An intersection recognition method and system based on multi-features of crowd-sourced trajectory data

Abstract

This invention proposes a method and system for intersection recognition based on multiple features of crowdsourced trajectory data, relating to the technical field of intersection recognition. The method includes: collecting crowdsourced vehicle trajectory data covering a target road area to obtain trajectory vector data; cropping the trajectory vector data to retain the target trajectory vector data; spatially rasterizing the target study area, mapping trajectory points to two-dimensional raster cells to construct multi-channel trajectory feature raster data; spatially aligning and overlaying multiple sets of multi-channel trajectory feature raster data; selecting multiple sample collection areas from the obtained unified multi-channel trajectory feature raster data; labeling the intersections within each sample collection area to form a sample dataset; and supervising the training of an intersection recognition model based on the sample dataset to obtain intersection recognition results and performance evaluation metrics. This application helps improve the accuracy and stability of intersection recognition.

Description

Technical Field

[0001] This invention relates to the field of intersection recognition technology, and in particular to an intersection recognition method and system based on multiple features of crowdsourced trajectory data. Background Technology

[0002] With the acceleration of urbanization and the continuous development of urban transportation systems, road intersections, as important nodes in urban road networks, play a crucial role in traffic organization, operational efficiency improvement, traffic safety management, and the construction of intelligent transportation systems. Intersections are typically areas where traffic conflicts are most concentrated, and where traffic accidents and congestion are most frequent. Their spatial distribution and structural characteristics have a significant impact on road network analysis, traffic simulation, signal control optimization, and intelligent driving applications. Therefore, achieving accurate and efficient identification of urban road intersections has significant practical importance and application value.

[0003] Chinese patent CN112364890B discloses a method for intersection guidance in creating a navigable urban road network using taxi trajectories. This method includes: introducing density features, intersection connection features, and direction features; performing multi-method integration for intersection identification and semi-supervised spurious removal; and detecting intersection locations. Based on the intersection identification results, considering Delaunay triangulation adjacency features, trajectory distribution features, and road geometric features, the method identifies the topological relationships between intersections and integrates morphological results to generate road segment geometry. Based on the identified intersection and road segment information, the method analyzes the road segment connection information of taxi trajectories at the intersections, and then calculates the intersection turning rules and the one-way / two-way attributes of the road segments. However, the above scheme relies on kernel density analysis, morphological refinement, and various manual threshold rules to screen candidate intersections in multiple rounds. The classification stage still uses traditional random forest priority to ensure the interpretability of rules and features, making the method highly sensitive to parameter settings and data quality. In urban road networks with uneven traffic flow distribution or complex intersection morphologies, it is prone to missed and false detections of intersections. Therefore, it is essential to provide a method and system for intersection recognition based on multiple features of crowdsourced trajectory data to improve the accuracy and stability of intersection recognition. Summary of the Invention

[0004] In view of this, the present invention proposes a method and system for intersection recognition based on multiple features of crowdsourced trajectory data, which helps to improve the accuracy and stability of intersection recognition.

[0005] This invention provides a method for intersection identification based on multiple features of crowdsourced trajectory data, the method comprising:

[0006] Collect crowdsourced vehicle trajectory data covering the target road area, preprocess the crowdsourced vehicle trajectory data, and obtain trajectory vector data under a unified geographic coordinate system;

[0007] The trajectory vector data is cropped to retain the target trajectory vector data corresponding to the target research area;

[0008] The target study area is spatially rasterized to map the trajectory points in the target trajectory vector data to two-dimensional raster cells to obtain two-dimensional mapping results. Feature extraction and normalization operations are then performed on the two-dimensional mapping results to construct multi-channel trajectory feature raster data.

[0009] Multiple sets of multi-channel trajectory feature raster data are spatially aligned and overlaid to obtain unified multi-channel trajectory feature raster data covering the target research area. Multiple sample collection areas are selected from the unified multi-channel trajectory feature raster data, and the intersections in each sample collection area are labeled to form a sample dataset.

[0010] The intersection recognition model is trained under supervision based on the sample dataset. Multi-channel trajectory feature raster data of the area to be recognized is input into the trained intersection recognition model to obtain the intersection recognition results and the performance evaluation index corresponding to the intersection recognition results.

[0011] Based on the above technical solutions, preferably, the collection of crowdsourced vehicle trajectory data covering the target road area specifically includes:

[0012] Normally operating taxis are selected as the data collection carriers for crowdsourced vehicle trajectory data. Initial vehicle trajectory data of taxis in a unified geographic coordinate system are collected at preset time intervals. The initial vehicle trajectory data includes spatial location coordinates, timestamps, and driving speed information.

[0013] The initial vehicle trajectory data is sequentially processed by removing outliers, filtering duplicates, and correcting the time sequence to obtain the vehicle trajectory data corresponding to each taxi. The vehicle trajectory data corresponding to all taxis are then integrated to obtain the crowdsourced vehicle trajectory data.

[0014] Based on the above technical solutions, preferably, the step of performing spatial rasterization processing on the target research area, mapping the trajectory points in the target trajectory vector data to two-dimensional raster cells, specifically includes:

[0015] Based on the horizontal coordinate range, vertical coordinate range, and preset grid resolution of the target study area, the target study area is divided into regularly arranged two-dimensional grid units.

[0016] Based on the spatial coordinates of each trajectory point in the target trajectory vector data, each trajectory point is mapped to the corresponding two-dimensional grid cell, and the number of trajectory points in each two-dimensional grid cell is counted to construct the trajectory point density feature.

[0017] The vehicle speeds corresponding to trajectory points falling within the same two-dimensional grid cell are statistically analyzed to obtain the vehicle speed characteristics corresponding to the two-dimensional grid cell.

[0018] Based on the coordinates of adjacent trajectory points in the target trajectory vector data, calculate the driving direction angle and direction angle change for each group of adjacent trajectory points, and statistically analyze the direction angle change within the same two-dimensional grid cell to obtain the average direction change characteristics corresponding to the two-dimensional grid cell.

[0019] More preferably, the step of spatially aligning and overlaying multiple sets of multi-channel trajectory feature raster data to obtain unified multi-channel trajectory feature raster data covering the target study area specifically includes:

[0020] Based on the target study area, spatial raster resolution, and two-dimensional raster row and column indexing rules, spatial alignment is performed on multi-channel trajectory feature raster data from different vehicles and different time periods so that raster units with the same row and column numbers in each group of multi-channel trajectory feature raster data correspond to the same geographic spatial location.

[0021] For multiple sets of multi-channel trajectory feature raster data located in the same two-dimensional raster unit, the trajectory point density feature, vehicle speed feature, and average direction change feature corresponding to the trajectory points in the multi-channel trajectory feature raster data are statistically averaged in the vehicle dimension to obtain the trajectory point density fusion feature, vehicle speed fusion feature, and average direction change fusion feature after the two-dimensional raster unit is fused. The trajectory point density fusion feature, vehicle speed fusion feature, and average direction change fusion feature are then integrated to form the fusion feature vector corresponding to the two-dimensional raster unit.

[0022] The fused feature vectors corresponding to all two-dimensional grid cells within the target study area are reorganized in a channel manner to construct unified multi-channel trajectory feature grid data covering the target study area.

[0023] More preferably, multiple sample collection areas are selected from the unified multi-channel trajectory feature raster data, and intersections within each sample collection area are labeled, specifically including:

[0024] In the unified multi-channel trajectory feature raster data, combined with the road structure distribution and the spatial variation characteristics of trajectory features, multiple independent sample collection areas are selected on the unified multi-channel trajectory feature raster data. The sample collection areas include dense intersection areas, main road areas, and ordinary road sections.

[0025] More preferably, the supervised training of the intersection recognition model based on the sample dataset specifically includes:

[0026] The sample dataset is divided into a training set and a validation set. The training set is used for parameter learning of the intersection recognition model, and the validation set is used for parameter tuning and overfitting monitoring of the intersection recognition model.

[0027] Using the multi-channel trajectory feature raster data in the sample dataset as model input, and the intersection annotation results corresponding to the multi-channel trajectory feature raster data as supervision labels, a joint loss function is constructed, and the parameters of the intersection recognition model are iteratively updated based on the joint loss function;

[0028] When the performance evaluation index of the intersection recognition model on the validation set reaches the preset convergence condition, the training is terminated and the trained intersection recognition model is obtained.

[0029] More preferably, the joint loss function includes a binary cross-entropy loss for measuring pixel-level classification error, a Dice loss for constraining the degree of region overlap, and a direction change weighted loss that assigns different weights to different grid positions based on the magnitude of trajectory direction change.

[0030] A second aspect of this application provides an intersection recognition system based on multiple features of crowdsourced trajectory data. The intersection recognition system includes a data acquisition module, a data processing module, and an intersection recognition module.

[0031] The data acquisition module is used to collect crowdsourced vehicle trajectory data covering the target road area, preprocess the crowdsourced vehicle trajectory data, and obtain trajectory vector data under a unified geographic coordinate system.

[0032] The data processing module is used to perform region clipping on the trajectory vector data to retain the target trajectory vector data corresponding to the target research area, perform spatial rasterization processing on the target research area, map the trajectory points in the target trajectory vector data to two-dimensional raster cells to obtain two-dimensional mapping results, and perform feature extraction and normalization operations on the two-dimensional mapping results in sequence to construct multi-channel trajectory feature raster data. Multiple sets of multi-channel trajectory feature raster data are spatially aligned and superimposed and fused to obtain unified multi-channel trajectory feature raster data covering the target research area. Multiple sample collection areas are selected in the unified multi-channel trajectory feature raster data, and intersections in each sample collection area are labeled to form a sample dataset.

[0033] The intersection recognition module is used to supervise the training of the intersection recognition model based on the sample dataset. The multi-channel trajectory feature raster data of the area to be recognized is input into the trained intersection recognition model to obtain the intersection recognition results and the performance evaluation index corresponding to the intersection recognition results.

[0034] A third aspect of this application provides an electronic device including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory.

[0035] A fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of a method for intersection recognition based on multiple features of crowdsourced trajectory data.

[0036] The intersection recognition method and system based on multiple features of crowdsourced trajectory data provided by this invention have the following advantages over existing technologies:

[0037] (1) By extracting multiple features from the trajectories of crowdsourced vehicles and rasterizing them into multi-channel feature data, compared with relying on a single feature or traditional map geometric information, it can more comprehensively depict the behavior patterns of vehicles near intersections, thereby significantly improving the accuracy and stability of intersection recognition. The spatial alignment and superposition fusion of multiple sets of trajectory features effectively weakens the adverse effects of noise and abnormal trajectories in single-collection data, making the recognition results more robust to data loss, noise and uneven sampling. Furthermore, by integrating trajectory data preprocessing, rasterization, multi-feature construction, sample labeling and supervised learning into one, it realizes the automated processing from the original crowdsourced trajectories to the intersection recognition results, significantly reducing the workload of manual interpretation and rule design. While outputting the intersection recognition results, it provides corresponding performance evaluation indicators, which facilitates the comparison, evaluation and targeted optimization of different model structures, different feature combinations and different parameter settings. It can form a closed-loop iterative mechanism to further improve the accuracy and stability of intersection recognition.

[0038] (2) By dividing the target study area into regular two-dimensional grids and statistically analyzing various features such as trajectory point density, vehicle speed and average direction change at the grid cell scale, fine-grained expression of vehicle spatial distribution, operating status and steering behavior can be achieved. The trajectory point density feature can highlight areas where traffic converges and diverges significantly, which is helpful for identifying the location of intersections where traffic gathers. Furthermore, by statistically aggregating within the grid cells, the influence of positioning error, instantaneous abnormal speed and individual abnormal steering in a single trajectory on the overall features can be effectively smoothed and weakened, making the constructed density, speed and direction change features more stable. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating the intersection identification method based on multiple features of crowdsourced trajectory data provided by the present invention.

[0041] Figure 2 This is a schematic diagram of the trajectory vector data rasterization process provided by the present invention;

[0042] Figure 3 A schematic diagram illustrating the construction of a multi-feature three-channel grid based on crowdsourced trajectory data provided by this invention;

[0043] Figure 4 This is a schematic diagram of manual marking of intersection areas provided by the present invention;

[0044] Figure 5 A schematic diagram illustrating the model training of the intersection recognition model provided by this invention;

[0045] Figure 6 This is a schematic diagram of the intersection recognition system provided by the present invention;

[0046] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention.

[0047] Explanation of reference numerals in the attached diagram: 1. Intersection recognition system; 11. Data acquisition module; 12. Data processing module; 13. Intersection recognition module; 2. Electronic equipment; 21. Processor; 22. Communication bus; 23. User interface; 24. Network interface; 25. Memory. Detailed Implementation

[0048] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0049] This invention discloses a method for intersection recognition based on multiple features of crowdsourced trajectory data, with reference to... Figure 1 The steps of this method include S1 to S5.

[0050] Step S1: Collect crowdsourced vehicle trajectory data covering the target road area, preprocess the crowdsourced vehicle trajectory data, and obtain trajectory vector data under a unified geographic coordinate system.

[0051] This step also includes steps S11 to S12.

[0052] Step S11: Select normally operating taxis as the data collection carriers for crowdsourced vehicle trajectory data, and collect the initial vehicle trajectory data of taxis in a unified geographic coordinate system at preset time intervals. The initial vehicle trajectory data includes spatial location coordinates, timestamps, and driving speed information.

[0053] Step S12: The initial vehicle trajectory data is sequentially processed by removing outliers, filtering duplicates, and correcting the time sequence to obtain the vehicle trajectory data corresponding to each taxi. The vehicle trajectory data corresponding to all taxis are then integrated to obtain the crowdsourced vehicle trajectory data.

[0054] In this step, the removal of abnormal trajectory points includes calculating the spatial distance between adjacent trajectory points, using the following formula:

[0055]

[0056] in, Indicates the spatial distance between adjacent trajectory points, ( ) indicates the first The x and y coordinates of each trajectory point, ( ) represents the next trajectory point The x and y coordinates.

[0057] The instantaneous speed of the vehicle is calculated using the time difference, and the specific formula is as follows:

[0058]

[0059] in, Indicates the first The instantaneous velocity of each trajectory point This represents the spatial distance between adjacent trajectory points. Indicates the first The timestamps corresponding to each trajectory point Indicates the first The timestamps corresponding to each trajectory point.

[0060] like If the speed exceeds the set maximum reasonable speed threshold, the trajectory point is determined to be an anomaly and removed. The maximum reasonable speed threshold is set based on the speed distribution characteristics statistically analyzed from historical trajectory data. After removing obvious outliers, the upper quantile value of the speed distribution is used, and the threshold can be set separately for different road types.

[0061] If adjacent trajectory points satisfy If the two trajectory points are considered to be duplicates, the duplicate trajectory points are removed, and only one trajectory point is retained for each spatial location.

[0062] Sort the trajectory points according to their timestamps to ensure the trajectory sequence satisfies:

[0063]

[0064] The preprocessed trajectory point data is converted into vector data files in the form of spatial point features. For multiple trajectory files generated by the same vehicle within the acquisition period, they are merged in chronological order to form a complete single-vehicle trajectory vector dataset, providing a unified data foundation for subsequent spatial clipping and feature construction.

[0065] Step S2 involves cropping the trajectory vector data to retain the target trajectory vector data corresponding to the target study area.

[0066] In this step, considering the significant structural differences in urban road networks across different areas, in order to improve the targeting and stability of intersection recognition, this invention selects the central urban area as the research area, and determines the spatial range of the research area to be 30720m×30720m.

[0067] The crowdsourced vehicle trajectory data for taxis obtained in step S1 is spatially cropped according to the study area, retaining only the trajectory points falling within the study area. This cropping operation effectively reduces the amount of data irrelevant to the research objective and ensures that subsequent rasterization processing and model training are performed within a unified spatial reference framework.

[0068] Step S3: Spatial rasterization processing is performed on the target study area to map the trajectory points in the target trajectory vector data to two-dimensional raster cells to obtain two-dimensional mapping results. Feature extraction and normalization operations are then performed on the two-dimensional mapping results to construct multi-channel trajectory feature raster data.

[0069] This step also includes steps S31 to S37.

[0070] Step S31: Based on the horizontal coordinate range, vertical coordinate range and preset grid resolution of the target study area, the target study area is divided into regularly arranged two-dimensional grid units.

[0071] In this step, a uniform spatial raster resolution is set within the target study area, dividing the area into regular two-dimensional raster cells. Assume the study area's extent in the planar coordinate system is... The raster resolution is Then the number of raster rows and columns in the study area are respectively:

[0072]

[0073] in, This indicates the number of two-dimensional grid cells in the horizontal direction. This indicates the number of two-dimensional grid cells in the vertical direction. and These represent the maximum and minimum coordinates of the target study area in the horizontal and vertical directions, respectively. r This represents the side length of a single grid cell. The symbol indicates rounding up.

[0074] Step S32: Based on the spatial coordinates of each trajectory point in the target trajectory vector data, map each trajectory point to the corresponding two-dimensional grid cell, and count the number of trajectory points in each two-dimensional grid cell to construct the trajectory point density feature.

[0075] In this step, for each vehicle trajectory point obtained after cropping in step two... According to its spatial coordinates ( ), mapping it to the corresponding raster cell ( ),like Figure 2 As shown, the mapping relationship is as follows:

[0076]

[0077] in,( ) is the first The x and y coordinates of each trajectory point, ( The numbers () represent the row and column numbers of the raster cell containing the trajectory point, respectively. and These represent the minimum values ​​of the x-coordinate and y-coordinate of the study area, respectively. r This represents the side length of a single grid cell. This indicates the floor function.

[0078] Using a single raster cell as the statistical unit, the number of trajectory points mapped to that raster is counted to construct the trajectory point density feature. Let the () The set of trajectory points contained within each grid cell is Then the trajectory point density feature of the grid Defined as ,in, This indicates the number of trajectory points in the set of trajectory points.

[0079] Step S33: Statistically analyze the vehicle speeds corresponding to trajectory points falling within the same two-dimensional grid cell to obtain the vehicle speed features corresponding to the two-dimensional grid cell.

[0080] In this step, the vehicle speeds corresponding to trajectory points within the same grid cell are statistically analyzed to construct vehicle speed features. Let the grid be... The set of vehicle speeds for the inner trajectory points is Then the vehicle speed characteristic of this grid cell The calculation is as follows:

[0081]

[0082] in, This indicates the vehicle speed characteristic of the grid cell. Indicates the first i The vehicle speed corresponding to each trajectory point. k This indicates the number of trajectory points within the two-dimensional grid cell.

[0083] Step S34: Based on the coordinates of adjacent trajectory points in the target trajectory vector data, calculate the driving direction angle and direction angle change corresponding to each group of adjacent trajectory points, and statistically analyze the direction angle change within the same two-dimensional grid cell to obtain the average direction change characteristics corresponding to the two-dimensional grid cell.

[0084] In this step, the driving direction angle is calculated based on the spatial coordinates of adjacent trajectory points of the same vehicle. The specific calculation formula is as follows:

[0085]

[0086] in, The azimuth of the trajectory point. Indicates the first The x and y coordinates of each trajectory point No. The x and y coordinates of each trajectory point.

[0087] Next, the change in the driving direction angle is calculated using the following formula:

[0088]

[0089] in, Indicates the first The change in the driving direction angle of each trajectory point and They represent the first The trajectory point and the first The driving direction angle of each trajectory point.

[0090] Then, the directional changes within the same grid cell are statistically analyzed to obtain the average directional change characteristic of the grid. The specific formula is as follows:

[0091]

[0092] in, Represents a two-dimensional grid cell directional change characteristics, For the first The change in the driving direction angle of each trajectory point This indicates the number of trajectory points within the grid.

[0093] Furthermore, the trajectory point density features, vehicle speed features, and direction change features are normalized respectively. For any feature... normalization results The calculation formula is as follows:

[0094]

[0095] in, This indicates the specific value of the feature within the raster. This represents the normalization result. and These represent the minimum and maximum values ​​of the feature within the target study area, respectively.

[0096] The normalized trajectory point density features, vehicle speed features, and direction change features are organized according to channels to construct multi-channel trajectory feature raster data, such as... Figure 3 As shown, each grid cell corresponds to a feature vector:

[0097]

[0098] in, This represents the normalized trajectory point density feature. This represents the normalized vehicle speed characteristic. This indicates the directional change characteristics after normalization. Represents a two-dimensional grid cell The corresponding multi-channel trajectory feature raster data.

[0099] Step S35: Based on the target study area, spatial raster resolution, and two-dimensional raster row and column indexing rules, spatial alignment is performed on multi-channel trajectory feature raster data from different vehicles and different time periods, so that raster units with the same row and column numbers in each group of multi-channel trajectory feature raster data correspond to the same geospatial location.

[0100] In this step, based on the construction of multi-channel trajectory feature raster data for single-vehicle trajectories, spatial alignment processing is performed on multi-channel raster data from different vehicles and time periods. Specifically, all single-vehicle trajectory raster data are generated based on the same study area, spatial resolution, and raster indexing rules; therefore, they can be directly aligned based on the raster row and column numbers of the two-dimensional raster units. Alignment is performed under a unified spatial coordinate system to ensure that the trajectory features of different vehicles have consistent spatial semantics within the same grid cell.

[0101] Step S36: For multiple sets of multi-channel trajectory feature raster data located in the same two-dimensional raster unit, the trajectory point density features, vehicle speed features, and average direction change features corresponding to the trajectory points in the multi-channel trajectory feature raster data are statistically averaged in the vehicle dimension to obtain the trajectory point density fusion features, vehicle speed fusion features, and average direction change fusion features after the two-dimensional raster unit is fused. The trajectory point density fusion features, vehicle speed fusion features, and average direction change fusion features are then integrated to form the fusion feature vector corresponding to the two-dimensional raster unit.

[0102] In this step, for cells located in the same two-dimensional grid cell The trajectory feature data of multiple vehicles are spatially overlaid to comprehensively reflect the traffic operation characteristics of that spatial location under conditions of multiple vehicles and multiple time periods. The fused trajectory feature vector is obtained by spatially overlaying the trajectory features of multiple vehicles using an averaging method, as shown in the following formula:

[0103]

[0104] in, Represents a two-dimensional grid cell The fused multi-vehicle trajectory features Indicates the first Vehicles in two-dimensional grid cells Trajectory point density features within, Indicates the first Vehicles in two-dimensional grid cells Vehicle speed characteristics within, Indicates the first Vehicles in two-dimensional grid cells Internal directional change characteristics, Represents a two-dimensional grid cell Number of vehicles in China.

[0105] Step S37: Reorganize the fused feature vectors corresponding to all two-dimensional grid cells within the target study area in a channel manner to construct unified multi-channel trajectory feature grid data covering the target study area.

[0106] In this step, the multi-vehicle trajectory feature results after spatial overlay processing are reorganized according to channel arrangement to construct unified multi-channel trajectory feature raster data covering the target study area. For any raster cell within the study area... Its corresponding multi-channel feature representation is as follows:

[0107]

[0108] in, This represents the trajectory density features after fusion. This indicates the vehicle speed characteristics after fusion. This indicates the directional change characteristics after fusion.

[0109] In this embodiment, by dividing the target study area into regular two-dimensional grids and statistically analyzing various features such as trajectory point density, vehicle speed, and average direction change at the grid cell scale, a fine-grained expression of vehicle spatial distribution, operating status, and steering behavior is achieved. Trajectory point density features highlight areas of significant traffic convergence and divergence, facilitating the identification of intersections where traffic congestion occurs. Furthermore, by statistically aggregating within the grid cells, the impact of positioning errors, instantaneous abnormal speeds, and individual abnormal turns on the overall features of a single trajectory is effectively smoothed and weakened, making the constructed density, speed, and direction change features more stable. Vehicle speed features reflect operational characteristics such as deceleration and stopping, helping to distinguish intersection areas from ordinary straight-ahead road sections. Average direction change features highlight areas with frequent vehicle turns, aiding in the identification of intersections with multiple merging directions or concentrated turns. The unified representation of multiple features in the same grid space improves the model's ability to distinguish between intersection and non-intersection areas.

[0110] Transforming complex trajectory vector data into regular two-dimensional raster feature data naturally adapts to the input format of rasterized deep learning models such as convolutional neural networks, simplifying feature organization and model design. Simultaneously, the regular raster structure facilitates parallel computation and rapid storage and retrieval, improving overall processing efficiency. By presetting the raster resolution and flexibly dividing the raster according to the target area's coordinate range, adjustments can be made based on different cities, road densities, and precision requirements. This ensures both the precision of feature representation and manages computational overhead, enhancing applicability and portability across various application scenarios.

[0111] Step S4 involves spatially aligning and overlaying multiple sets of multi-channel trajectory feature raster data to obtain unified multi-channel trajectory feature raster data covering the target study area. Multiple sample collection areas are then selected from the unified multi-channel trajectory feature raster data, and intersections within each sample collection area are labeled to form a sample dataset.

[0112] In this step, based on the road structure distribution and spatial variation characteristics of trajectory features, multiple independent sample collection areas are selected in the unified multi-channel trajectory feature raster data. These sample collection areas include dense intersection areas, main road areas, and ordinary road sections.

[0113] In one example, based on the fused unified multi-channel trajectory feature raster data, and considering the road structure distribution and spatial variation characteristics of trajectory features, several representative raster datasets with a size of 7680m were selected. The sample area is 7680m. The sample area preferably includes areas with dense intersections, main roads, and ordinary road sections to ensure the diversity and representativeness of the sample data in terms of spatial structure and traffic behavior characteristics.

[0114] The selected sample region is processed for intersection labeling. The intersection locations in the sample region are labeled to generate a binary labeled image. ,like Figure 4 As shown, where:

[0115]

[0116] Among them, the intersection labeling process refers to the use of a variable radius circular area to represent intersections of different structural scales during the labeling stage. The radius of the circle is determined based on the road geometry and three-lane feature data.

[0117] Step S5: Supervised training of the intersection recognition model is performed based on the sample dataset. Multi-channel trajectory feature raster data of the area to be recognized is input into the trained intersection recognition model to obtain the intersection recognition results and the corresponding performance evaluation indicators.

[0118] This step also includes steps S51 to S53.

[0119] Step S51: Divide the sample dataset into a training set and a validation set. The training set is used for parameter learning of the intersection recognition model, and the validation set is used for parameter tuning and overfitting monitoring of the intersection recognition model.

[0120] Step S52: Using the multi-channel trajectory feature raster data in the sample dataset as the model input, and the intersection annotation results corresponding to the multi-channel trajectory feature raster data as the supervision label, a joint loss function is constructed, and the parameters of the intersection recognition model are iteratively updated based on the joint loss function.

[0121] In this step, the joint loss function includes a binary cross-entropy loss to measure pixel-level classification error, a Dice loss to constrain the degree of region overlap, and a direction change weighted loss that assigns different weights to different grid positions based on the magnitude of trajectory direction change.

[0122] Step S53: When the performance evaluation index of the intersection recognition model on the validation set reaches the preset convergence condition, the training is terminated and the trained intersection recognition model is obtained.

[0123] In this embodiment, by extracting multiple features from the trajectories of crowdsourced vehicles and rasterizing them into multi-channel feature data, compared to relying solely on a single feature or traditional map geometric information, this approach can more comprehensively depict the behavioral patterns of vehicles near intersections, thereby significantly improving the accuracy and stability of intersection recognition. The spatial alignment and overlay fusion of multiple sets of trajectory features effectively mitigate the adverse effects of noise and abnormal trajectories in single-collection data, making the recognition results more robust to data loss, noise, and uneven sampling. Furthermore, by integrating trajectory data preprocessing, rasterization, multi-feature construction, sample labeling, and supervised learning into one system, the process from raw crowdsourced trajectories to intersection recognition results is automated, significantly reducing the workload of manual interpretation and rule design. While outputting the intersection recognition results, corresponding performance evaluation indicators are provided, facilitating comparative evaluation and targeted optimization of different model structures, feature combinations, and parameter settings. This can form a closed-loop iterative mechanism, further improving the accuracy and stability of intersection recognition.

[0124] In one example, such as Figure 5 As shown, each sample consists of input features and output labels. The input features are multi-channel trajectory feature raster data, and the output labels are binary label masks that spatially correspond one-to-one with the input raster. The input sample is represented as follows: The output labels are represented as: .

[0125] in, and These represent the height and width of the raster sample in space, respectively. This represents the number of trajectory feature channels, which includes... , as well as , Represents the density channels of trajectory points. Indicates the average vehicle speed channel. Indicates a channel for changing direction; A value of 1 indicates an intersection area, and a value of 0 indicates a non-intersection area.

[0126] The sample dataset was divided into training, validation, and test sets at a ratio of 70% / 15% / 15% for model training, parameter tuning, and performance evaluation.

[0127] By setting up feature-aware convolutional layers, weighted mapping is performed on different trajectory feature channels to obtain the initial feature representation, as follows:

[0128]

[0129] in, Indicates the first Each trajectory feature channel Indicates the relationship with the first The convolution kernel weights corresponding to each trajectory feature channel This represents the convolution operation. Indicates the bias term. This represents a non-linear activation function. It enables the model to distinguish and represent different types of trajectory behavior features in the initial stage of feature encoding.

[0130] In each skip connection of the Attention UNet++ network, attention weights are calculated based on encoder features, decoder guidance features, and trajectory behavior features. The calculation method for attention weights is as follows:

[0131]

[0132] in, This represents the feature map from the encoder. This represents the guiding feature map from the decoder. This represents a behavior feature map formed by fusing trajectory direction change features with vehicle speed features. This represents the convolution weights corresponding to the encoder feature map. This represents the convolution weights corresponding to the decoder's guiding feature map. This represents the convolution weights corresponding to the behavioral feature map. This represents the joint attention weight graph.

[0133] By applying attention weights to the encoder features, the enhanced feature representation is obtained, as follows:

[0134]

[0135] in, This indicates element-wise multiplication.

[0136] To address the diversity of intersections in terms of spatial scale and morphological structure, multi-scale convolution processing is performed on the feature map at the decoding node. The specific formula is as follows:

[0137]

[0138] in, Indicates the expansion rate Convolution operation, i =1, 2, 3 , as well as Represents parameters at different scales. This indicates a feature concatenation operation. This represents the multi-scale decoding features after fusion.

[0139] The model is trained using a joint loss function, which includes the basic segmentation loss, the region overlap constraint loss, and the trajectory behavior weighted loss, specifically:

[0140] The binary cross-entropy loss function is expressed as:

[0141]

[0142] The Dice loss function is expressed as:

[0143]

[0144] The weighted loss function based on direction change is expressed as:

[0145]

[0146] Among them, the weighting coefficient Defined as: In each loss function In the intersection recognition model, the first... Predicted values ​​for each trajectory point Indicates the first The true labeled value of each trajectory point This represents a constant used to prevent the denominator from being zero. This represents the weight adjustment parameter. Indicates the first The change in the driving direction angle of each trajectory point This represents the total number of samples involved in the calculation of the loss function.

[0147] The joint loss function is expressed as:

[0148]

[0149] in, These represent the weight coefficients of the binary cross-entropy loss function. The weight coefficients of the Dice loss function are represented. This represents the weight coefficients of the weighted loss function based on direction change.

[0150] After the intersection recognition model is trained, test set samples are input into the trained model to obtain the corresponding predicted probability map; by setting a threshold... The prediction results are binarized to obtain the final intersection recognition result.

[0151] The model's prediction results are compared with the actual labeled results, and the model performance is evaluated using precision, recall, and F1 score. The calculation formulas are as follows:

[0152]

[0153]

[0154]

[0155] Wherein, TP represents the number of grid cells that are actually labeled as intersections and the intersection recognition model predicts as intersections; FP represents the number of grid cells that are actually labeled as non-intersections but the intersection recognition model predicts as intersections; and FN represents the number of grid cells that are actually labeled as intersections but the intersection recognition model predicts as non-intersections.

[0156] By fusing trajectory point density features, vehicle speed features, and directional angle change features, this invention comprehensively characterizes intersections from both spatial distribution and traffic behavior perspectives. Compared to existing methods that rely solely on single trajectory density or simple statistical features, this invention can more accurately distinguish intersection areas from ordinary road areas, especially under conditions of uneven traffic flow, significant differences in vehicle travel paths, or complex road structures, maintaining high recognition accuracy and stability while effectively reducing the probability of false and missed identifications. By fusing trajectory data collected from multiple taxis over a long time scale, the invention effectively mitigates interference from the randomness of individual vehicle travel paths, short-term abnormal traffic behavior, and uneven data sampling. The spatial characteristics of intersection areas are continuously enhanced by vehicle convergence, deceleration, and directional changes in multi-vehicle, multi-time-period data, making them more significant and stable, thereby improving the reliability and generalization ability of intersection recognition results in large-scale urban areas.

[0157] This invention transforms the intersection recognition problem into a semantic segmentation problem based on multi-channel trajectory raster data. An attention mechanism is introduced to adaptively weight different feature channels and their spatial locations, enabling the model to focus on key discriminative feature regions within the intersection area. This technique allows the invention to maintain high recognition accuracy and completeness even at multi-way intersections, irregular intersections, and densely road-bound urban centers, significantly outperforming traditional rule-based or simple classification-based recognition methods. It does not rely on complete, high-precision road centerlines or prior intersection location data, but rather on crowdsourced trajectory data generated during natural vehicle movement for intersection recognition. This technique makes the invention applicable to areas with incomplete, outdated, or non-existent high-precision maps, significantly improving the applicability and promotional value of the method in different cities and road environments.

[0158] Based on the above method, this application discloses an intersection recognition system based on multiple features of crowdsourced trajectory data, with reference to... Figure 6 The intersection recognition system 1 includes a data acquisition module 11, a data processing module 12, and an intersection recognition module 13, wherein...

[0159] Data acquisition module 11 is used to collect the trajectory data of multiple vehicles covering the target road area, preprocess the trajectory data of multiple vehicles, and obtain trajectory vector data under a unified geographic coordinate system.

[0160] The data processing module 12 is used to perform regional cropping on the trajectory vector data to retain the target trajectory vector data corresponding to the target study area, perform spatial rasterization processing on the target study area, map the trajectory points in the target trajectory vector data to two-dimensional raster cells to obtain two-dimensional mapping results, and perform feature extraction and normalization operations on the two-dimensional mapping results in sequence to construct multi-channel trajectory feature raster data. Multiple sets of multi-channel trajectory feature raster data are spatially aligned and superimposed and fused to obtain unified multi-channel trajectory feature raster data covering the target study area. Multiple sample collection areas are selected in the unified multi-channel trajectory feature raster data, and the intersections in each sample collection area are labeled to form a sample dataset.

[0161] The intersection recognition module 13 is used to supervise the training of the intersection recognition model based on the sample dataset. It inputs the multi-channel trajectory feature raster data of the area to be recognized into the trained intersection recognition model to obtain the intersection recognition results and the corresponding performance evaluation indicators.

[0162] In one example, the data acquisition module 11 selects normally operating taxis as the data acquisition carriers for crowdsourced vehicle trajectory data. It collects the initial vehicle trajectory data of taxis in a unified geographic coordinate system at preset time intervals. The initial vehicle trajectory data includes spatial location coordinates, timestamps, and driving speed information. The initial vehicle trajectory data is then processed sequentially to remove outliers, filter duplicate points, and correct the time sequence to obtain the vehicle trajectory data corresponding to each taxi. Finally, the vehicle trajectory data corresponding to all taxis are integrated to obtain crowdsourced vehicle trajectory data.

[0163] In one example, the data processing module 12 is used to divide the target study area into regularly arranged two-dimensional grid cells according to the horizontal coordinate range, vertical coordinate range, and preset grid resolution of the target study area; based on the spatial coordinates of each trajectory point in the target trajectory vector data, each trajectory point is mapped to the corresponding two-dimensional grid cell, and the number of trajectory points in each two-dimensional grid cell is counted to construct the trajectory point density feature; the vehicle speed corresponding to the trajectory points falling in the same two-dimensional grid cell is counted to obtain the vehicle speed feature corresponding to the two-dimensional grid cell; based on the coordinates of adjacent trajectory points in the target trajectory vector data, the driving direction angle and direction angle change corresponding to each group of adjacent trajectory points are calculated, and the direction angle change falling in the same two-dimensional grid cell is counted to obtain the average direction change feature corresponding to the two-dimensional grid cell.

[0164] In one example, the data processing module 12 spatially aligns multi-channel trajectory feature raster data from different vehicles and time periods based on the target study area, spatial raster resolution, and two-dimensional raster row and column indexing rules, so that raster cells with the same row and column numbers in each group of multi-channel trajectory feature raster data correspond to the same geographic spatial location. For multiple groups of multi-channel trajectory feature raster data located in the same two-dimensional raster cell, the trajectory point density feature, vehicle speed feature, and average direction change feature corresponding to the trajectory points in the multi-channel trajectory feature raster data are statistically averaged in the vehicle dimension to obtain the trajectory point density fusion feature, vehicle speed fusion feature, and average direction change fusion feature after the two-dimensional raster cell is fused. The trajectory point density fusion feature, vehicle speed fusion feature, and average direction change fusion feature are then integrated to form the fusion feature vector corresponding to the two-dimensional raster cell. The fusion feature vectors corresponding to all two-dimensional raster cells in the target study area are reorganized in a channel manner to construct unified multi-channel trajectory feature raster data covering the target study area.

[0165] In one example, the data processing module 12 is used to select multiple independent sample collection areas in the unified multi-channel trajectory feature raster data, combining the road structure distribution and the spatial variation characteristics of trajectory features. The sample collection areas include dense intersection areas, main road areas and ordinary road sections.

[0166] In one example, the intersection recognition module 13 is used to divide the sample dataset into a training set and a validation set. The training set is used for parameter learning of the intersection recognition model, and the validation set is used for parameter tuning and overfit monitoring of the intersection recognition model. The multi-channel trajectory feature raster data in the sample dataset is used as the model input, and the intersection annotation results corresponding to the multi-channel trajectory feature raster data are used as supervision labels to construct a joint loss function. The parameters of the intersection recognition model are iteratively updated based on the joint loss function. When the performance evaluation index of the intersection recognition model on the validation set reaches the preset convergence condition, the training is terminated, and the trained intersection recognition model is obtained.

[0167] In one example, the joint loss function includes a binary cross-entropy loss to measure pixel-level classification error, a Dice loss to constrain the degree of region overlap, and a direction change weighted loss that assigns different weights to different grid positions based on the magnitude of trajectory direction change.

[0168] Please see Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 As shown, the electronic device 2 may include: at least one processor 21, at least one network interface 24, user interface 23, memory 25, and at least one communication bus 22.

[0169] The communication bus 22 is used to enable communication between these components.

[0170] The user interface 23 may include a display screen and a camera. Optionally, the user interface 23 may also include a standard wired interface and a wireless interface.

[0171] The network interface 24 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0172] The processor 21 may include one or more processing cores. The processor 21 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 25, and by calling data stored in the memory 25. Optionally, the processor 21 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 21 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor 21.

[0173] The memory 25 may include random access memory (RAM) or read-only memory. Optionally, the memory 25 may include non-transitory computer-readable storage medium. The memory 25 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 25 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 25 may also be at least one storage device located remotely from the aforementioned processor 21. Figure 7 As shown, the memory 25, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for an intersection recognition method based on multiple features of crowdsourced trajectory data.

[0174] exist Figure 7In the electronic device 2 shown, the user interface 23 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 21 can be used to call an application program stored in the memory 25 that is a crossroads recognition method based on multiple features of crowdsource trajectory data. When executed by one or more processors, the electronic device executes one or more methods as described in the above embodiments.

[0175] A non-transitory computer-readable storage medium stores instructions that, when executed by one or more processors, cause a computer to perform one or more methods as described in the above embodiments.

[0176] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0177] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0178] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

[0179] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0180] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0181] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0182] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intersection recognition based on multiple features of crowdsourced trajectory data, characterized in that, The method includes: Collect crowdsourced vehicle trajectory data covering the target road area, preprocess the crowdsourced vehicle trajectory data, and obtain trajectory vector data under a unified geographic coordinate system; The trajectory vector data is cropped to retain the target trajectory vector data corresponding to the target study area; Spatial rasterization is performed on the target study area to map the trajectory points in the target trajectory vector data to two-dimensional raster cells to obtain two-dimensional mapping results. Feature extraction and normalization operations are then performed on the two-dimensional mapping results to construct multi-channel trajectory feature raster data. Multiple sets of multi-channel trajectory feature raster data are spatially aligned and overlaid to obtain unified multi-channel trajectory feature raster data covering the target study area. Multiple sample collection areas are selected from the unified multi-channel trajectory feature raster data, and the intersections in each sample collection area are labeled to form a sample dataset. The intersection recognition model is trained under supervision based on the sample dataset. Multi-channel trajectory feature raster data of the area to be recognized is input into the trained intersection recognition model to obtain the intersection recognition results and the corresponding performance evaluation indicators. Supervised training includes: The sample dataset is divided into a training set and a validation set; Using multi-channel trajectory feature raster data from the sample dataset as model input, and the intersection annotation results corresponding to the multi-channel trajectory feature raster data as supervision labels, a joint loss function is constructed, and the parameters of the intersection recognition model are iteratively updated based on the joint loss function. The joint loss function includes binary cross-entropy loss, Dice loss, and direction change weighted loss that assigns different weights to different raster positions based on the magnitude of trajectory direction change. When the performance evaluation index of the intersection recognition model on the validation set reaches the preset convergence condition, the training is terminated and the trained intersection recognition model is obtained. Joint loss function Represented as: ； ； ； The binary cross-entropy loss function is... The Dice loss function, The loss function is weighted by the change of direction. These are weighting coefficients. In the intersection recognition model, the first Predicted values ​​for each trajectory point For the first The true labeled value of each trajectory point For weight adjustment parameters, For the first Change in the driving direction angle of each trajectory point The total number of samples participating in the loss function calculation. These are the weighting coefficients of the binary cross-entropy loss function. These are the weighting coefficients of the Dice loss function. The weighting coefficients are used for the weighted loss function based on the change of direction.

2. The intersection recognition method based on multiple features of crowdsourced trajectory data as described in claim 1, characterized in that, The collection of crowdsourced vehicle trajectory data covering the target road area specifically includes: Normally operating taxis are selected as the data collection carriers for crowdsourced vehicle trajectory data. Initial vehicle trajectory data of taxis in a unified geographic coordinate system are collected at preset time intervals. The initial vehicle trajectory data includes spatial location coordinates, timestamps, and driving speed information. The initial vehicle trajectory data is sequentially processed by removing outliers, filtering duplicates, and correcting the time sequence to obtain the vehicle trajectory data corresponding to each taxi. The vehicle trajectory data corresponding to all taxis are then integrated to obtain the crowdsourced vehicle trajectory data.

3. The intersection recognition method based on multiple features of crowdsourced trajectory data as described in claim 1, characterized in that, The step of performing spatial rasterization processing on the target study area, mapping trajectory points in the target trajectory vector data to two-dimensional raster cells, specifically includes: Based on the horizontal coordinate range, vertical coordinate range, and preset grid resolution of the target study area, the target study area is divided into regularly arranged two-dimensional grid units. Based on the spatial coordinates of each trajectory point in the target trajectory vector data, each trajectory point is mapped to the corresponding two-dimensional grid cell, and the number of trajectory points in each two-dimensional grid cell is counted to construct the trajectory point density feature. The vehicle speeds corresponding to trajectory points falling within the same two-dimensional grid cell are statistically analyzed to obtain the vehicle speed characteristics corresponding to the two-dimensional grid cell. Based on the coordinates of adjacent trajectory points in the target trajectory vector data, calculate the driving direction angle and direction angle change for each group of adjacent trajectory points, and statistically analyze the direction angle change within the same two-dimensional grid cell to obtain the average direction change characteristics corresponding to the two-dimensional grid cell.

4. The intersection recognition method based on multiple features of crowdsourced trajectory data as described in claim 3, characterized in that, The step of spatially aligning and overlaying multiple sets of multi-channel trajectory feature raster data to obtain unified multi-channel trajectory feature raster data covering the target study area specifically includes: Based on the target study area, spatial raster resolution, and two-dimensional raster row and column indexing rules, spatial alignment is performed on multi-channel trajectory feature raster data from different vehicles and different time periods so that raster units with the same row and column numbers in each group of multi-channel trajectory feature raster data correspond to the same geographic spatial location. For multiple sets of multi-channel trajectory feature raster data located in the same two-dimensional raster unit, the trajectory point density feature, vehicle speed feature, and average direction change feature corresponding to the trajectory points in the multi-channel trajectory feature raster data are statistically averaged in the vehicle dimension to obtain the trajectory point density fusion feature, vehicle speed fusion feature, and average direction change fusion feature after the two-dimensional raster unit is fused. The trajectory point density fusion feature, vehicle speed fusion feature, and average direction change fusion feature are then integrated to form the fusion feature vector corresponding to the two-dimensional raster unit. The fused feature vectors corresponding to all two-dimensional grid cells within the target study area are reorganized in a channel manner to construct unified multi-channel trajectory feature grid data covering the target study area.

5. The intersection recognition method based on multiple features of crowdsourced trajectory data as described in claim 1, characterized in that, Multiple sample collection areas are selected from the unified multi-channel trajectory feature raster data, and intersections within each sample collection area are labeled, specifically including: In the unified multi-channel trajectory feature raster data, combined with the road structure distribution and the spatial variation characteristics of trajectory features, multiple independent sample collection areas are selected on the unified multi-channel trajectory feature raster data. The sample collection areas include dense intersection areas, main road areas, and ordinary road sections.

6. A crossroads recognition system based on multiple features of crowdsourced trajectory data, characterized in that, The intersection recognition system (1) includes a data acquisition module (11), a data processing module (12), and an intersection recognition module (13), wherein, The data acquisition module (11) is used to collect the trajectory data of multiple vehicles covering the target road area, preprocess the trajectory data of multiple vehicles, and obtain trajectory vector data under a unified geographic coordinate system. The data processing module (12) is used to perform regional cropping on the trajectory vector data to retain the target trajectory vector data corresponding to the target research area, perform spatial rasterization on the target research area, map the trajectory points in the target trajectory vector data to two-dimensional raster cells to obtain two-dimensional mapping results, and perform feature extraction and normalization operations on the two-dimensional mapping results in sequence to construct multi-channel trajectory feature raster data, perform spatial alignment and superposition fusion on multiple sets of multi-channel trajectory feature raster data to obtain unified multi-channel trajectory feature raster data covering the target research area, and select multiple sample collection areas in the unified multi-channel trajectory feature raster data, and mark the intersections in each sample collection area to form a sample dataset; Supervised training includes: The sample dataset is divided into a training set and a validation set; Using multi-channel trajectory feature raster data from the sample dataset as model input, and the intersection annotation results corresponding to the multi-channel trajectory feature raster data as supervision labels, a joint loss function is constructed, and the parameters of the intersection recognition model are iteratively updated based on the joint loss function. The joint loss function includes binary cross-entropy loss, Dice loss, and direction change weighted loss that assigns different weights to different raster positions based on the magnitude of trajectory direction change. When the performance evaluation index of the intersection recognition model on the validation set reaches the preset convergence condition, the training is terminated and the trained intersection recognition model is obtained. Joint loss function Represented as: ； ； ； The binary cross-entropy loss function is... The Dice loss function, The loss function is weighted by the change of direction. These are weighting coefficients. In the intersection recognition model, the first Predicted values ​​for each trajectory point For the first The true labeled value of each trajectory point For weight adjustment parameters, For the first Change in the driving direction angle of each trajectory point The total number of samples participating in the loss function calculation. These are the weighting coefficients of the binary cross-entropy loss function. These are the weighting coefficients of the Dice loss function. The weighting coefficients of the direction-weighted loss function; The intersection recognition module (13) is used to supervise the training of the intersection recognition model based on the sample dataset. The multi-channel trajectory feature raster data of the area to be recognized is input into the trained intersection recognition model to obtain the intersection recognition result and the performance evaluation index corresponding to the intersection recognition result.

7. An electronic device, characterized in that, The device includes a processor (21), a memory (25), a user interface (23), and a network interface (24). The memory (25) is used to store instructions. The user interface (23) and the network interface (24) are used to communicate with other devices. The processor (21) is used to execute the instructions stored in the memory (25) to cause the electronic device (2) to perform the method as described in any one of claims 1-5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-5.

Images