Outdoor three-dimensional walking navigation road network map generation method fusing public source trajectory data

Abstract

The invention relates to the technical field of computers, and discloses an outdoor three-dimensional walking navigation road network map generation method fusing public source trajectory data. The method comprises the following steps: carrying out Kalman filtering and spline interpolation preprocessing on a public source track; performing track segment semantic segmentation based on a motion state discrimination criterion; extracting multi-dimensional behavior characteristics of geometry, kinematics, direction stability, time sequence distribution and the like of the track segment; generating a candidate path cluster through an improved space-time clustering algorithm; constructing a traffic intention comprehensive scoring model, and screening a high-intention path cluster; a main curve algorithm is adopted to extract a center line and construct a two-dimensional topology road network; and finally, fusing three-dimensional elevation information of an original track, endowing a road network with a slope, accumulating climbing attributes, and generating a three-dimensional walking road network map supporting refined navigation. The accuracy and practicability of the road network are improved.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to a method for generating outdoor three-dimensional pedestrian navigation road network maps that integrates crowdsourced trajectory data. Background Technology

[0002] With the rapid development of smart cities and location services, building high-precision pedestrian navigation road networks based on crowdsourced trajectory data has become the core foundation of outdoor 3D navigation systems.

[0003] These systems rely on massive amounts of pedestrian movement trajectories, automatically extracting effective travel paths through data mining and spatial analysis to generate dynamic road network maps that require no manual surveying. This method offers advantages in improving mapping efficiency and reducing maintenance costs, and is particularly suitable for open areas such as squares, parks, and campuses.

[0004] However, the core challenge lies in how to accurately identify "intentional paths" with navigational significance from unstructured trajectory flows and eliminate interference trajectories caused by random pedestrian behavior.

[0005] The key to generating pedestrian road networks by integrating crowdsourced trajectory data lies in the path validity assessment mechanism. Current mainstream methods generally employ density-based clustering algorithms to aggregate trajectory points, treating high-density areas as potential paths. This strategy assumes that frequently used areas are valid roads, essentially equating spatial usage frequency with path semantics. However, in unpaved surface areas such as lawns, plazas, and landscape water features, pedestrians often engage in short-term, low-frequency crossing behaviors in various directions due to shortcuts, sightseeing, or temporary avoidance.

[0006] Although these random trajectories may form density clusters in certain areas, they do not have the intention of continuous passage and do not conform to actual walking norms. If they are misjudged as valid paths, the generated road network will contain a large number of false connections, which will seriously damage the accuracy and reliability of the navigation system.

[0007] Existing technologies have significant shortcomings in the path filtering stage: Relying solely on the spatial density of trajectory points ignores the fundamental constraint of surface physical properties on the feasibility of passage. The lack of coupled analysis of trajectory direction consistency and surface material compatibility makes it impossible to distinguish between passable but unplanned reasonable shortcuts and inaccessible but occasionally crossed invalid trajectories.

[0008] Especially in mixed terrain scenarios, such as areas where grasslands and trails intersect, traditional methods struggle to establish a logical connection between travel intentions and environmental constraints, leading to non-path areas being incorrectly included in the road network topology.

[0009] Therefore, there is an urgent need for a path feasibility assessment mechanism that can integrate prior knowledge of surface material and trajectory movement characteristics to accurately suppress invalid crossing trajectories and reliably extract the true intended path. Summary of the Invention

[0010] The technical problem to be solved by this invention is that, when generating pedestrian road networks using crowdsourced GPS trajectory data, existing technologies mainly rely on density-based spatial clustering noise application algorithms or their variants. These algorithms only rely on the spatial density of trajectory points for path identification, thus incorrectly identifying the accidental and high-density crossing behaviors of pedestrians due to temporary activities or shortcuts as fixed and valid walking paths. This results in the generated road network map containing a large number of false road networks, which cannot accurately reflect the pedestrian infrastructure with clear travel intentions in the real world.

[0011] To address the aforementioned technical issues, this invention provides a method for generating outdoor 3D pedestrian navigation road network maps that integrates crowdsourced trajectory data. This method aims to establish a multi-dimensional, hierarchical path intent discrimination and filtering mechanism by introducing behavioral semantics and temporal distribution features of trajectory segments. This fundamentally distinguishes between real paths with long-term, stable travel intentions and accidental, random traversal trajectories, thereby generating a high-precision, high-reliability outdoor 3D pedestrian navigation road network.

[0012] The method for generating an outdoor 3D pedestrian navigation road network map that integrates crowdsourced trajectory data, provided by this invention, specifically includes the following steps: Acquire and preprocess crowdsourced trajectory data, which is a set of trajectory point sequences containing user identifiers, timestamps, and three-dimensional coordinates; the preprocessing includes noise filtering and smoothing of the original trajectory point sequences using the Kalman filter algorithm, and resampling of the processed trajectory point sequences using cubic spline interpolation to generate standardized trajectory data with uniform time intervals.

[0013] Furthermore, the method includes performing trajectory segment semantic segmentation on the standardized trajectory data. Specifically, each standardized trajectory is traversed, and according to a preset motion state discrimination criterion, the continuous sequence of trajectory points is segmented into trajectory segments with independent motion semantics.

[0014] The motion state discrimination criteria include: when the instantaneous velocity of N consecutive points in the trajectory is less than a preset static velocity threshold and the duration is greater than a preset dwell time threshold, it is determined to be a dwell point, and this point is used as a dividing point; when the rate of change of the heading angle of a certain point in the trajectory is greater than a preset sharp turn angle threshold, it is determined to be a turning point, and this point is used as a dividing point.

[0015] This segmentation decomposes the original long trajectory into a set of trajectory segments that represent a continuous, single-intention movement process.

[0016] Furthermore, the method includes extracting multi-dimensional behavioral features for each trajectory segment in the set of trajectory segments to construct a trajectory segment behavioral feature vector that can quantify its travel intention. The trajectory segment behavioral feature vector specifically includes: Geometric features: Calculate the straight-line distance of the trajectory segment and the actual path length, and obtain the path curvature based on the ratio of the two; Kinematic characteristics: Calculate the average velocity, velocity variance, average acceleration, and acceleration variance of all points within the trajectory segment; Orientation stability characteristics: Calculate the cyclic standard deviation of all heading angle data points within the trajectory segment to quantify the consistency of the trajectory segment's direction of travel; Temporal distribution characteristics: The timestamps of the trajectory segments are mapped to standardized time periods, their distribution frequency in different time units is statistically analyzed, and the Shannon entropy of their temporal distribution is calculated.

[0017] Furthermore, the method includes preliminary trajectory segment clustering based on spatiotemporal proximity.

[0018] An improved density-based spatial clustering noise application algorithm is adopted. When calculating the core object, the algorithm not only requires that the number of trajectory segments in the neighborhood meets the density threshold, but also requires that the timestamps of these trajectory segments are dispersed within a preset time span, so as to initially filter out dense trajectories caused by single, short-term events.

[0019] This step clusters the set of trajectory segments by setting three parameters: spatial search radius, minimum number of trajectory segments, and minimum time span, to generate a set of candidate path clusters.

[0020] Furthermore, the method includes performing a comprehensive scoring and filtering of travel intention for each candidate path cluster in the candidate path cluster set. For each candidate path cluster, the behavioral feature vectors of all trajectory segments within it are first aggregated, and the mean and variance of the behavioral features within the cluster are calculated.

[0021] Subsequently, based on preset weighting coefficients, the average speed, speed variance, directional stability, path curvature, and time distribution entropy within the cluster are weighted and summed to construct a comprehensive scoring model for travel intention.

[0022] The calculation formula for the scoring model is as follows: ; in The score is based on the intentionality of the individual. The average velocity within the cluster. This is the normalized value of the intra-cluster velocity variance. This is a normalized value for intra-cluster directional stability. This is the normalized value of the path curvature within the cluster. This is the normalized value of the temporal distribution entropy within the cluster. to These are pre-calibrated weighting coefficients.

[0023] Based on the calculated comprehensive score of travel intent, a global intent scoring threshold is set. Candidate path clusters with scores below the threshold are judged as accidental crossing trajectories and are eliminated, while those with scores above the threshold are retained as high intent path clusters.

[0024] Furthermore, the method includes centerline extraction and network topology construction for the selected high-intent path clusters. For each high-intent path cluster, the master curve algorithm is used to fit the centerline of its spatial point set to generate an initial path skeleton representing the path geometry.

[0025] Then, all initial path skeletons are processed, and their spatial intersections are calculated. Each intersection and the endpoint of each path skeleton is defined as a road network topology node.

[0026] The path skeleton segment connecting any two topological nodes is defined as a topological edge of the road network. This step transforms discrete path point cloud data into a two-dimensional topological road network graph with node and edge connections.

[0027] Furthermore, the method includes assigning three-dimensional elevation and slope attributes to the two-dimensional topological network map. For each topological edge in the topological network map, the three-dimensional coordinate data of all original trajectory segments constituting that edge are backtracked.

[0028] The original elevation data are smoothed using the moving average method, and the smoothed elevation values ​​are assigned to each discrete point of the topological edge to generate a three-dimensional path.

[0029] Based on this 3D path, the total length, cumulative elevation gain, cumulative elevation loss, and average slope of each topological edge are calculated, and these 3D attributes are stored as weight information in the topological road network map, ultimately generating an outdoor pedestrian navigation road network map containing complete geometric, topological, and 3D physical attributes.

[0030] As one embodiment of the present invention, the Kalman filter algorithm in the preprocessing is configured with a state vector that is a six-dimensional vector containing three-dimensional position coordinates and three-dimensional velocity components, and the observation vector is the three-dimensional coordinates directly output by the global positioning system module.

[0031] The state transition matrix is ​​constructed based on the uniform motion model, the observation matrix is ​​an identity matrix, and the parameters of the process noise covariance matrix and the observation noise covariance matrix are set based on prior sensor characteristic data.

[0032] In one embodiment of the present invention, the static speed threshold for trajectory segment semantic segmentation is set to 0.5 meters per second, the dwell time threshold is set to 180 seconds, and the sharp turning angle threshold is set to 45 degrees per second. These thresholds are empirical values ​​obtained based on statistical analysis of a large number of pedestrian walking behaviors, aiming to maximize the separation of the three basic behaviors of walking, dwelling, and turning.

[0033] As one implementation method described in this log, the weight coefficients in the traffic intentionality comprehensive scoring model... to The model is trained on a dataset containing labeled real-path and non-path samples through an optimization process based on the gradient descent algorithm. The objective function of this optimization process is to maximize the model's classification accuracy on the training samples, so that each weight coefficient can optimally reflect the importance of different behavioral features in intent discrimination.

[0034] In one embodiment of the present invention, the principal curve algorithm used in the centerline extraction initializes the curve as the first-order principal component axis of the path cluster point set. The algorithm continuously adjusts the curve shape through iterative projection and smoothing steps to pass through the "middle" of the data points until the curve position converges or reaches the preset maximum number of iterations. This method ensures that the extracted centerline smoothly and accurately reflects the overall direction of the path cluster.

[0035] In one embodiment of the present invention, the data structure of the final generated outdoor 3D pedestrian navigation road network map follows the storage paradigm of a graph database. Each topological node is stored as a node object, containing a unique node identifier, longitude, latitude, and altitude attributes. Each topological edge is stored as a relation object, containing a unique relation identifier, a start node identifier, an end node identifier, and a series of navigation-required weight information as attributes such as path length, average slope, estimated travel time, and cumulative elevation gain.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a multi-dimensional behavioral feature vector that includes geometric, kinematic, directional stability and temporal distribution features, and establishes a comprehensive scoring model for travel intentions. This enables in-depth mining and quantitative evaluation of the inherent behavioral intentions of trajectory data, fundamentally solving the technical problem that existing technologies rely solely on spatial density and cannot distinguish between real paths and accidental crossing trajectories.

[0037] 2. The hierarchical screening mechanism proposed in this invention first performs coarse screening using an improved spatiotemporal clustering algorithm, and then performs fine filtering using an intentional scoring model. This effectively eliminates false paths generated by temporary activities without stable travel intentions, such as square gatherings and group crossings of lawns, thereby improving the accuracy and reliability of the generated road network.

[0038] 3. Based on the generation of a two-dimensional topological road network, this invention further integrates the three-dimensional elevation information of the original trajectory, endowing the road network with key three-dimensional navigation attributes such as slope and cumulative elevation gain. This enables more refined pedestrian navigation applications that take into account terrain undulations, meeting the planning needs of special groups, such as people with disabilities and users pushing strollers, for barrier-free or gentle slope paths.

[0039] 4. This invention decomposes the complex trajectory data processing flow into a series of clearly defined modular steps, such as trajectory segment semantic segmentation, multi-dimensional feature extraction, intent scoring, centerline extraction and topology construction, forming a complete, robust and engineerable outdoor 3D pedestrian road network generation technology solution. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the overall technical solution architecture of the outdoor three-dimensional pedestrian navigation road network map generation method that integrates crowdsourced trajectory data proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the comprehensive scoring model for travel intention in this invention; Figure 3 This is a flowchart illustrating the logical process of trajectory segment semantic segmentation and multi-dimensional behavioral feature extraction in this invention. Figure 4 This is a flowchart illustrating the logical process of preliminary trajectory segment clustering and high-intent path cluster selection based on spatiotemporal proximity in this invention. Figure 5 This is a logical flowchart of the centerline extraction and two-dimensional topological road network construction of the high-intent path cluster in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal and the cloud in this invention. Detailed Implementation

[0041] Please refer to Figures 1 to 6 This invention provides a method for generating outdoor three-dimensional pedestrian navigation road network maps by integrating crowdsourced trajectory data, aiming to solve the technical defect in the prior art that misjudges accidental crossing behavior as a valid path because it relies solely on the spatial density of trajectory points for clustering.

[0042] This method introduces behavioral semantics and temporal distribution features of trajectory segments to construct a multi-dimensional, hierarchical path intent discrimination mechanism, thereby accurately identifying real walking paths with long-term and stable travel intentions, and generating a high-precision outdoor walking navigation road network map containing complete geometric shape, topological structure and three-dimensional physical attributes.

[0043] The method first performs the steps of acquiring and preprocessing crowdsourced trajectory data.

[0044] The crowdsourced trajectory data comes from the historical trajectory records of users collected and uploaded by a large number of mobile terminal devices. Each trajectory consists of a series of trajectory points arranged in chronological order. Each trajectory point contains a unique user identifier, a timestamp accurate to the millisecond level, and three-dimensional coordinate information composed of longitude, latitude, and altitude.

[0045] Raw trajectory data often suffers from noise interference such as positioning drift, signal loss, or high-frequency jitter, so systematic preprocessing is necessary to improve data quality.

[0046] The preprocessing process includes two core sub-steps: First, the Kalman filter algorithm is used to suppress noise and smooth the trajectory of the original trajectory point sequence. Second, the filtered trajectory point sequence is resampled to ensure a uniform time interval.

[0047] In the Kalman filtering stage, the state vector is set as a six-dimensional vector, with the first three components representing the three-dimensional position coordinates at the current moment and the last three components representing the corresponding three-dimensional velocity components.

[0048] The observation vector is directly taken from the three-dimensional coordinate values ​​output by the Global Positioning System module.

[0049] The state transition matrix is ​​constructed based on a uniform linear motion model, which assumes that the user's motion state remains approximately unchanged between two adjacent sampling periods. The observation matrix is ​​set as an identity matrix, indicating that the observed values ​​directly correspond to the positional portion of the state vector.

[0050] The parameters of the process noise covariance matrix and the observation noise covariance matrix are calibrated based on prior sensor performance indicators. For example, the standard deviation of the position error of a typical civilian-grade GPS module is about 5 meters, and the standard deviation of the velocity error is about 0.5 meters per second.

[0051] By recursively calculating the predicted and updated states, the Kalman filter can suppress high-frequency noise while preserving the overall motion trend of the trajectory.

[0052] After filtering is completed, the trajectory resampling stage begins.

[0053] Because different devices have different sampling frequencies, the time intervals between the original trajectory points are inconsistent, which is not conducive to subsequent temporal feature extraction and trajectory segmentation.

[0054] Therefore, cubic spline interpolation is used to interpolate the filtered trajectory point sequence. Cubic spline interpolation can generate position estimates at any specified time point while ensuring the continuity and smoothness of the trajectory.

[0055] In this embodiment, all trajectories are uniformly resampled to a standard frequency of trajectory points per second, thereby forming a standardized trajectory dataset.

[0056] Each trajectory in this dataset consists of trajectory points at equal time intervals, and each point has undergone noise suppression and smoothing, laying a high-quality data foundation for subsequent semantic segmentation and feature extraction.

[0057] After obtaining standardized trajectory data, a trajectory segment semantic segmentation step is performed. The goal of this step is to decompose each long trajectory into several short trajectory segments with independent motion semantics. Each trajectory segment represents a continuous, single-intention movement process. The segmentation operation is based on a preset motion state discrimination criterion, traversing all trajectory points of each standardized trajectory and dynamically determining whether there are state transition points. Specifically, two types of key segmentation points are defined: dwell points and turning points.

[0058] The criteria for determining a stopping point are as follows: when the instantaneous velocity of all trajectory points within a continuous 180-second period is less than 0.5 meters per second, the starting point of that time period is marked as a stopping point. Instantaneous velocity is calculated by dividing the Euclidean distance between adjacent trajectory points by the time interval. This threshold is based on statistical analysis of large-scale pedestrian behavior data and can distinguish between slow walking and stationary standing. Once a stopping interval meeting the criteria is detected, dividing markers are inserted at both ends of that interval, splitting the original trajectory into two segments.

[0059] The turning point is determined as follows: when the rate of change of the heading angle at a point in the trajectory exceeds 45 degrees per second, that point is marked as a turning point. The heading angle is determined by the azimuth angle between the current trajectory point and its predecessor, and the rate of change is calculated by dividing the difference in heading angles between two adjacent points by the time interval. This threshold is also based on empirical research and is sufficient to capture common pedestrian sharp-turning behaviors, such as turning at intersections or navigating around obstacles, while avoiding misjudging minor directional fluctuations as valid turning events. Whenever a turning point is detected, the trajectory is cut at that point.

[0060] Through the aforementioned dual discrimination mechanism, each original long trajectory is systematically decomposed into a series of trajectory segments. Each trajectory segment contains no stopping or turning behavior, representing a continuous straight-line or gently curving journey. The resulting set of trajectory segments constitutes the basic unit for subsequent feature extraction and cluster analysis.

[0061] Next, for each trajectory segment in the trajectory segment set, multi-dimensional behavioral feature extraction is performed to construct a trajectory segment behavioral feature vector. This vector is used to quantify the strength of the travel intention embodied by the trajectory segment and is the core input for subsequent intentionality scoring. The feature vector contains four dimensions: geometric features, kinematic features, directional stability features, and temporal distribution features.

[0062] In terms of geometric features, the straight-line distance between the start and end points of the calculated trajectory segment is... The actual path length is the sum of the distances between all adjacent trajectory points in the trajectory segment. Path curvature Defined as and The ratio, which ranges from 0 to 1, indicates that the trajectory segment is closer to a straight line and the travel intention is clearer. Conversely, if the ratio is less than 1, it may reflect exploratory walking or detour behavior.

[0063] Regarding kinematic characteristics, the mean instantaneous velocity of all trajectory points within the trajectory segment is calculated. With variance and the mean instantaneous acceleration With variance Instantaneous acceleration is obtained by dividing the difference between adjacent velocities by the time interval. Stable high-speed walking typically corresponds to higher... With lower This indicates that the user has a clear destination orientation; while walking at low speed and with frequent speed changes may be considered wandering or temporary crossing.

[0064] Regarding directional stability characteristics, the cyclic standard deviation of all heading angle data points within the calculated trajectory segment was used. Cyclic standard deviation is a statistic specifically designed for handling angular data, accurately measuring the dispersion of a set of angular values ​​around their mean. The smaller the value, the more consistent the direction of travel, and the more likely the trajectory segment is to be part of a fixed path; The larger the value, the more frequently the direction changes, and the weaker the intention to travel.

[0065] In terms of time distribution characteristics, the timestamps of all trajectory points in the trajectory segment are mapped to a standardized 24-hour cycle, which is divided into 96 15-minute time units.

[0066] The number of trajectory points appearing within each time unit is counted to form a time distribution histogram. The Shannon entropy of this histogram is then calculated. Shannon entropy is used to measure the uniformity of time distribution: if trajectory segments appear concentrated in a few time units, A lower value may reflect a temporary crossing driven by a specific event; if the trajectory segments are distributed across multiple time periods throughout the day... A higher level is more likely to represent daily commuting or regular route usage behavior.

[0067] The above four dimensions, comprising seven specific indicators, together constitute a seven-dimensional trajectory segment behavior feature vector, comprehensively depicting the travel intention attributes of the trajectory segment.

[0068] After feature extraction, preliminary trajectory segment clustering based on spatiotemporal proximity is performed. This step employs an improved density-based spatial clustering noise application algorithm. Its core idea is to add temporal distribution dispersion requirements to the traditional spatial density constraints, thereby filtering dense trajectory clusters caused by a single short-term event. The clustering process uses the geometric center point of each trajectory segment (i.e., the midpoint between the start and end points) as the spatial reference point.

[0069] The algorithm sets three key parameters: spatial search radius. Minimum number of trajectory segments and minimum time span For any trajectory segment, if its spatial search radius... There are no less than within the range There are several other trajectory segments, and the time span covered by the timestamps of these trajectory segments is not less than [a certain number of] time segments. If so, the trajectory segment is identified as the core object. In this embodiment, Set to 10 meters. Set to 20. The time frame is set to 12 hours. This means that an area is considered a potentially valid path area only if more than 20 track segments pass through it continuously for at least 12 hours.

[0070] By iteratively expanding the neighborhood of the core object, the algorithm groups trajectory segments that meet the conditions into the same candidate path cluster. Trajectory segments not included in any cluster are considered noise points and are directly eliminated. The candidate path cluster set generated in this step has initially eliminated false paths generated by short-term high-density events such as square gatherings and temporary construction detours, but it may still contain some trajectory clusters that, although they have spatiotemporal persistence, lack genuine travel intentions and require further screening.

[0071] Subsequently, a comprehensive scoring and filtering of travel intent is performed on each candidate path cluster in the candidate path cluster set. This step is the core innovation of this invention, aiming to achieve refined filtering by quantitatively evaluating the overall travel intent strength of each cluster.

[0072] For any candidate path cluster, the behavioral feature vectors of all trajectory segments within it are first aggregated. The mean and variance within each feature dimension are calculated to form cluster-level statistical features. For example, the average velocity of all trajectory segments within a cluster is calculated as the cluster's mean. The mean of the calculated velocity variance is used as And so on.

[0073] Based on this, a comprehensive scoring model for travel intention is constructed. The calculation formula for this model is as follows: ; The score is based on the applicant's intentions. , , , , These are the normalized values ​​of the cluster average velocity, velocity variance, directional stability, path curvature, and time distribution entropy, respectively, with a normalization range of 0-1. to For the pre-calibrated weighting coefficients, satisfying .

[0074] The weight coefficients were calibrated through supervised learning. On a training set containing manually labeled real path clusters and non-path clusters, gradient descent was used to optimize the weight combination, ensuring the model's output score maximizes the differentiation between the two classes of samples. After training, to These weights are fixed for actual inference. In this embodiment, the typical weight configuration obtained after training is as follows: This indicates that average velocity and directional stability dominate the intentionality discrimination.

[0075] Calculate the value of each candidate path cluster After setting the value, a global intentionality scoring threshold is set. In this embodiment, Set to 0.65. All The candidate path clusters were determined to be accidental crossing trajectories and were eliminated; Clusters of high intent are retained. This selection mechanism ensures that the final retained path clusters are not only persistent in time and space, but also reflect clear and stable travel intent in terms of behavioral semantics.

[0076] After obtaining a set of highly intentional path clusters, centerline extraction and network topology construction are performed. This step transforms the discrete trajectory point cloud into a structured road network map.

[0077] For each high-intent path cluster, the principal curve algorithm is used to fit the centerline of its spatial point set. The principal curve is a smooth curve passing through the "middle" of the data, effectively representing the overall direction of the point cloud. During algorithm initialization, the first-order principal component axes of all points in the path cluster are used as the initial curve. Then, an iterative process begins: first, each data point is projected onto the current curve to obtain projected points; then, using these projected points as control points, a new smooth curve is generated through local weighted regression. The above projection and smoothing steps are repeated until the change in curve position is less than a preset convergence threshold or the maximum number of iterations (e.g., 100) is reached. The final principal curve obtained is the initial path skeleton for that path cluster.

[0078] After extracting all path skeletons, topological relationships are constructed. All initial path skeletons are traversed, and the spatial intersections between any two skeletons are calculated. Intersections are determined using a line segment intersection detection algorithm, considering intersections on the projection plane (usually a horizontal plane) in 3D space, and verifying that their vertical elevation differences are within tolerance (e.g., 2 meters). All detected intersections, along with the two endpoints of each path skeleton, constitute the topological node set of the road network.

[0079] Next, each path skeleton is segmented according to its internal topological nodes. The skeleton segment between any two adjacent topological nodes is defined as a topological edge. Thus, the entire road network is represented as a two-dimensional topological graph composed of nodes and edges, where nodes represent intersections or path endpoints, and edges represent pedestrian segments connecting these nodes.

[0080] Finally, the two-dimensional topological road network map is assigned three-dimensional elevation and slope attributes to generate a complete outdoor three-dimensional pedestrian navigation road network map. For each topological edge, its original composition is traced back, that is, all trajectory segments belonging to the high-intentionality path cluster to which the edge belongs. All original three-dimensional coordinate points within the spatial range of the edge in these trajectory segments are extracted to form an elevation observation sequence.

[0081] The elevation sequence is smoothed using a moving average method. The window size is dynamically adjusted according to the edge length, for example, a smoothing window is set every 10 meters. The smoothed elevation values ​​are assigned to discrete sampling points of the topological edge, thereby reconstructing the three-dimensional geometry of the edge.

[0082] Based on this 3D path, several key navigation attributes are calculated: path length is the sum of distances of all segments in 3D space; cumulative elevation gain is the sum of all positive elevation increments; cumulative elevation loss is the sum of the absolute values ​​of all negative elevation increments; and average slope is the ratio of the total elevation change to the horizontal projection length. These attributes are stored as weight information along with the identifiers of the topological edges and the identifiers of the start and end nodes.

[0083] The final generated outdoor 3D pedestrian navigation road network map is organized using a graph database structure. Each topological node is stored as a node object, containing a node identifier, longitude, latitude, and altitude; each topological edge is stored as a relation object, containing a relation identifier, start node identifier, end node identifier, and attributes such as path length, average slope, estimated travel time, cumulative elevation gain, and cumulative elevation loss. This data structure supports advanced functions such as efficient path querying, shortest path calculation, and barrier-free navigation.

[0084] In summary, this embodiment achieves fully automated and intelligent generation of a high-precision 3D pedestrian road network map from raw crowdsourced trajectory data through a series of rigorous and coherent steps, including trajectory preprocessing, semantic segmentation, multi-dimensional feature extraction, spatiotemporal clustering, intent scoring, centerline fitting, and 3D attribute assignment. The entire process not only eliminates accidental trajectories but also preserves the geometric details and 3D terrain information of the actual paths, providing a solid data foundation for subsequent personalized and barrier-free pedestrian navigation.

[0085] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0086] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for generating outdoor 3D pedestrian navigation road network maps by integrating crowdsourced trajectory data, characterized in that, include: Acquire and preprocess crowdsourced trajectory data to generate standardized trajectory data with uniform time intervals; The standardized trajectory data is subjected to trajectory segment semantic segmentation, and each standardized trajectory is divided into trajectory segments with independent motion semantics according to the preset motion state discrimination criteria. For each trajectory segment in the set of trajectory segments, its multi-dimensional behavioral features are extracted to construct a trajectory segment behavioral feature vector; Generate a set of candidate path clusters; Obtain the comprehensive score of intent to travel, eliminate candidate path clusters with scores lower than the global intent score threshold, and retain high intent path clusters; The centerline of the high-intention path cluster is extracted and the network topology is constructed. The master curve algorithm is used to fit the centerline of the spatial point set of each high-intention path cluster to generate an initial path skeleton. The spatial intersections between all initial path skeletons are calculated, and the intersections and skeleton endpoints are defined as topology nodes. The skeleton segment connecting any two topology nodes is defined as a topology edge to form a two-dimensional topology road network diagram. The two-dimensional topological road network map is given three-dimensional elevation and slope attributes. The original trajectory segment three-dimensional coordinate data corresponding to each topological edge is traced back. The elevation data is smoothed by the moving average method and the discrete points of the topological edge are assigned to generate a three-dimensional path. The path length, cumulative elevation gain, cumulative elevation loss and average slope of each topological edge are calculated as three-dimensional navigation attributes. Finally, an outdoor three-dimensional pedestrian navigation road network map is generated.

2. The method for generating an outdoor three-dimensional pedestrian navigation road network map by fusing crowdsourced trajectory data according to claim 1, characterized in that, The crowdsourced trajectory data is a set of trajectory point sequences containing user identifiers, timestamps, and three-dimensional coordinates; the preprocessing includes noise filtering and smoothing of the original trajectory point sequences using the Kalman filter algorithm, and resampling of the processed trajectory point sequences using cubic spline interpolation.

3. The method for generating an outdoor three-dimensional pedestrian navigation road network map by fusing crowdsourced trajectory data according to claim 1, characterized in that, The motion state discrimination criteria include: when the instantaneous speed of N consecutive points in the trajectory is less than a preset stationary speed threshold and the duration is greater than a preset dwell time threshold, it is determined to be a dwell point; when the rate of change of the heading angle of a certain point in the trajectory is greater than a preset sharp turn angle threshold, it is determined to be a turning point.

4. The method for generating an outdoor three-dimensional pedestrian navigation road network map by fusing crowdsourced trajectory data according to claim 1, characterized in that, The trajectory segment behavior feature vector includes: path curvature, average velocity, velocity variance, average acceleration, acceleration variance, directional stability cyclic standard deviation, and time distribution Shannon entropy.

5. The method for generating an outdoor three-dimensional pedestrian navigation road network map by fusing crowdsourced trajectory data according to claim 1, characterized in that, The steps for generating a candidate path cluster set include: The trajectory segment set is initially clustered based on spatiotemporal proximity. An improved density-based spatial clustering noise application algorithm is adopted to simultaneously satisfy the following conditions when calculating the core object: the number of trajectory segments in the spatial neighborhood is greater than the density threshold and the time span covered by the timestamp is not less than the preset minimum time span.

6. The method for generating an outdoor three-dimensional pedestrian navigation road network map by fusing crowdsourced trajectory data according to claim 1, characterized in that, The steps for obtaining the comprehensive score of traffic intention include: performing comprehensive scoring and screening of each candidate path cluster in the candidate path cluster set, aggregating the behavioral feature vectors of all trajectory segments within it, calculating the mean and variance of the behavioral features within the cluster, and weighting and summing the normalized average speed, speed variance, directional stability, path curvature and time distribution entropy according to preset weight coefficients to obtain the comprehensive score of traffic intention.

7. The method for generating an outdoor three-dimensional pedestrian navigation road network map by fusing crowdsourced trajectory data according to claim 2, characterized in that, The state vector of the Kalman filter algorithm is a six-dimensional vector containing three-dimensional position coordinates and three-dimensional velocity components. The observation vector is the three-dimensional coordinates output by the global positioning system module. The state transition matrix is ​​constructed based on the uniform motion model. The observation matrix is ​​an identity matrix. The process noise covariance matrix and the observation noise covariance matrix are set according to the sensor characteristic parameters.

8. The method for generating an outdoor three-dimensional pedestrian navigation road network map by fusing crowdsourced trajectory data according to claim 6, characterized in that, The weight coefficients are obtained by training on a dataset containing labeled real paths and non-path samples, using a gradient descent algorithm to optimize classification accuracy.

9. The method for generating an outdoor three-dimensional pedestrian navigation road network map by fusing crowdsourced trajectory data according to claim 1, characterized in that, The master curve algorithm uses the first-order principal component axis of the path cluster point set as the initial curve, and generates the center line by iteratively executing projection and local weighted regression smoothing steps until the curve converges or reaches the maximum number of iterations.

10. The method for generating an outdoor three-dimensional pedestrian navigation road network map by fusing crowdsourced trajectory data according to claim 1, characterized in that, The determination of spatial intersection points is based on the detection of line segment intersections on the horizontal projection plane.

Images