Spatial syntax-based traffic station pedestrian path selection prediction method and system

Abstract

The invention discloses a traffic station pedestrian path selection prediction method and system based on a space syntax, and the method comprises the steps: carrying out the matching of dynamic pedestrian trajectory data and a static station space structure based on the same coordinate system, and constructing a target traffic station mapping database; performing full-dimensional parameter extraction and screening removal processing on the target traffic station mapping database through a spatial syntax to obtain key spatial syntax features; and constructing a hierarchical heterogeneous nested Logit decision model, and performing pedestrian path selection prediction on the key space syntactic features to obtain a pedestrian path selection prediction result. According to the method, the pedestrian path selection prediction precision of the traffic station can be improved through the quantitative relationship between the space structure in the rail traffic station and the pedestrian behavior. The traffic station pedestrian path selection prediction method and system based on the space syntax can be widely applied to the technical field of traffic station pedestrian planning.

Description

Technical Field

[0001] This invention relates to the field of pedestrian planning technology at transportation hubs, and more particularly to a method and system for predicting pedestrian path selection at transportation hubs based on space syntax. Background Technology

[0002] As key nodes in modern urban transportation systems, the efficient and safe operation of urban rail transit stations is crucial for ensuring the overall functioning of the city. Rail transit stations typically exhibit highly dynamic passenger flow characteristics, with passengers moving in various directions and experiencing large volumes of traffic. This presents significant challenges to traffic management and space optimization within the stations.

[0003] Extensive research has been conducted using methods such as video capture, simulation modeling, and questionnaires to study pedestrian walking speed and spatial needs, emergency evacuation exit selection behavior, evacuation behavior and psychological reactions, and facility layout and evacuation efficiency, striving to ensure the smooth operation of urban rail transit systems and a positive passenger experience. Among these studies, discrete choice models and their extended models, based on utility maximization theory, are widely used due to their solid theoretical foundation and explanatory power. Scholars have used these models to explore numerous influencing factors. For example, at the strategic level, they focus on the trade-off between activity type and destination; at the tactical level, they focus on external environmental factors such as shortest physical distance, estimated travel time, and the number of turns along the route. Internal factors such as individual attributes (e.g., age, familiarity with the environment) and travel purpose also play a crucial role throughout the research.

[0004] However, these mainstream models have a fundamental limitation when dealing with large-scale, structurally complex built environments such as subway stations: they tend to abstract and simplify physical space into a "node-arc" topological network. While this simplification improves the computational efficiency of the model to some extent, it may also lead to the neglect of the geometry of the space itself, the complexity of the topological structure, and, more importantly, the pedestrian visual environment experience. This directly results in insufficient exploration of the deeper "space-society" relationship behind pedestrian path selection behavior.

[0005] Space syntax, as a theory for studying the relationship between space and human behavior, offers a new perspective for analyzing pedestrian path selection. Based on graph theory and network analysis, space syntax transforms complex spaces into analyzable networks, revealing the topological properties and accessibility of spaces through parameters such as integration and selectivity. It has achieved great success in macro-level urban studies such as urban planning, transportation planning, and the preservation of traditional villages and historic districts. However, empirical research on space syntax for micro-level, complex, small-scale pedestrian spaces such as rail transit stations is relatively limited. Related techniques have also attempted to apply space syntax to analyze the intrinsic connection between spatial structure and pedestrian behavior, but have not deeply revealed the intrinsic relationship between spatial structure and pedestrian path selection.

[0006] Most existing research focuses on verifying the macroscopic correlation between spatial syntactic parameters and observed passenger flow distribution, lacking the integration of these spatial quantitative indicators as endogenous variables into pedestrian path selection models. This results in insufficient explanatory correlation between spatial syntactic features and behavioral models. This lack of interpretability limits the practical application of research findings and also affects the accurate understanding and prediction of pedestrian behavior in subway stations. Summary of the Invention

[0007] To address the aforementioned technical problems, the present invention aims to provide a method and system for predicting pedestrian path selection at transportation stations based on spatial syntax. This method and system can improve the accuracy of pedestrian path selection prediction at transportation stations by leveraging the quantitative relationship between the spatial structure within the rail transit station and pedestrian behavior.

[0008] The first technical solution adopted in this invention is: a pedestrian path selection prediction method for traffic stations based on space syntax, comprising the following steps: Based on the same coordinate system, dynamic pedestrian trajectory data is matched with static station spatial structure to construct a target transportation station mapping database; By using spatial syntax to extract parameters from the target traffic station mapping database in all dimensions and perform filtering and elimination processes, key spatial syntax features are obtained. A hierarchical and heterogeneous nested Logit decision model is constructed to predict pedestrian path selection based on key spatial syntactic features, and the prediction results are obtained.

[0009] Furthermore, the step of matching dynamic pedestrian trajectory data with static station spatial structures under the same coordinate system to construct a target transportation station mapping database specifically includes: Obtain the CAD plan of the target transportation station, and divide the CAD plan of the target transportation station into several grid units based on ergonomics and the average stride length of adults. Based on the CAD plan of the divided target traffic stations, key nodes and path segments are extracted, and the start and end points of path selection are defined to construct the network topology diagram of the target traffic stations. The system acquires pedestrian trajectory videos from cameras at the target transportation station, treats each pedestrian as a point mass, and tracks their trajectory with their right heel as the center to obtain the position coordinates of each pedestrian in the target measurement area. The position coordinates of each pedestrian in the target measurement area are converted into actual physical plane coordinates by using an affine transformation matrix. Then, a fuzzy matching algorithm is used to map the actual physical plane coordinates to the network topology diagram of the target traffic station, thus constructing a target traffic station mapping database.

[0010] Furthermore, the step of extracting and filtering parameters from the target traffic station mapping database using spatial syntax to obtain key spatial syntax features specifically includes: Based on the Depthmap software, spatial syntactic analysis was performed on the target traffic station mapping database to extract parameters and obtain initial spatial syntactic features. The initial spatial syntactic features are screened by feature gain using the XGBoost regression model, and the features with a cumulative contribution rate greater than a preset value are retained to form a candidate set. Based on the candidate set, features exhibiting directional paradoxes and features with variance inflation factors greater than a preset value are removed to obtain key space syntactic features.

[0011] Furthermore, the spatial syntactic analysis of the target traffic station mapping database specifically includes convex space analysis, axis analysis, and view domain analysis. The convex space analysis is used to identify continuous passable areas, the axis analysis is used to extract the main path skeleton, and the view domain analysis is used to calculate the reachable range of the line of sight.

[0012] Furthermore, the key spatial syntax features specifically include convex space selectivity, axis control degree squared, axis length squared, field integration degree square root, and step shortest path length squared.

[0013] Furthermore, the hierarchical heterogeneous nested Logit decision model specifically includes an upper-layer policy network and a lower-layer path network, wherein: The upper-layer policy network is used to simulate the macro-classification and pre-screening of travel strategies by pedestrians based on spatial perception. It includes three mutually exclusive policy nodes: structure priority nest, time optimization nest, and non-optimal nest. The structure priority nest corresponds to the set of main paths with good visual accessibility, the time optimization nest corresponds to the set of shortcuts with short physical distance, and the non-optimal nest corresponds to the set of random walk paths. The lower-level path network is the micro-execution layer for decision-making, which includes several specific physical path nodes; An inclusion coefficient is constructed by quantifying the overall utility contribution of lower-level path groups to upper-level strategy nodes, and then the upper-level strategy nodes and lower-level path nodes are connected.

[0014] Furthermore, the step of constructing a hierarchical heterogeneous nested Logit decision model to predict pedestrian path selection based on key spatial syntactic features and obtain the pedestrian path selection prediction result specifically includes: A hierarchical heterogeneous nested Logit decision model is constructed by inputting key spatial syntactic features into the hierarchical heterogeneous nested Logit decision model. The lower-level path network based on the hierarchical heterogeneous nested Logit decision model performs linear weighting operations on the key space syntactic features and preset weight coefficients to obtain the deterministic utility value of each specific path. By using a logarithmic summation algorithm, the utility values ​​of all lower-level paths belonging to the same policy node are non-linearly aggregated to calculate the coverage value of that policy node. The upper-level policy network based on the hierarchical heterogeneous nested Logit decision model calculates the marginal probabilities of various policies, including structure priority, time optimization, and non-optimal, based on the inclusion value of the policy node and the Logit probability transformation formula. Conditional probability processing is applied to the marginal probabilities of various strategies to obtain the conditional selection probability of each specific path within the nest; The pedestrian path selection prediction result is obtained by multiplying the marginal probabilities of various strategies with the conditional selection probabilities of each specific path within the nest.

[0015] The second technical solution adopted in this invention is: a pedestrian path selection prediction system for traffic stations based on space syntax, comprising: The first module is used to match dynamic pedestrian trajectory data with static station spatial structure under the same coordinate system to build a target transportation station mapping database. The second module is used to extract and filter parameters from the target traffic station mapping database in all dimensions using spatial syntax to obtain key spatial syntax features. The third module is used to construct a hierarchical heterogeneous nested Logit decision model to predict pedestrian path selection based on key spatial syntactic features, and obtain the pedestrian path selection prediction results.

[0016] The beneficial effects of the method and system of this invention are as follows: This invention constructs a target transportation station mapping database by matching dynamic pedestrian trajectory data with static station spatial structure under the same coordinate system; further, it extracts and filters parameters from the target transportation station mapping database in all dimensions through spatial syntax to obtain key spatial syntax features, clarifying the significant quantitative relationship between parameters such as vision integration degree in spatial syntax and pedestrian path selection behavior, and improving the robustness of the model; finally, it constructs a hierarchical heterogeneous nested Logit decision model to predict pedestrian path selection based on key spatial syntax features, obtaining pedestrian path selection prediction results. The core parameters of spatial syntax selected by machine learning are used as input variables, and the path selection probability of the nested Logit model is used as the prediction output, thereby establishing a quantitative relationship between the spatial structure within rail transit stations and pedestrian behavior, improving the scientific nature of passenger flow distribution prediction in the station planning and operation management stages. Attached Figure Description

[0017] Figure 1 This is a flowchart of the steps of the pedestrian path selection prediction method at traffic stations based on space syntax in this invention; Figure 2 This is a structural block diagram of the pedestrian path selection prediction system for traffic stations based on space syntax, as described in this invention. Figure 3 This is a schematic diagram of pedestrian path selection prediction provided in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the spatial layout and grid division of a subway station provided in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the station network topology provided in a specific embodiment of the present invention; Figure 6 This is a schematic diagram of the implementation process for quantification and screening of spatial syntactic feature parameters provided in a specific embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the relationship between visual integration and pedestrian path in a subway station concourse, provided in a specific embodiment of the present invention. Figure 8 This is a schematic diagram illustrating the relationship between pedestrian path and visual integration degree provided in a specific embodiment of the present invention; Figure 9 This is a schematic diagram of the ranking of the importance (Gain value) of convex space features provided in a specific embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of a hierarchical heterogeneous nested Logit decision model provided in a specific embodiment of the present invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0019] First, the technical terminology of this invention needs to be explained: 1) Discrete choice model: Based on the principle of utility maximization, it calculates the utility value and probability of a pedestrian choosing each path in a finite number of mutually exclusive paths.

[0020] 2) Space Syntax: A method for calculating the spatial structural relationships between roads within rail transit stations based on graph theory.

[0021] 3) Spatial syntactic feature parameters: Spatial feature parameters that significantly affect pedestrian path selection within urban rail transit stations.

[0022] 4) Spatial syntactic feature parameter selection method: First, spatial analysis is performed to obtain the feature parameter set. The principle is to correspond to the three types of behavior in pedestrian path selection: "pedestrian dwell time, pedestrian trajectory ratio, and pedestrian turning frequency". Then, the parameters are preprocessed by data standardization, correlation analysis, nonlinear transformation and preliminary screening. Next, the XGBoost feature selection algorithm model is selected. Through the triple evaluation mechanism of feature importance assessment, SHAP parsing and collinearity test, the final selection results are determined.

[0023] In the planning and operation management of urban rail transit, how to accurately predict the path selection behavior of pedestrians inside the station and evaluate the rationality of the facility layout has always been a core issue in traffic simulation and hub design.

[0024] Discrete choice models (especially Logit models and their variants) are widely used in route selection analysis in the transportation field due to their strong interpretability and high predictive accuracy. This method, based on utility maximization theory, analyzes the core influencing factors of route selection within subway stations. It assumes that pedestrians, as decision-makers, choose the "optimal" or "satisfactory" option from a finite number of alternative routes, and evaluates the model's rationality and the accuracy of its predictions using model evaluation indicators. This method transforms the subjective and objective factors influencing pedestrian route selection into measurable indicators. Existing research typically focuses on factors such as pedestrians' personal attributes, travel characteristics, route length, and psychological state, with less emphasis on the spatial layout within the station.

[0025] With the development of spatial perception and environmental psychology, research has found that pedestrians' path selection behavior within stations is also influenced by the station's spatial layout and morphological changes. In information-rich and structurally complex environments like rail transit stations, the visual characteristics of space are key factors affecting pedestrians' wayfinding efficiency, psychological comfort, and sense of security. Space syntax, as a theory studying the relationship between space and human behavior, provides a new perspective for analyzing pedestrian path selection behavior through quantitative analysis of spatial structure.

[0026] Accurate prediction of pedestrian path selection behavior is a core challenge in the planning and operation analysis of urban rail transit stations. Although discrete choice models (such as the Logit model) are widely used, existing technologies suffer from the following significant drawbacks: First, insufficient quantification of spatial features and a lack of coupling with microscopic perception. Existing models often simplify complex station spaces into a "node-arc" topological network, considering only surface parameters such as physical distance and number of turns. This approach ignores visual perception features such as visual field integration and spatial closure, leading to significant discrepancies between model predictions and reality in areas with obstructed views or complex spatial structures. Second, the selection of feature indicators is subjective, lacking a scientific screening mechanism. When constructing utility functions, existing technologies often rely on researchers' subjective experience for variable selection, making it difficult to accurately identify core influencing factors from a vast amount of spatial morphological parameters. This not only easily introduces multicollinearity problems but also fails to quantify the specific contribution of each spatial feature to the decision. Finally, single decision assumptions do not conform to real-world logic. Traditional Logit models typically assume that each path option is independent of the others, failing to reflect the hierarchical decision-making mechanism of pedestrians who "first determine the macro strategy and then choose the specific path," leading to distorted predictions of the probability of choosing similar paths.

[0027] Based on this, such as Figure 3 As shown, the purpose of this embodiment is to construct a method for predicting pedestrian path selection at urban rail transit stations based on spatial syntactic feature filtering and a nested Logit model. This embodiment constructs a coupled model between syntactic feature parameters describing the microstructure of the station's space and pedestrian path selection behavior. The core spatial syntactic parameters, filtered through machine learning, are used as input variables, and the path selection probability from the nested Logit model is used as the prediction output. This establishes a quantitative relationship between the spatial structure within rail transit stations and pedestrian behavior, improving the scientific rigor of passenger flow distribution prediction during station planning and operation management.

[0028] Reference Figure 1 This invention provides a pedestrian path selection prediction method for traffic stations based on space syntax, which includes the following steps: S100. Based on the same coordinate system, the dynamic pedestrian trajectory data is matched with the static station spatial structure to construct a target transportation station mapping database. Specifically, the CAD plan of the target transportation station is obtained, and divided into several grid cells according to ergonomics and the average stride length of adults. Based on the divided CAD plan of the target transportation station, key nodes and path segments are extracted, and the start and end points of path selection are defined to construct the network topology diagram of the target transportation station. Pedestrian trajectory videos are acquired from the cameras at the target transportation station, and pedestrians are treated as point masses, with the right heel as the center for trajectory tracking, to obtain the position coordinates of each pedestrian in the target measurement area. The position coordinates of each pedestrian in the target measurement area are converted into actual physical plane coordinates through an affine transformation matrix, and the actual physical plane coordinates are mapped to the network topology diagram of the target transportation station using a fuzzy matching algorithm to construct a target transportation station mapping database.

[0029] In this embodiment, the primary task in constructing a high-precision prediction model is to solve the problem of accurately matching dynamic pedestrian trajectory data with the static spatial structure of the station within the same coordinate system. This technical solution selects a subway station in a certain city as the analysis object. This station is characterized by its multi-layered, multi-functional, and densely populated nature. First, the CAD plan of the subway station is acquired and processed. Based on ergonomics and the average adult stride, the station's CAD plan is divided into 0.5m × 0.5m micro-grid units, serving as the basis for spatial analysis and data storage. Specifically, as follows... Figure 4 As shown.

[0030] Furthermore, such as Figure 5 As shown, based on this, key nodes such as turnstiles and stairs are identified according to the functional zoning of the station, and the complex station space is abstracted into a network topology structure composed of "nodes" and "path segments" to define the start and end points of path selection.

[0031] Furthermore, multi-angle high-definition cameras were deployed in the H2 area of ​​a subway station concourse (covering all staircases, escalators, turnstiles, and main entrances / exits) to record pedestrian trajectory videos at a frame rate of 60fps and a resolution of 1920*1080 pixels. The video data was analyzed, treating pedestrians as point masses, and tracing their paths centered on their right heels to obtain the position coordinates of each pedestrian in the measurement area. To address the perspective distortion issue in the pixel coordinates of the pedestrian trajectories captured in the video, this technical solution uses an affine transformation matrix to convert them into actual physical plane coordinates, and then uses a fuzzy matching algorithm to map the cleaned pedestrian trajectory points to the corresponding spatial grid. A total of 307 pedestrian trajectories and 23,635 trajectory points were extracted from the video. This established a precise index relationship between dynamic behavioral data and static spatial units, laying the data foundation for subsequent model construction.

[0032] S200. Through spatial syntax, the target traffic station mapping database is subjected to full-dimensional parameter extraction and filtering to obtain key spatial syntax features. It should be noted that, given the complex spatial structure of urban rail transit stations and the multi-source, heterogeneous nature of pedestrian behavior data, a standardized processing chain from mapping raw data to identifying core features is necessary to build a high-precision path selection prediction model. Therefore, as... Figure 6 As shown, this embodiment proposes a "quantification and screening implementation process for spatial syntax feature parameters". The overall data processing strategy adopts "full-dimensional divergence-intelligent convergence": first, a unified spatiotemporal benchmark is established based on multi-source data mapping; then, spatial syntax is used to extract parameters in all dimensions to maximize the capture of environmental features; finally, redundancy is eliminated through a machine learning screening mechanism to output the core variables.

[0033] S210. Based on Depthmap software, spatial syntactic analysis is performed on the target traffic station mapping database to extract parameters and obtain initial spatial syntactic features; Specifically, in order to comprehensively capture the spatial factors affecting pedestrians, this embodiment utilizes spatial syntax theory to perform a full-dimensional analysis of the site space, transforming the abstract spatial structure into 59 calculable quantitative indicators, as shown in Tables 1, 2, and 3. These indicators cover convex space analysis parameters reflecting the geometric closure and connectivity of the space, axis analysis parameters reflecting the linear guidance capability of the space, and visual field analysis parameters reflecting the pedestrian's visual perception range and global accessibility.

[0034] Table 1 List of characteristic parameters of convex space ; Table 2 List of axis characteristic parameters ; Table 3 List of View Feature Parameters ; However, modeling solely based on theoretical indicators lacks empirical evidence; it is necessary to first verify whether there is a significant correlation between spatial parameters and pedestrian behavior. Therefore, this embodiment introduces the Jaccard coefficient for quantitative analysis. Regions with high visual integration are defined as set A, and high-density areas of actual pedestrian trajectories are defined as set B. The overlap between the two is calculated. The analysis results show (e.g.) Figure 7 , Figure 8 As shown in the figure, when the 5% region with the highest visual integration is selected as the core region, its Jaccard coefficient with the pedestrian trajectory coverage area is as high as 0.79. This data strongly demonstrates that "visual integration" has a dominant influence on pedestrian distribution, thus providing a solid scientific basis for introducing this parameter for subsequent modeling.

[0035] S220. The initial spatial syntactic features are initially screened by feature gain using the XGBoost regression model, and the features with a cumulative contribution rate greater than the preset value are retained to form a candidate set. Specifically, directly inputting 59 high-dimensional spatial features into the model inevitably leads to overfitting and computational redundancy. Therefore, selecting a scientific feature selection mechanism is crucial for accurately identifying core influencing factors from a massive set of parameters. This embodiment proposes a triple selection mechanism of "gain-explanation-collinearity." The selection of convex spatial feature sets is used as an example for illustration.

[0036] To capture the potential nonlinear relationship between spatial syntax indicators and the dependent variable (i.e., pedestrian dwell time), the initial feature set was first subjected to nonlinear transformations (logarithmic transformation, square root transformation, and square transformation). To reduce data dimensionality, a filtering method based on the Pearson correlation coefficient was used for preliminary screening. In this step, only the 13 features with the highest correlation to the dependent variable were retained, forming a preliminary optimized candidate feature set (as shown in Table 4).

[0037] Table 4 Candidate feature set after preliminary screening of convex space ; Then, the XGBoost regression algorithm is used for initial feature selection. Pedestrian path selection behavior is used as the target variable for training, and the average gain contribution of each feature in the decision tree construction process is calculated. The calculation results show (e.g.) Figure 9 As shown in the figure, the gain contributions of features such as "selectivity" and "connection weight" are the most significant.

[0038] S230. Based on the candidate set, features with directional paradoxes and features with variance inflation factors greater than preset values ​​are removed to obtain key space syntactic features.

[0039] Secondly, to ensure that the selected features are interpretable and statistically valid, this scheme introduces SHAP value analysis to determine the positive and negative influence of features on the prediction results, and eliminates features with unclear physical meaning.

[0040] Finally, a multicollinearity test was performed on the retained features. Redundant features with a VIF value greater than 10 were removed by calculating the variance inflation factor (VIF). After the above rigorous triple screening, five core features were finally selected as the input variables of the model, as shown in Table 5, namely: convex space selectivity (C_Cho), square of axis control degree (square_A_Ctrl), square root of axis length (sqrt_LL), square root of view integration degree (sqrt_VI_HH), and square of shortest path length at step size (square_MSSP).

[0041] It should be noted that: Convex space selectivity (C_Cho): Characterizes the "traversal potential" of a space, that is, the frequency with which the space is traversed in the set of shortest paths in the entire station network. The higher the value, the more important the area is as a necessary hub connecting various functional areas.

[0042] The square of the axis control degree (square_A_Ctrl): represents the "local guidance power" of the space, reflecting the current space's control over directly connected spaces. The higher the value, the more branch roads the space connects to, and the more nodes it is a local traffic conversion node.

[0043] The square root of the axis length (sqrt_LL): represents the "sense of visual transparency" in space, reflecting the physical extension distance of the visual channel. A longer axis usually means better long-distance visibility and a clearer sense of direction.

[0044] The square root of visual integration (sqrt_VI_HH): represents the "global visual accessibility" of the space, reflecting the ease or difficulty of observing the entire station from that point. The higher the value, the wider the field of vision, and the easier it is for pedestrians to establish global spatial cognition.

[0045] Squared Shortest Path Length (SMSSP): This is a measure of spatial "travel cost," reflecting the average physical distance required to reach other locations in the network. It is an intuitive physical indicator for measuring spatial depth.

[0046] Table 5. Five core spatial syntactic features finally selected, their descriptions, and VIF values. ; S300. Construct a hierarchical heterogeneous nested Logit decision model to predict pedestrian path selection based on key spatial syntactic features, and obtain the pedestrian path selection prediction results.

[0047] Specifically, a hierarchical heterogeneous nested Logit decision model is constructed, and key spatial syntactic features are input into the hierarchical heterogeneous nested Logit decision model. Based on the lower-level path network of the hierarchical heterogeneous nested Logit decision model, the key spatial syntactic features are linearly weighted with preset weight coefficients to obtain the deterministic utility value of each specific path. Through a logarithmic summation algorithm, the utility values ​​of all lower-level paths belonging to the same policy node are nonlinearly aggregated to calculate the coverage value of that policy node. Based on the upper-level policy network of the hierarchical heterogeneous nested Logit decision model, based on the coverage value of the policy node, the marginal probabilities of various policies such as pedestrian choice structure priority, time optimization, and non-optimal are calculated using the Logit probability transformation formula. Conditional probability processing is applied to the marginal probabilities of various policies to obtain the conditional selection probability of each specific path within the nest. The marginal probabilities of various policies are multiplied by the conditional selection probabilities of each specific path within the nest to obtain the pedestrian path selection prediction result.

[0048] Specifically, existing multinomial Logit models often assume that path options are independent, ignoring the correlation between paths. However, pedestrian decisions in complex environments are often hierarchical, i.e., "strategy first, path selection later." To address the technical shortcomings of traditional models in handling complex path correlations, this invention constructs a two-layer heterogeneous nested decision network structure, as follows: Figure 10 As shown.

[0049] This network structure includes clearly defined hierarchical structures, connection relationships, and data processing and flow mechanisms, which are described in detail below: Regarding the network hierarchy and connections, the model constructed in this embodiment is logically divided into two cascaded topological levels: an "upper-layer policy network" and a "lower-layer path network." Upper-layer policy network: This layer serves as a macro-level filtering layer for decision-making, containing three mutually exclusive policy nodes: the "structure priority nest" (corresponding to the set of main paths with superior visual accessibility), the "time optimization nest" (corresponding to the set of shortcuts with shorter physical distances), and the "non-optimal nest" (corresponding to the set of random walk paths). The role of this layer is to simulate the macro-level classification and pre-screening of travel strategies by pedestrians based on spatial perception.

[0050] Lower-level path network: This level is the micro-execution layer of decision-making, containing several specific physical path nodes. Each path node is uniquely assigned to and connected to a policy node in the upper level based on its physical and spatial attributes. Cross-nesting connections are strictly prohibited, thus forming a tree-like decision branching structure.

[0051] Inter-level connection mechanism: Upper-level strategy nodes and lower-level path nodes are connected through a "containment coefficient". Containment coefficient As a flow control valve connecting channels, it quantifies the overall utility contribution of lower-level path groups to upper-level strategy nodes, and its value range is strictly constrained within a certain range. Intervals are used to ensure the statistical validity of the network structure.

[0052] Data Input and Network Layer Processing Flow. After the five selected core spatial syntactic feature vectors are input into the model as raw data, they need to undergo the following two stages of computational processing to finally generate the prediction result: Phase 1: Bottom-up utility aggregation calculation. First, the data enters the lower-level path network. The system calls the utility function calculation module to perform a linear weighted calculation on the input feature vectors (such as view integration degree, axis control degree, etc.) and preset weight coefficients to obtain the deterministic utility value of each specific path. Subsequently, using a logarithmic summation algorithm, the utility values ​​of all lower-level paths belonging to the same policy node are nonlinearly aggregated to calculate the "inclusion value (IV)" of that policy node. This process propagates the attractiveness of micro-paths upwards, transforming it into the comprehensive attractiveness of macro-policy nodes.

[0053] The specific expression of the utility function is as follows: ; In the above formula, Indicates belonging to the first Nest's first The deterministic utility value of a path, to The weighting coefficients to be calibrated For the selectivity of convex space, This is the square of the axis control degree. It is the square root of the axis length. The square root of the visual integration degree. The step size is the square of the shortest path length. For the specific constant of the nest Phase Two: Top-Down Probability Distribution Prediction. Based on the coverage values ​​calculated in Phase One, the system first processes the data in the upper-layer policy network. Using the Logit probability transformation formula, the marginal probabilities of pedestrians choosing various strategies such as "structure-first," "time-optimized," or "non-optimal" are calculated. Next, the system performs conditional probability processing on the lower-level path network. Given a known strategy, it calculates the conditional selection probability of each specific path within that nest. Finally, the system performs a joint probability operation, multiplying the upper-level strategy selection probability by the lower-level conditional selection probability, and outputs the final predicted selection probability for each physical path. .

[0054] The specific expression of the Logit probability transformation formula is as follows: ; In the above formula, Indicates that the pedestrian chose the first The marginal probability of a strategy nest (such as a "structure-first nest") This is the containment value of the strategy nest, which is the lower-level utility aggregation value calculated in Phase 1. The inclusion factor of the strategy nest (i.e., the calibrated structural parameters such as 0.55 and 0.49) reflects the comprehensive influence of lower-level paths on upper-level strategies. This is a specific constant representing the alternative nesting of the strategy, used to characterize the pedestrian's inherent preference for the strategy, in addition to spatial features. The total number of strategy nests (in this embodiment) ).

[0055] Further calculations of model weight coefficients and instantiation of the prediction function are performed. The constructed nested Logit network structure only provides a mathematical framework for path selection (including undetermined coefficients). To achieve quantitative prediction of specific passenger flows, the specific weight values ​​of each feature variable in the utility function and the inclusion coefficients between levels must be determined. The purpose of this step is to use the collected sample data to solve and lock the model parameters through algorithms, thereby transforming the abstract mathematical model into an executable prediction formula.

[0056] Parameter solving algorithm implementation. This embodiment uses maximum likelihood estimation (MLE) as the parameter solving tool. The acquired "space-behavior mapping database" is used as the training set. A log-likelihood function is constructed. This function represents the probability of observing actual sample data under a given model structure. It obtains the utility weight coefficients corresponding to each spatial syntactic feature (such as the sqrt VI_HH degree of view integration). And the inclusion coefficient connecting the upper-layer strategy and the lower-layer path. .

[0057] Parameter calibration results and model locking. After MLE iterative calculation, the model parameter estimation results obtained in this embodiment are shown in Table 6.

[0058] Table 6. Final parameter estimation results of the nested Logit (NL) model ; In Table 6: * indicates Value < 0.05; ** indicates Value < 0.01; *** indicates Value < 0.001.

[0059] Based on the calculation results in Table 6, perform the following model locking operation: The first step is to verify the structural validity by checking the calculated containment coefficient. The values ​​(0.55, 0.49, 0.45 respectively) were checked to see if they were within the interval (0, 1]. The results confirmed this. The significant deviation from 1 and greater than 0 validates the statistical validity of the hierarchical decision structure, confirming that the system adopts a nested structure rather than a degenerate multinomial Logit structure.

[0060] The second step is to instantiate the prediction function, using the significance levels in Table 6. Characteristic coefficients with a value less than 0.05 (such as the field integration coefficient of 0.105) are used as fixed constants and substituted into the utility function formula.

[0061] Thus, the system completes the transformation from an "abstract model" to a "concrete prediction operator". The generated prediction function can receive spatial feature data input from any new site and directly output the quantized selection probability of each path.

[0062] Finally, the superiority of the model and the prediction accuracy were evaluated. To verify the superiority of this scheme, 159 valid path selection samples extracted from real pedestrian trajectory data were used to estimate the parameters of the three models. The comparison results (as shown in Table 7) show that the NL model constructed in this scheme achieves the minimum values ​​for both the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC), and has the highest pseudo-determination coefficient (Pseudo R²) at 0.425. This indicates that the nested Logit model (NL) exhibits the best performance in all three metrics, not only having the strongest interpretability of the data, but also achieving the best balance between model complexity and goodness of fit.

[0063] Table 7 Comparison of performance metrics for the three models: MNL, ML, and NL ; This demonstrates that the model of this invention has a stronger explanatory power for pedestrian behavior compared to traditional models. Subsequently, this embodiment calculates the model's performance on a series of performance metrics from a classification prediction perspective (see Table 8). The results show that the overall accuracy of the model reaches 0.743. More importantly, its precision is 0.880, recall is 0.827, and the combined F1 score reaches 0.852. These metrics are all at a high level, fully demonstrating that the nested Logit model constructed in this paper, based on endogenized spatial syntax core features, not only has strong explanatory power in theory but also possesses reliable predictive ability in practice.

[0064] Table 8 Summary of Model Prediction Accuracy Evaluation Indicators ; Therefore, we can conclude that: 1) Spatial visual perception within urban rail transit stations has a dominant influence on pedestrian path selection. Empirical data shows that the overlap (Jaccard coefficient) between areas with high visual integration and high-density passenger flow areas is as high as 0.79, demonstrating the necessity of introducing visual parameters for prediction.

[0065] 2) Traditional empirical parameter selection methods are difficult to handle high-dimensional data, while the triple screening mechanism of "XGBoost gain screening + SHAP interpretation and verification + VIF collinearity elimination" can effectively lock 5 core variables from 59 initial indicators, solving the problems of model overfitting and insufficient interpretability.

[0066] 3) The nested Logit model (NL) is the optimal model for predicting pedestrian behavior at complex sites. Its AIC / BIC ratio is significantly lower than that of the MNL and ML models, and its classification accuracy reaches 74.3%, proving that the hierarchical decision structure can more accurately reflect the psychological mechanisms of pedestrians and is suitable as the core algorithm for site planning and evaluation.

[0067] Therefore, the technical features that distinguish the embodiments of the present invention from the prior art are as follows: 1) Analyze the coupling relationship between spatial microstructure and pedestrian path selection.

[0068] 2) Feature selection method based on the triple mechanism of "gain-interpretation-collinearity".

[0069] 3) The constructed hierarchical heterogeneous nested Logit decision structure.

[0070] 4) Quantitative data confirmed that "visual integration" is the primary parameter affecting the weight of pedestrian path selection in subway stations, and established its functional relationship with path selection probability.

[0071] In summary, the embodiments of the present invention have the following advantages compared with the prior art: 1) This embodiment clarifies the significant quantitative relationship between parameters such as visual integration degree in spatial syntax and pedestrian path selection behavior. This relationship was empirically verified through Jaccard coefficient analysis, overcoming the deficiency of traditional traffic simulation that only focuses on physical distance and ignores visual perception.

[0072] 2) This embodiment establishes a scientific feature selection mechanism. By combining machine learning algorithms with statistical tests, five core variables are selected from 59 high-dimensional features. This not only ensures the simplicity of the model but also solves the multicollinearity problem and improves the robustness of the model.

[0073] 3) This embodiment adopts a nested Logit model structure, which significantly improves prediction accuracy. Empirical data shows that compared with traditional models, the accuracy, precision, recall and F1 score of this embodiment are 0.743, 0.880, 0.827 and 0.852, respectively, which can more realistically reproduce pedestrian path preferences in complex spatial environments.

[0074] Reference Figure 2 A pedestrian path selection prediction system for transportation stations based on space syntax includes: The first module 201 is used to match dynamic pedestrian trajectory data with static station spatial structure based on the same coordinate system to construct a target transportation station mapping database. The second module 202 is used to extract parameters from the target traffic station mapping database in all dimensions and perform filtering and elimination processing through spatial syntax to obtain key spatial syntax features; The third module 203 is used to construct a hierarchical heterogeneous nested Logit decision model to predict pedestrian path selection based on key spatial syntactic features, and obtain the pedestrian path selection prediction results.

[0075] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0076] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A pedestrian path selection prediction method for traffic stations based on space syntax, characterized in that, Includes the following steps: Based on the same coordinate system, dynamic pedestrian trajectory data is matched with static station spatial structure to construct a target transportation station mapping database; By using spatial syntax to extract parameters from the target traffic station mapping database in all dimensions and perform filtering and elimination processes, key spatial syntax features are obtained. A hierarchical and heterogeneous nested Logit decision model is constructed to predict pedestrian path selection based on key spatial syntactic features, and the prediction results are obtained.

2. The pedestrian path selection prediction method for traffic stations based on space syntax according to claim 1, characterized in that, The step of matching dynamic pedestrian trajectory data with static station spatial structures under the same coordinate system to construct a target transportation station mapping database specifically includes: Obtain the CAD plan of the target transportation station, and divide the CAD plan of the target transportation station into several grid units based on ergonomics and the average stride length of adults. Based on the CAD plan of the divided target traffic stations, key nodes and path segments are extracted, and the start and end points of path selection are defined to construct the network topology diagram of the target traffic stations. The system acquires pedestrian trajectory videos from cameras at the target transportation station, treats each pedestrian as a point mass, and tracks their trajectory with their right heel as the center to obtain the position coordinates of each pedestrian in the target measurement area. The position coordinates of each pedestrian in the target measurement area are converted into actual physical plane coordinates by using an affine transformation matrix. Then, a fuzzy matching algorithm is used to map the actual physical plane coordinates to the network topology diagram of the target traffic station, thus constructing a target traffic station mapping database.

3. The pedestrian path selection prediction method for traffic stations based on space syntax according to claim 2, characterized in that, The step of extracting and filtering parameters from the target traffic station mapping database using spatial syntax to obtain key spatial syntax features specifically includes: Based on the Depthmap software, spatial syntactic analysis was performed on the target traffic station mapping database to extract parameters and obtain initial spatial syntactic features. The initial spatial syntactic features are screened by feature gain using the XGBoost regression model, and the features with a cumulative contribution rate greater than a preset value are retained to form a candidate set. Based on the candidate set, features exhibiting directional paradoxes and features with variance inflation factors greater than a preset value are removed to obtain key space syntactic features.

4. The pedestrian path selection prediction method for traffic stations based on space syntax according to claim 3, characterized in that, The spatial syntax analysis of the target traffic station mapping database specifically includes convex space analysis, axis analysis, and view domain analysis. The convex space analysis is used to identify continuous passable areas, the axis analysis is used to extract the main path skeleton, and the view domain analysis is used to calculate the reachable range of the line of sight.

5. The pedestrian path selection prediction method for traffic stations based on space syntax according to claim 4, characterized in that, The key spatial syntax features specifically include convex space selectivity, axis control degree squared, axis length squared, field integration degree square root, and step shortest path length squared.

6. The pedestrian path selection prediction method for traffic stations based on space syntax according to claim 5, characterized in that, The hierarchical heterogeneous nested Logit decision model specifically includes an upper-layer policy network and a lower-layer path network, wherein: The upper-layer policy network is used to simulate the macro-classification and pre-screening of travel strategies by pedestrians based on spatial perception. It includes three mutually exclusive policy nodes: structure priority nest, time optimization nest, and non-optimal nest. The structure priority nest corresponds to the set of main paths with good visual accessibility, the time optimization nest corresponds to the set of shortcuts with short physical distance, and the non-optimal nest corresponds to the set of random walk paths. The lower-level path network is the micro-execution layer for decision-making, which includes several specific physical path nodes; An inclusion coefficient is constructed by quantifying the overall utility contribution of lower-level path groups to upper-level strategy nodes, and then the upper-level strategy nodes and lower-level path nodes are connected.

7. The pedestrian path selection prediction method for traffic stations based on space syntax according to claim 6, characterized in that, The step of constructing a hierarchical heterogeneous nested Logit decision model to predict pedestrian path selection based on key spatial syntactic features and obtain the pedestrian path selection prediction results specifically includes: A hierarchical heterogeneous nested Logit decision model is constructed by inputting key spatial syntactic features into the hierarchical heterogeneous nested Logit decision model. The lower-level path network based on the hierarchical heterogeneous nested Logit decision model performs linear weighting operations on the key space syntactic features and preset weight coefficients to obtain the deterministic utility value of each specific path. By using a logarithmic summation algorithm, the utility values ​​of all lower-level paths belonging to the same policy node are non-linearly aggregated to calculate the coverage value of that policy node. The upper-level policy network based on the hierarchical heterogeneous nested Logit decision model calculates the marginal probabilities of various policies, including structure priority, time optimization, and non-optimal, based on the inclusion value of the policy node and the Logit probability transformation formula. Conditional probability processing is applied to the marginal probabilities of various strategies to obtain the conditional selection probability of each specific path within the nest; The pedestrian path selection prediction result is obtained by multiplying the marginal probabilities of various strategies with the conditional selection probabilities of each specific path within the nest.

8. A pedestrian path selection prediction system for traffic stations based on space syntax, characterized in that, Includes the following modules: The first module is used to match dynamic pedestrian trajectory data with static station spatial structure under the same coordinate system to build a target transportation station mapping database. The second module is used to extract and filter parameters from the target traffic station mapping database in all dimensions using spatial syntax to obtain key spatial syntax features. The third module is used to construct a hierarchical heterogeneous nested Logit decision model to predict pedestrian path selection based on key spatial syntactic features, and obtain the pedestrian path selection prediction results.

Images