is used to model the subway construction early warning area, and the hierarchical analysis method is used to divide the early warning sub-regions. The on-site geological conditions (C1), structural risk level (C2), and external operation impact level (C3) are the first-level evaluation indicators. The on-site geological conditions (C1) include soil and rock stability (C11), groundwater distribution (C12), and adverse geological impact (C13). The structural risk level (C2) includes structural type (C21), construction stage (C22), and stress state (C23). The external operation impact level (C3) includes construction machinery operation intensity (C31), vehicle traffic density (C32), and personnel activity frequency (C33). The early warning area includes three-level early warning sub-regions: core area, warning area, and monitoring area.

[0026] Furthermore, in step S2, lidar equipment is deployed along the warning area and the deployment location covers the entire subway warning area; point cloud data is collected in real time and the collected raw point cloud data is uploaded to the platform system in real time; the raw point cloud data uploaded to the platform system is preprocessed, including removing invalid noise points, registering point cloud coordinates, unifying the coordinate reference of the point cloud data, and standardizing the point cloud data.

[0027] Furthermore, in step S2, the preprocessing includes: removing invalid noise points through a combination of statistical filtering and radius filtering algorithms, and achieving point cloud coordinate registration through an iterative nearest point algorithm.

[0028] Further, in step S3, the PointRCNN point cloud 3D target detection algorithm is selected to extract the target depth features and contour information, and to obtain candidate detection states including the target's three-dimensional coordinates, contour size, appearance features, and confidence level; a three-dimensional constant velocity angular rate motion model is used to estimate the motion state of the intruding target, and the prediction results are smoothed and optimized by combining the capacitive Kalman filter algorithm; an affinity model that integrates the target's appearance features, geometric parameters, and spatial distance correlation is constructed, and a greedy algorithm is used to solve for the optimal matching pair; an adaptive capacitive Kalman filter algorithm is used to dynamically update the tracking state of the intruding target by real-time correction of the innovation sequence and dynamic adjustment of the filter parameters, and to record the target's motion speed, motion direction, and acceleration.

[0029] Furthermore, in step S4, the static risk value of the foreign object body includes volume, material type, and intrusion location; the dynamic risk value of the foreign object movement includes movement speed and intrusion duration; and the risk value associated with the subway environment is set in combination with the characteristics of the subway scene's subway environment.

[0030] Furthermore, in step S5, the user display interface supports a variety of real-time interactive operations, including rapid location of warning points, target trajectory tracing, warning details query, and multi-view review.

[0031] The present invention also provides a subway encroachment warning system based on trajectory prediction and risk mapping, which uses the above method to issue warnings for subway encroachments. Specific Implementation Example 2 A subway encroachment warning method based on trajectory prediction and risk mapping includes the following steps: Step S1: Construct a multi-level 3D model of the early warning area. High-precision 3D reconstruction is performed on the subway construction early warning area. Taking into account geological conditions, structural risk levels, and the impact of nearby external work disturbances, the early warning space is divided into three progressive levels: a core domain, a warning domain, and a monitoring domain. A lightweight BIM model is imported and integrated into a 3D rendering engine to build a high-fidelity visualization simulation scene. Each sub-region is distinguished by differentiated colors and unique identifier codes, supporting interactive operations such as scene roaming, sectioning, and object picking.

[0033] Step S2: LiDAR Point Cloud Data Acquisition and Preprocessing. Multiple LiDAR sensors are optimally deployed along the warning area to acquire 3D point cloud data of the surrounding environment in real time and upload it to the edge processing node. Preprocessing steps such as statistical filtering for noise reduction, multi-source point cloud spatiotemporal registration, and spatial standardization are sequentially performed on the raw point cloud to generate high-quality structured point cloud data, providing a reliable data foundation for subsequent intrusion analysis and target identification.

[0034] Step S3: Intrusion target identification, tracking, and trajectory prediction. Based on point cloud clustering and feature recognition algorithms, heterogeneous targets in the intrusion warning area are detected. An adaptive capacitive Kalman filter algorithm is used to recursively estimate the state of multiple targets and continuously track them, acquiring key state information such as target category, 3D spatial position, and velocity vector in real time. Historical motion parameters are extracted and a kinematic extrapolation model is established to predict the target's future trajectory and potential intrusion range.

[0035] Step S4: Dynamic Risk Mapping and Early Warning Level Classification. A multi-level hierarchical early warning classification system is constructed, introducing a dynamic adaptive weighted fusion model to collaboratively assess the static attribute risk of foreign objects, the dynamic risk of target movement, and the associated risk from the railway line environment. Based on this, the comprehensive risk value of the target is quantitatively calculated. Early warnings are divided into four levels according to the comprehensive risk value of the target, and visualized and labeled using red, orange, yellow, and green colors. For suspected risk targets that have not yet been clearly identified, they are initially labeled in gray, and their category attributes and early warning levels are dynamically updated after manual confirmation.

[0036] Step S5: Visualization and Interactive Operation. The integrated risk information is pushed to the front end of the visualization platform via real-time data channels, achieving precise overlay display with the 3D subway track scene; it provides functions such as spatial positioning of warning points, target trajectory tracing, attribute detail query, and free multi-view viewing, assisting emergency command personnel in conducting analysis and decision-making in a realistic 3D environment.

[0037] Specifically, in step S1, the core of constructing a multi-level early warning area 3D model is to achieve accurate digital replication of the early warning area and risk-layered management. Its technical process covers four key links: high-precision modeling, regional layer division, immersive scene construction, and visualization annotation. Each link is closely connected to ensure the practicality and intuitiveness of the 3D model.

[0038] High-precision 3D modeling relies on detailed survey data of the subway construction early warning area, including geological survey reports, construction design drawings, and on-site measurement data. It adopts a combination of 3D laser scanning technology and BIM modeling technology to conduct a comprehensive scan and modeling of subway stations, tunnels, surrounding buildings, underground pipelines, and topography within the early warning area. The modeling accuracy is controlled at the millimeter level to ensure that the model can accurately reproduce the actual working conditions on site, providing a precise spatial carrier for subsequent risk monitoring and early warning.

[0039] After completing the basic 3D modeling, the analytic hierarchy process (AHP) was used to divide the warning areas into sub-regions, taking into account the actual site conditions. During the division, an evaluation index system was established: site geological conditions (C1), structural risk level (C2), and external operation impact level (C3) were selected as primary evaluation indicators. Each primary indicator had secondary sub-indicators. Specifically, C1 included soil and rock stability (C11), groundwater distribution (C12), and adverse geological impacts (C13); C2 included structural type (C21), construction stage (C22), and stress state (C23); and C3 included construction machinery operation intensity (C31), vehicle traffic density (C32), and personnel activity frequency (C33). Construct a judgment matrix A, and use the 1-9 scaling method to determine the weight of each indicator. The judgment matrix has the following form: in, For the first The first indicator and the first The importance ratio of each indicator satisfies 1 indicates equal importance, and 9 indicates extreme importance; Calculate the weight vector The largest eigenvalue of the judgment matrix is ​​found by using the eigenvalue method. And the corresponding feature vectors, after normalization, yield the weight vector. ,in ; The consistency check is calculated using the following formula: in, To evaluate the number of indicators (in this embodiment) ), The average random consistency index ( hour ),Require This ensures the consistency of the judgment matrix; The scores of each indicator are quantified, and each secondary sub-indicator is scored on a 10-point scale. The overall regional risk value is then calculated by weighted summation. : in, , , These are the weights of each secondary sub-indicator; regions are divided according to the comprehensive risk value: For the core domain, For the warning zone, This is the monitoring domain.

[0040] To achieve intuitive and visual display of early warning information, an immersive visualization simulation scenario needs to be built. Specifically, based on the complete BIM model of the subway station and its sections, lightweight processing is first performed. This is done by removing redundant model components, simplifying model details, and compressing model data to reduce the model's memory footprint and ensure that the model can be quickly imported into a high-precision 3D rendering engine. In the 3D rendering engine, materials are applied to the model, lighting and shadow rendering is performed, and the environment is simulated to recreate the realistic scene's lighting, terrain, and surrounding environment, thus building an immersive simulation environment. This scenario supports multi-dimensional interactive operations for users, including real-time roaming, zooming, rotating, and clicking to query.

[0041] To visually represent the risk levels of different warning sub-regions, differentiated colors and labels are required for each sub-region. Adopting industry-standard risk color coding and considering the warning requirements of this patent, the core area is labeled red, representing extremely high risk requiring focused monitoring and real-time warnings; the alert area is labeled orange, representing high risk requiring enhanced monitoring and timely warnings; and the monitoring area is labeled yellow, representing general risk requiring routine monitoring and periodic checks. Simultaneously, clear markings are placed at the boundaries of each sub-region, indicating the sub-region name, risk level, and monitoring focus, ensuring users can quickly distinguish the risk differences between areas and intuitively grasp the overall risk distribution of the warning area.

[0042] Specifically, in step S2, the acquisition and preprocessing of lidar point cloud data is the foundation for accurate identification and tracking of intruding targets. Its core objective is to acquire complete, accurate, and standardized three-dimensional point cloud data to provide reliable support for subsequent data processing and analysis. The entire process is divided into three core stages: sensor deployment, data acquisition, and data preprocessing. Each stage requires strict control of technical parameters to ensure data quality.

[0043] During the sensor deployment phase, high-resolution lidar sensors should be rationally deployed based on the actual conditions of the subway early warning area, including its topography, structural layout, and monitoring range. The selection of lidar sensors must meet the early warning monitoring requirements, prioritizing high-resolution lidar with high measurement accuracy, fast response speed, and strong anti-interference capabilities. These sensors should have a measurement range of 0.3-100m, a point cloud density of 100 points / cm², a sampling frequency of 10Hz, a measurement accuracy of ±2cm, an anti-interference rating of IP67, and an operating temperature range of -40℃ to 85℃.

[0044] The sensor placement is optimized using the principle of "multi-point deployment and cross-coverage". The formula for calculating the sensor placement spacing D is as follows: in, This is the maximum measurement range of the sensor. This represents the sensor's horizontal field of view.

[0045] Sensors should be deployed at high points, key corners, and around core areas along the warning zone to ensure complete coverage and eliminate blind spots. For complex terrain or obstructed areas, the number of sensors can be increased, or their height and angle adjusted, to ensure comprehensive monitoring. Furthermore, after deployment, sensors must be calibrated to ensure consistent measurement accuracy, laying the foundation for subsequent point cloud data registration.

[0046] The lidar sensor collects high-density 3D point cloud data of the surrounding construction environment in real time according to a preset sampling frequency. This includes 3D spatial information of various targets within the warning area, such as buildings, structures, construction machinery, personnel, and foreign objects. During the collection process, the sensor uploads the raw point cloud data to the platform system in real time via wired or wireless transmission. The transmission process uses an encryption protocol to ensure data security and integrity, preventing data loss or tampering. Simultaneously, the system monitors the data collection status in real time. If data collection is interrupted or abnormal, it promptly issues an alert signal, allowing staff to quickly troubleshoot and ensure continuous and stable data collection.

[0047] Raw point cloud data contains a large number of invalid noise points caused by environmental interference and equipment errors. Furthermore, point cloud data collected by multiple sensors suffers from inconsistencies in coordinates and formats. Therefore, targeted preprocessing of the raw point cloud data is necessary. The preprocessing process mainly includes three steps: First, noise removal, employing a combination of statistical filtering and radius filtering. Statistical filtering calculates the number and distance of neighboring points for each point, eliminating noise points that are too far away or have an abnormal number. Radial filtering sets a reasonable radius threshold to eliminate isolated noise points with insufficient neighboring points within the radius range. The combination of these two methods effectively removes invalid noise points and retains valid point cloud data. Second, point cloud registration, using an iterative nearest-point algorithm to register point cloud coordinates between multiple lidar sensors. This algorithm, through iterative optimization, aligns point cloud data collected by different sensors in the same spatial coordinate system, unifying the coordinate reference and ensuring spatial consistency of the point cloud data. Third, data standardization, uniformly processing point cloud data with different sampling frequencies and formats, converting the point cloud data format to a standard format supported by the platform system, and adjusting the sampling frequency to ensure consistent sampling intervals for all point cloud data, achieving unified data format standardization. The preprocessed point cloud data is complete, accurate, and standardized, and can effectively support subsequent work on intruding target identification, tracking, and trajectory prediction.

[0048] Statistical filtering calculates the average distance between the neighboring points of each point. and standard deviation The formula is as follows: , in, Number of neighboring points (set) ), For the current point and the first The distance between neighboring points, when When this happens, it is identified as noise and removed; Among them, radius filtering sets a radius threshold. When a point is within radius If the number of neighboring points within the range is less than 3, it is judged as an isolated noise point and removed; The point cloud registration process employs an iterative nearest-point algorithm to unify point cloud data collected by multiple sensors into the same spatial coordinate system. The specific iterative formula is as follows: (1) Objective function: (2) Solve for the rotation matrix Translation vector , so that the objective function Minimize, the iteration termination condition is: ,in For the first The objective function value of the next iteration; Among these measures, data standardization converts the point cloud data format to the LAS 1.4 standard format, and uniformly adjusts the sampling frequency to 10Hz, with a sampling interval of... ,in To ensure that the data format is consistent with the sampling frequency, the sampling frequency must be standardized.

[0049] Specifically, in step S3, a target recognition, tracking, and trajectory prediction algorithm is designed based on the detection-tracking framework, and the state vector of the intruding target is: The state vector consists of position, velocity, and angular velocity. The modeling of the intrusion target tracking system comprises two main parts: a predictive model to estimate the target's motion state and a measurement model to observe the target's current motion state. Its state-space model is as follows: in, and These are the state transition function and the measurement function, respectively, and the sampling time is the scanning interval of the lidar sensor. for Real-time tracking status, For 3D object detection algorithms to obtain Continuously monitor the status. and These are the process noise and measurement noise vectors of the system, respectively.

[0050] Assuming the speed at which the intruding target moves along a spatial arc remains constant, i.e. angular velocity A constant-rate angular-rate model is used to establish the kinematic model of the intrusion target: in, , .

[0051] An adaptive capacitive Kalman filter algorithm is used to optimize the predicted state to obtain higher prediction accuracy. The volume point is calculated based on the third-order spherical radial volume criterion, and then, by substituting the system nonlinear function, the volume point is propagated according to the three-dimensional constant-rate angular-rate model. calculate Using the predicted motion state value and the prediction error covariance matrix at time 1, state prediction is performed: in, Indicates the volume point. This represents the dimension of the tracking state vector. This represents the prior estimate of the state vector of the intruding target. This is for prior estimation of the covariance matrix.

[0052] An affinity model is constructed that integrates target appearance features, geometric parameters, and spatial distance. The matching degree between the predicted state and the candidate detection state is calculated using the following formula: in, For appearance-related, For geometric correlation, For distance correlation, the calculation methods are as follows: Among them, new information , This is the regularization function.

[0053] Maintaining the identity and temporal continuity of targets through data association ensures that the same target can be continuously and correctly tracked in time, thus serializing isolated detections and forming stable trajectories. To this end, based on an affinity model between predicted and detected states, the association problem is formalized as bipartite graph matching. A target trajectory is assigned to each candidate detection, and the total cost defined by the affinity model is minimized to obtain the optimal inter-frame matching result. To solve the minimum weight matching problem in a bipartite graph, a greedy algorithm is used. After data association is completed, a series of matching pairs can be obtained. This represents the near-optimal association between the predicted state and the candidate detected state. Meanwhile, unmatched detected states are used to initialize new trajectories. Furthermore, the association process may generate predicted states that fail to match due to the target naturally disappearing or being missed. To distinguish between these two cases, a threshold is set. If continuous If a predicted state within a frame cannot match any detected state, the target is determined to have disappeared naturally, and its trajectory is deleted; otherwise, the predicted state is considered a missed detection and will continue to participate in data association in the next moment.

[0054] The innovation covariance representation of volumetric Kalman filtering reflects the impact of unknown errors: The innovation error covariance matrix can be estimated as follows: in, This is a weighting factor; adjust the weights accordingly.

[0055] To address the unknown biases caused by inaccuracies in the motion model and the statistical characteristics of system noise, and to improve the tracking accuracy of intruding targets, an adaptive mechanism is introduced into the capacitive Kalman filter algorithm. This mechanism utilizes an adaptive factor... Corrected error covariance matrix: in, This represents the trace of the covariance matrix.

[0056] The volume points are calculated based on the corrected prediction error covariance matrix and the prediction state: Calculate the new information covariance matrix: According to the measurement model, volumetric point propagation can be expressed as: Calculate the prior detection state: The cross-covariance matrix and the covariance matrix are as follows: Finally, the equations for updating the posterior estimation of the tracking state and the error covariance matrix are expressed as follows: Wherein, the Kalman gain matrix is , The motion state vector of the intruding target.

[0057] Intrusion range determination is divided into real-time intrusion determination and intrusion prediction. The motion state of the intruding target is calculated based on adaptive capacitive Kalman filtering. The system determines the spatial range of the intruding target by using three-dimensional position coordinates, and uses three-dimensional velocity and angular velocity to assist in risk prediction.

[0058] Among them, the real-time intrusion limit determination is based on the inclusion judgment algorithm of spatial points and polyhedra, and selects the reference plane of the three-dimensional coordinate system of the warning area (such as...). (Horizontal plane) to determine the three-dimensional position of the target Projection matching is performed on the boundary polyhedra corresponding to the core domain, warning domain, and monitoring domain defined in step 1, and the number of intersections between the target point and the polyhedron boundary is calculated using the ray casting method: in, For the regional boundary vertex coordinates For sign functions. When The number is odd, and the target coordinates satisfy the region boundary. If the axis range is specified, then the target falls into that area.

[0059] Intrusion prediction is based on the three-dimensional constant velocity and angular velocity motion model in step 3 to predict the target's future. After determining the trajectory of the movement within a time period, the aforementioned spatial point and polyhedron inclusion judgment algorithm is used to determine whether the target falls into the corresponding area.

[0060] Specifically, in step S4, a dynamic adaptive weighted fusion model is constructed to achieve comprehensive risk quantification and early warning level classification of intrusion targets, and three types of sub-risk items are constructed, namely, the static risk of the foreign object itself. Dynamic risks of foreign object movement Risks associated with the line environment The quantification methods for each sub-risk item are as follows: Static risks of foreign objects : in, The target volume is scored (1-10 points, the larger the volume, the higher the score). Score based on material type (8-10 points for metals, 1-7 points for non-metals). Scoring is given for the intrusion location (core domain 10 points, vigilance domain 6-9 points, monitoring domain 1-5 points). Dynamic risks of foreign object movement : in, Score the target speed (1-10 points, the faster the speed, the higher the score, v≥10m / s gets 10 points). The score is based on the duration of the violation (1-10 points, the longer the duration, the higher the score, t≥30s gets 10 points). Line environment associated risks : in, The subway operation status is scored (8-10 points during operating hours, 1-7 points during non-operating hours). The score is based on the construction conditions (8-10 points for high-risk operations, 1-7 points for routine operations). Score environmental factors (severe weather 8-10 points, normal weather 1-7 points).

[0061] To address the limitation of static weights in adapting to changes in operating conditions, a dynamic weight adjustment mechanism is introduced, enabling the weights to be automatically adjusted according to the actual on-site conditions. in, After normalization, it satisfies , To dynamically adjust the factors, when a sub-risk value suddenly increases beyond a threshold, its weight is automatically amplified. The weights of the corresponding sensor risk items are adjusted according to environmental changes, and "inertial weighting" is applied to items with persistently high risk to prevent misjudgment due to instantaneous fluctuations. The calculation method is as follows: in, , This is an indicator function.

[0062] This is the instantaneous high-risk incentive coefficient, a gain coefficient used to adjust the contribution intensity of instantaneous high risk to the dynamic adjustment factor. It is dimensionless and ranges from [0, 1], with a typical value of 0.5 to 1.0. The larger this coefficient is, the stronger the response to a single sub-risk exceeding the high-risk threshold. This is the trend change excitation coefficient, used to adjust the gain coefficient of the contribution of the rapid deterioration trend of the risk value to the dynamic adjustment factor. It is dimensionless and ranges from [0, 1], with a typical value of 0.3 to 0.6. The larger the coefficient, the more sensitive it is to the response to a sharp increase in the risk value in a short period of time. The cumulative sustained incentive coefficient is a gain coefficient used to adjust the contribution intensity of the cumulative effect of the sustained high level of sub-risk to the dynamic adjustment factor. It is dimensionless and ranges from [0, 1], with a typical value of 0.1 to 0.3. The larger the coefficient, the more significant the weighting of targets that are in a high-risk state for a long time. The high-risk threshold is a pre-set sub-risk value used to determine high risk. It is dimensionless, ranges from 0 to 1, and typically has a value of 0.7. At that time, high-risk incentives are activated instantaneously; Let be the change in sub-risk value, representing the difference between the i-th sub-risk indicator at the current time t and the previous time t-1. , dimensionless, ranges from [-1, 1], positive values ​​indicate increased risk, negative values ​​indicate decreased risk; The high rate of change threshold is a pre-set threshold for determining high risk based on the change in sub-risk values. It is dimensionless, ranges from (0, 1], and typically takes the value 0.2 / sampling period. When the trend change stimulus is activated; The duration of sustained high value represents the i-th sub-risk value. Continuously greater than or equal to the attention threshold The duration is expressed in units of sampling periods; once Falling back to the following, Reset immediately and start accumulating again; The inertial activation threshold is a pre-set minimum effective threshold for the duration of continuous high-level states, expressed in the number of sampling periods, with a typical value of 3. When T_i ≥ T_min, the cumulative continuous excitation term is activated. Its function is to filter out short-term instantaneous fluctuations and avoid frequent false triggers. The attention threshold is a pre-set threshold for determining the continuous attention of sub-risk values. It is dimensionless, ranges from (0,1), and typically takes the value of 0.5. At that time, the system begins to accumulate the duration of the sustained high level of this sub-risk.

[0063] The various coefficients can be automatically adjusted according to the on-site environment. For example, when the noise of the lidar increases in rainy or foggy weather, the coefficients can be temporarily reduced. Coefficients are used to suppress spurious trends.

[0064] Specifically, in step S5, warning data such as the warning level, target information, movement trajectory, and risk assessment results are pushed in real time to the display interfaces of relevant personnel through the platform system, including the monitoring center's large screen, staff's mobile phones, tablets, and other terminals, ensuring that relevant personnel can obtain the core warning content in a timely manner. During the push process, a tiered push mechanism is adopted: Level 1 and Level 2 warnings use pop-up notifications and sound alarms to ensure staff are aware of them immediately; Level 3 and Level 4 warnings use message pushes to remind staff to pay attention. At the same time, the frequency of warning data pushes is consistent with the data collection frequency to ensure the real-time nature of the warning information. The pushed content includes the specific location of the intruding target, the warning level, risk parameters, and movement trajectory prediction results, providing staff with comprehensive warning information support.

[0065] Based on the 3D simulation scene built in step S1, the warning information is precisely overlaid with the 3D scene of the subway track, achieving an intuitive and visual display of the location, movement status, and warning level of the intruding target. During the overlay process, coordinate alignment technology is used to unify the spatial coordinates of the warning data with the coordinate system of the 3D simulation scene, ensuring that the location of the intruding target is accurately presented in the scene, with a deviation of no more than 5 centimeters from the actual location on site. Specifically, in the 3D scene, the intruding target is presented as a highlighted model, with its color consistent with the warning level color. At the same time, key information such as the warning level, target name, and movement speed are marked around the target. The target's movement trajectory is presented as dynamic lines, with different transparency levels used to distinguish the trajectory at different time periods, making it easy for staff to view the target's movement process. The risk assessment results are presented in a floating window; clicking on the target allows viewing detailed risk parameters and assessment basis, providing an intuitive understanding of the target's risk status.

[0066] To facilitate rapid investigation and emergency response by staff, the system supports multiple real-time interactive operations, each highly targeted and user-friendly. First, it offers rapid early warning point location: staff can click on an early warning message on the display interface, and the system can instantly pinpoint the location of the intruding target, automatically adjusting the 3D scene view to focus on the target area, eliminating the need for manual searching and significantly improving investigation efficiency. Second, it provides target trajectory playback: staff can select any intruding target and replay its entire movement process from entering the warning area to the current moment, with playback accuracy down to 0.1 seconds. This assists staff in analyzing the movement patterns and causes of the intrusion, providing a basis for subsequent prevention and control measures. Third, it offers early warning details query: staff can click on an intruding target or early warning message to view various risk parameters (such as volume, material, and movement speed), risk assessment basis, and early warning level classification process, gaining a comprehensive understanding of the target's risk status. Fourth, it provides multi-view inspection: staff can switch between different perspectives (such as overhead, side, front, and close-up views) to comprehensively inspect the situation on-site, clearly viewing the relative positional relationship between the target and surrounding structures and equipment, providing decision support for quickly developing emergency response plans.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A subway encroachment warning method based on trajectory prediction and risk mapping, characterized in that, Includes the following steps: S1. Create a 3D model of the subway construction early warning area, divide the subway early warning area into multi-level early warning sub-areas, and build a visualization scene based on the BIM model of subway stations and sections. S2. Deploy lidar within the subway early warning area to collect point cloud data and perform preprocessing; S3. The preprocessed point cloud data is analyzed and processed using an adaptive capacitive Kalman filter algorithm to identify and track intruding targets within the subway warning area. Extract the historical motion parameters of the intruding target, establish a kinematic model of the intruding target, and combine the historical motion parameters to predict the target's trajectory and determine the possible intrusion range of the target; S4. Based on the possible intrusion range of the target, calculate the comprehensive risk value of the intruding target, and automatically classify the intruding target according to the comprehensive risk value of the target to determine the corresponding early warning level; S5. Push the warning data, including the warning level, target information, movement trajectory and risk assessment results, to the user display interface in real time, and overlay the warning information with the scene. The target comprehensive risk value mentioned in step S4 is calculated using a dynamic adaptive weighted fusion model. The sub-risk items of the dynamic adaptive weighted fusion model are divided into three categories: static risk value of the foreign object itself, dynamic risk value of the foreign object's movement, and risk value associated with the line environment. The dynamic adaptive weighted fusion model is implemented through dynamic adjustment factors, wherein the calculation method of the dynamic adjustment factors is as follows: ; in, Dynamic adjustment factor For sub-risk items, risk value , For indicator functions; For instantaneous high-risk incentive coefficient, This represents the stimulus coefficient for trend abrupt change. To accumulate the continuous incentive coefficient, The high-risk threshold The change in sub-risk value, For high rate of change threshold, To maintain a high level for a prolonged period, The threshold for inertial activation. This is the threshold for attention.

2. The subway encroachment warning method based on trajectory prediction and risk mapping according to claim 1, characterized in that, In step S1, high-precision 3D modeling technology is used to model the subway construction early warning area, and the hierarchical analysis method is used to divide the early warning sub-regions. On-site geological conditions, structural risk level, and external operation impact level are the first-level evaluation indicators. The on-site geological conditions include the stability of soil and rock mass, groundwater distribution, and adverse geological effects. The structural risk level includes the structural type, construction stage, and stress state. The external operation impact level includes the intensity of construction machinery operation, vehicle traffic density, and personnel activity frequency. The early warning area includes three-level early warning sub-regions: core area, warning area, and monitoring area.

3. The subway encroachment warning method based on trajectory prediction and risk mapping according to claim 1, characterized in that, In step S2, lidar equipment is deployed along the warning area and the deployment location covers the entire subway warning area; Real-time acquisition of point cloud data, and real-time uploading of the acquired raw point cloud data to the platform system; The raw point cloud data uploaded to the platform system is preprocessed. The preprocessing process includes removing invalid noise points, registering point cloud coordinates, unifying the coordinate reference of the point cloud data, and standardizing the point cloud data.

4. The subway encroachment warning method based on trajectory prediction and risk mapping according to claim 3, characterized in that, In step S2, the preprocessing includes: removing invalid noise points by combining statistical filtering and radius filtering algorithms, and achieving point cloud coordinate registration by iterative nearest point algorithm.

5. The subway encroachment warning method based on trajectory prediction and risk mapping according to claim 1, characterized in that, In step S3, the PointRCNN point cloud 3D target detection algorithm is selected to extract the target depth features and contour information, and to obtain candidate detection states including the target's three-dimensional coordinates, contour size, appearance features, and confidence level. A three-dimensional constant velocity angular rate motion model is used to estimate the motion state of the intruding target, and the prediction results are smoothed and optimized by combining the capacitive Kalman filter algorithm. An affinity model that integrates the target's appearance features, geometric parameters, and spatial distance correlation is constructed, and a greedy algorithm is used to solve for the optimal matching pair. An adaptive capacitive Kalman filter algorithm is used to dynamically update the tracking state of the intruding target by real-time correction of the innovation sequence and dynamic adjustment of the filter parameters, and to record the target's motion speed, motion direction, and acceleration.

6. The subway encroachment warning method based on trajectory prediction and risk mapping according to claim 1, characterized in that, In step S4, the static risk value of the foreign object body includes volume, material type, and intrusion location; the dynamic risk value of the foreign object movement includes movement speed and intrusion duration; and the risk value associated with the subway environment is set in combination with the characteristics of the subway scene's subway environment.

7. The subway encroachment warning method based on trajectory prediction and risk mapping according to claim 1, characterized in that, In step S5, the user interface supports a variety of real-time interactive operations, including rapid location of warning points, target trajectory tracing, warning details query, and multi-view review.

8. A subway encroachment warning system based on trajectory prediction and risk mapping, characterized in that, The method described in any one of claims 1-7 is used to provide early warning of subway encroachment.

Images