A road speed limit hierarchical mining method and system considering adjacent road segment information

Abstract

The application discloses a kind of road speed limit level mining method and system considering adjacent road section information.It includes: data preprocessing: carry out road network topology processing, trajectory data cleaning and map matching;Multi-modal feature design: analysis and mining the multiple, multi-order complementary features of road network and trajectory data about road speed limit, form a high discriminant description feature set about road speed limit, as the basis and basis for identifying road speed limit;Road speed limit mining: comprehensively consider the road network features, trajectory features of current road section, left and right adjacent road section, consider high and low speed information, carry out the level mining of road speed limit.Compared with single feature, the integrated road network and trajectory feature improves the accuracy of road speed limit identification;Compared with only considering the target road section for road speed limit identification, considering spatial adjacent information for road speed limit identification has better effect;Compared with directly identifying road speed limit, the road speed limit identification method considering high and low speed information has higher accuracy.

Description

Technical Field

[0001] This invention belongs to the field of navigation information mining, and in particular relates to a method and system for mining road speed limit levels that takes into account information of adjacent road segments. Background Technology

[0002] As a complex graph structure, urban road networks require not only static geometric shapes and spatial topology but also rich navigation attributes to enable core road network applications such as route planning, vehicle navigation, and location services. However, most attribute updates for commercial and free maps still heavily rely on manual annotation, leading to low map data completeness and inaccuracies. For example, the "maxspeed" label defined by OSM (OpenStreetMap) is missing to varying degrees in different cities. Although road speed limit information can be obtained based on the identification of road speed limit signs or the analysis and prediction of factors affecting speed limits, these methods require on-site surveys, which are not only costly to acquire and update but also frequently affected by environmental factors such as weather and lighting. Summary of the Invention

[0003] With the development of location acquisition technology and the widespread use of location acquisition devices, crowdsourced trajectory data has provided new opportunities for the updating and maintenance of navigable road networks. Therefore, considering the need for "speed limit reminders" in navigation systems, this invention integrates trajectory data information with road network data to develop a method for mining speed limit information, which is of great significance for map improvement.

[0004] In view of this, the present invention provides a method for mining road speed limit levels that takes into account information of adjacent road segments, comprising the following steps:

[0005] Step 1, Data Preprocessing: First, perform topology checks on the road network data and quality cleaning on the trajectory data. Then, perform map matching on the road network and trajectory data to connect trajectory points with their respective road segments.

[0006] Step 2, Multimodal Feature Design: Design road network features based on the road network structure, including road geometric features, road distribution features, and road topological connection features; design trajectory features based on road network matching trajectory information, including one-way and two-way road characteristics, road width, road traffic flow, and road operating speed.

[0007] Step 3, Road speed limit mining: Combining the multimodal features in Step 2, a random forest classification model is used as the basic classifier. The road speed limit classifier is obtained by training the model using road speed limit training samples. Then, taking into account the road network and trajectory features of the test sample road segment and its upstream and downstream adjacent road segments, the road speed limit level is identified based on the voting method.

[0008] Furthermore, the road geometric features in step 2 include road geometric length and road curvature;

[0009] Road distribution characteristics include the area of ​​adjacent meshes on the road, the average density of n nodes on the road, and the average line density of n-1 road segments on the road;

[0010] The road topology features include node clustering coefficient, degree centrality, proximity centrality, and betweenness centrality as measured from the perspective of the constituent objects (nodes), and node clustering coefficient, degree centrality, proximity centrality, and betweenness centrality as measured from the perspective of the constituent objects (links).

[0011] Furthermore, the one-way and two-way road features in step 2 include one-way and two-way road information;

[0012] Road traffic flow characteristics include the number of trajectories matched to the road;

[0013] Road width features include the mean, mode, standard deviation, median, and maximum of the absolute values ​​of the distances from the road matching trajectory points to the road centerline, and the 5th to 90th percentiles of the absolute values ​​of the distances from the road matching trajectory points to the road centerline;

[0014] Road speed characteristics include the mean, mode, standard deviation, median, minimum, and maximum speeds of the road matching trajectory points, the 5th to 90th percentiles of the road matching trajectory point speeds, the average instantaneous speeds of the six largest time periods within 24 hours, and the mean, mode, standard deviation, median, minimum, and maximum speed characteristics of the high-speed features.

[0015] Furthermore, the specific implementation method of step 3 is as follows:

[0016] Step 3.1, Adjacent road segment identification: Identify left and right adjacent road segments based on the road stroke construction method;

[0017] Step 3.2, Feature preprocessing: Perform missing value processing and feature selection on the designed multimodal features;

[0018] Step 3.3, Model Training: First, L0-level speed limit recognition is performed based on road network features and trajectory features. Then, L1-level speed limit recognition is optimized by modeling high and low speed limits based on trajectory features. Based on road network features, a random forest model is used to perform L1-level speed limit recognition. Finally, considering the current road segment and its left and right adjacent road segments, a vote is taken on the following six results: L1-level speed limit recognition results for the current road segment based on trajectory features, L1-level speed limit recognition results for the current road segment based on road network features, L1-level speed limit recognition results for the left adjacent road segment based on trajectory features, L1-level speed limit recognition results for the left adjacent road segment based on road network features, L1-level speed limit recognition results for the right adjacent road segment based on trajectory features, and L1-level speed limit recognition results for the right adjacent road segment based on road network features. The final L1-level speed limit recognition result is obtained.

[0019] Furthermore, the specific implementation method of step 3.1 is as follows:

[0020] Based on the road stroke construction method, approximate upstream and downstream adjacent road segments are obtained. The road segment connected to the starting point with the smallest deflection angle of the road segment under study and less than a certain angle is identified as the upstream adjacent road segment, and the road segment connected to the corresponding ending point is identified as the downstream adjacent road segment. For the case where the road network is represented by a double line and there are short connecting road segments, the identification of upstream and downstream adjacent road segments is optimized. That is, if the adjacent road segment is identified as a short connecting road segment, the search continues; if the search finds a long connecting road segment, it is identified as an adjacent road segment; otherwise, the current road segment is identified as an adjacent road segment.

[0021] Furthermore, the specific implementation method of step 3.2 is as follows:

[0022] (3.2.1) Missing value handling: When calculating the relevant features for road attribute identification, the road segment features with missing values ​​based on trajectory data are filled with 0;

[0023] (3.2.2) Feature selection: First, Spearman correlation coefficient is used to analyze the correlation of design features, then important features are extracted based on this, and finally recursive feature elimination method is used for final feature selection.

[0024] The formula for calculating the Spearman correlation coefficient is as follows:

[0025] Ρ =Cov(r X ,r Y ) / σ rX σ rY Formula (1)

[0026] Where X and Y are the original variables, r X r Y The order of the original variables after sorting them is called the rank variable, Cov(r). X ,r Y ) is the covariance of the rank variable, σ rX ,σ rY It is the standard deviation of the rank variable.

[0027] Furthermore, the specific implementation method of step 3.3 is as follows:

[0028] (3.3.1) Speed ​​limit L0 and L1 classification: First, the speed limit is roughly divided into four categories: ultra-low speed, low speed, high speed and ultra-high speed, which is the L0 level speed limit information. Then, each category is further refined to obtain the L1 level speed limit information.

[0029] (3.3.2) Level 0 speed limit recognition: p% of the samples are randomly selected as the training set and (100-p)% are selected as the test set; the training samples in the training set are divided into Level 0 segments to form new training samples. Based on the trajectory features, a random forest classifier is trained to form a Level 0 speed limit recognition classifier; considering the influence of the left and right adjacent road segments on the current road segment speed limit, based on the trajectory features, the Level 0 speed limit recognition classifier is used to perform Level 0 speed limit recognition on the current road segment and the left and right adjacent road segments in the test set respectively. Then, based on the voting method, the three results are fused to realize the Level 0 speed limit recognition of the test set.

[0030] (3.3.3) L1 speed limit recognition: When recognizing road speed limits based on trajectory features, it is difficult to accurately identify ultra-low speed limits and ultra-high speed limits. Only the above two types of speed limits are recognized to the L0 level. For the low-speed and high-speed road sections after L0 classification, a random forest classification model is used to model low classifiers and high classifiers respectively to carry out L1 speed limit recognition optimization.

[0031] When identifying road speed limits based on road network features, a random forest classifier is directly used to train the labeled L1-level speed limit samples to obtain a speed limit classification model, and then the test samples are tested.

[0032] Considering the influence of adjacent road segments on the left and right, the identification results are optimized based on the trajectory characteristics of the current road segment and the adjacent road segments on the left and right, as well as the road network characteristics identification results of the current road segment and the adjacent road segments on the left and right. The voting method is used to fuse the six results and finally obtain the L1 level speed limit identification result.

[0033] Furthermore, in the random forest model, the maximum feature determination method, impurity, maximum tree depth, minimum number of samples contained in a leaf node, and class weights are set to None, gini, None, 1, and balanced, respectively, while the number of weak classifiers is searched based on a fine grid.

[0034] This invention also provides a road speed limit hierarchy discovery system that takes into account information about adjacent road segments, comprising the following modules:

[0035] The preprocessing module is used for data preprocessing: first, it performs topology checks on the road network data and quality cleaning on the trajectory data; then, it performs map matching of the road network and trajectory data to connect the trajectory points with their respective road segments.

[0036] The feature design module is used for multimodal feature design: designing road network features based on the road network structure, including road geometric features, road distribution features, and road topological connection features; and designing trajectory features based on road network matching trajectory information, including one-way and two-way road features, road width, road traffic flow, and road operating speed.

[0037] A speed limit hierarchy mining module is used for road speed limit mining. It combines multimodal features, uses a random forest classification model as the basic classifier, and trains the model using road speed limit training samples to obtain the road speed limit classifier. Then, considering the road network and trajectory features of the test sample road segment and its upstream and downstream adjacent road segments, it identifies the road speed limit hierarchy based on a voting method. Compared with existing technologies, the advantages and beneficial effects of this invention are: This invention comprehensively considers the road geometry, road distribution, road topology, one-way and two-way road features presented by the road network, as well as the road width, traffic flow, and road speed reflected in trajectory data. It also considers the upstream and downstream adjacency information of road segments and the high and low speed information of the trajectory, using a random forest classification model as the basic classifier for road segment speed limit hierarchy identification research. This makes up for the shortcomings of existing work in road speed limit information mining and can effectively solve the problems of long mining cycles and high costs. The method proposed in this invention provides an effective technical reference for speed limit information mining. Attached Figure Description

[0038] Figure 1 This is a flowchart of a road speed limit hierarchy mining method that takes into account information of adjacent road segments, according to an embodiment of the present invention;

[0039] Figure 2 This is a flowchart of the downstream adjacent road segment identification process in this invention.

[0040] Figure 3 This is a flowchart of the L0-level speed limit recognition process in this invention.

[0041] Figure 4 This is a flowchart of the L1-level speed limit recognition optimization process based on trajectory features in this invention.

[0042] Figure 5 This is a flowchart of the L1-level speed limit recognition based on road network features in this invention.

[0043] Figure 6 This is a flowchart of the speed limit recognition process in this invention.

[0044] Figure 7 The experimental comparison diagram is provided for the embodiments of the present invention. Detailed Implementation

[0045] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0046] like Figure 1 As shown, the present invention provides a method for mining road speed limit levels that takes into account information of adjacent road segments, including the following steps:

[0047] Step 1, Data Preprocessing: First, perform topology checks on the road network data and quality cleaning on the trajectory data. Then, perform map matching on the road network and trajectory data to connect trajectory points with their respective road segments.

[0048] Step 2, Multimodal Feature Design: Design road network features based on the road network structure, including road geometric features, road distribution features, and road topological connection features; design trajectory features based on road network matching trajectory information, including one-way and two-way road characteristics, road width, road traffic flow, and road operating speed.

[0049] Step 3, Road speed limit mining: Combining the multimodal features in Step 2, a random forest classification model is used as the basic classifier. The road speed limit classifier is obtained by training the model using road speed limit training samples. Then, taking into account the road network and trajectory features of the test sample road segment and its upstream and downstream adjacent road segments, the road speed limit level is identified based on the voting method.

[0050] Furthermore, the multimodal feature design described in step 2 specifically includes:

[0051] Different roads exhibit different characteristics such as road geometry, road distribution, road topology, one-way and two-way traffic, road flow, and road speed. These characteristics are strongly correlated with road grade and speed limit, and can be mined based on road network and trajectory data. Therefore, a series of corresponding features were designed based on road network and trajectory data as the basis for road grade and speed limit identification, as shown in Table 1.

[0052] Table 1 Statistical information on multimodal features

[0053]

[0054] Furthermore, the road speed limit excavation described in step 3 specifically includes:

[0055] Step 3.1, Adjacent road segment identification: For a road, the attributes of the current road segment are likely to be the same as the corresponding attributes of its left and right adjacent road segments. Therefore, the left and right adjacent road segments are identified based on the road stroke construction method.

[0056] Step 3.2, Feature Preprocessing: In order to ensure the integrity and usability of the data and eliminate feature redundancy during the road speed limit recognition process, it is necessary to process missing values ​​and select features.

[0057] Step 3.3, Model Training: First, L0-level speed limit recognition is performed based on road network features and trajectory features. Then, L1-level speed limit recognition is optimized by modeling high and low speed limits based on trajectory features. In addition, L1-level speed limit recognition is also performed using a random forest model based on road network features. Finally, considering the current road segment and the left and right adjacent road segments, six results are voted on, including the L1-level speed limit recognition results based on trajectory features, the L1-level speed limit recognition results based on road network features, the L1-level speed limit recognition results based on trajectory features, the L1-level speed limit recognition results based on road network features, the L1-level speed limit recognition results based on trajectory features, the L1-level speed limit recognition results based on road network features, the L1-level speed limit recognition results based on trajectory features, the L1-level speed limit recognition results based on road network features, the L1-level speed limit recognition results based on trajectory features, the L1-level speed limit recognition results based on road network features, the L1-level speed limit recognition results based on trajectory features, and the L1-level speed limit recognition results based on road network features for the right adjacent road segment. The final L1-level speed limit recognition result is obtained.

[0058] Furthermore, the specific implementation method of step 3.1 is as follows:

[0059] (1) Based on the road stroke construction method, the approximate upstream and downstream adjacent road segments are obtained. The road segment connected to the starting point with the smallest deflection angle of the road segment under study and less than 30° is identified as the upstream adjacent road segment, and the road segment connected to the corresponding ending point is identified as the downstream adjacent road segment.

[0060] (2) Considering that this paper uses a double-line representation of the road network, there are short connecting road segments with a length of less than 50m. Therefore, this paper optimizes the identification of upstream and downstream adjacent road segments. That is, if an adjacent road segment is identified as a short connecting road segment, the search continues downward. See details below. Figure 2 Flowchart for identifying downstream adjacent road segments.

[0061] Furthermore, the specific implementation method of step 3.2 is as follows:

[0062] (1) Missing value handling: Considering that the missing values ​​are due to the lack of traffic flow, when calculating the road attribute identification related features, the road segment features with missing values ​​based on trajectory data will be filled with 0.

[0063] (2) Feature Selection: First, Spearman correlation coefficient is used to analyze the correlation of design features. Then, based on this, important features are extracted. Finally, recursive feature elimination method is used for final feature selection. The formula for calculating Spearman correlation coefficient is as follows:

[0064] Ρ=Cov(r X ,r Y ) / σ rX σ rY Formula (1)

[0065] Where X and Y are the original variables, r X rY The order of the original variables after sorting them is called the rank variable. Cov(r) X ,r Y ) is the covariance of the rank variable, σ rX ,σ rY It is the standard deviation of the rank variable.

[0066] Furthermore, the specific implementation method of step 3.3 is as follows:

[0067] (1) Classification of Speed ​​Limits (L0 and L1): First, speed limits are roughly divided into four categories: ultra-low speed, low speed, high speed, and ultra-high speed, which constitutes the L0 level speed limit information. Then, each category is further refined. For example, 10km / h, 15km / h, 20km / h, and 25km / h are classified as ultra-low speed; 30km / h, 40km / h, and 50km / h as low speed; 60km / h, 70km / h, and 80km / h as high speed; and 90km / h, 100km / h, 110km / h, and 120km / h as ultra-high speed, which constitutes the L1 level classification. See Table 2 for details.

[0068] Table 2 Classification of High and Low Speed

[0069]

[0070] (2) Level 0 speed limit recognition

[0071] 70% of the randomly selected samples are used as training samples, and 30% are used as test samples. The training samples are then divided into L0-level classes to form new training samples. Based on road network trajectory features, a random forest classifier is trained to form an L0-level speed limit recognition classifier. Considering the influence of adjacent road segments on the current road segment's speed limit, based on road network trajectory features, the L0-level speed limit recognition classifier is used to perform L0-level speed limit recognition for the current road segment (test set) and its adjacent road segments. Then, based on a voting method, the three results are merged to achieve L0-level speed limit recognition for the test set. See details below. Figure 3 The flowchart for L0-level speed limit recognition is shown. Five-fold cross-validation is used during model training. The maximum feature determination method, impurity, maximum tree depth, minimum number of samples contained in a leaf node, and class weights are set to "None", "gini", "None", 1, and "balanced" respectively. The number of weak classifiers is searched based on a fine-grained grid.

[0072] (3) Level 1 speed limit recognition

[0073] The speed distribution patterns of low-speed road sections differ from those of high-speed road sections. Therefore, this invention proposes a separate modeling method for high-speed and low-speed road sections to optimize speed limit identification based on trajectory features, thereby ensuring the accuracy of L1-level speed limit identification results. Ultra-low and ultra-high speed limits are difficult to identify with fine precision; therefore, these two types of speed limits are only identified to the L0 level. For the low-speed and high-speed road sections classified at L0 level above, a random forest classification model is still used to model low-level and high-level classifiers based on trajectory features, respectively, to optimize L1-level speed limit identification. See details... Figure 4 Flowchart for L1-level speed limit recognition optimization based on trajectory features.

[0074] The road speed limit identification method based on road network features directly trains the model on the labeled L1-level speed limit samples to obtain a speed limit classification model, and then tests it on the test samples. See details... Figure 5 L1 speed limit recognition process based on road network features.

[0075] Considering the influence of adjacent road segments, this invention optimizes the identification results based on the trajectory features of the current road segment and its adjacent segments, as well as the road network feature identification results of the current road segment and its adjacent segments. A voting method is used to fuse these six results, ultimately yielding an L1-level speed limit identification result. This method not only ensures the accuracy of the identification results but also enables hierarchical identification of road speed limits. See the speed limit identification process for details. Figure 6 .

[0076] In practical implementation, the method proposed in this invention can be automated by those skilled in the art using computer software technology. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of this invention and computer equipment including the computer program running the corresponding computer program, should also be within the scope of protection of this invention. This invention also provides a road speed limit hierarchy discovery system that takes into account adjacent road segment information, including the following modules:

[0077] The preprocessing module is used for data preprocessing: first, it performs topology checks on the road network data and quality cleaning on the trajectory data; then, it performs map matching of the road network and trajectory data to connect the trajectory points with their respective road segments.

[0078] The feature design module is used for multimodal feature design: designing road network features based on the road network structure, including road geometric features, road distribution features, and road topological connection features; and designing trajectory features based on road network matching trajectory information, including one-way and two-way road features, road width, road traffic flow, and road operating speed.

[0079] The speed limit level mining module is used for road speed limit mining: combining multimodal features, using a random forest classification model as the basic classifier, and using road speed limit training samples to train the model to obtain the road speed limit classifier; then taking into account the road network and trajectory features of the test sample road segment and its upstream and downstream adjacent road segments, the road speed limit level is identified based on the voting method.

[0080] The specific implementation methods of each module and the corresponding steps are not described in this invention.

[0081] To verify the effectiveness of the method, the study selected the second ring road areas of a certain city and two other cities as experimental areas. For the basic road network, the OSM road network, which has low acquisition cost and high geometric accuracy, was adopted. For trajectory data, crowdsourced trajectory data from July to September 2019 was used for the first city, and crowdsourced trajectory data from November 30, 2018, was used for the second city. The data distribution is as follows: Figure 7 (a), (b), (c), and (d) are shown in the figures. The roads in the selected experimental area are winding and diverse, without a fixed geometric shape, and exhibit a clear radial distribution. In one city, the trajectory data used had high noise and a low sampling frequency, while in two other cities, the trajectory data used had low noise and a high sampling frequency. The selected experimental area and the data used are representative of the analysis and mining of road attribute information using crowdsourced trajectories.

[0082] Compared with previous methods, this method integrates complementary features of the road network and trajectory, considers adjacent road segments, and takes into account the differences between high and low speeds, proposing a road speed limit hierarchy mining method that takes into account information from adjacent road segments. To verify this method, it was compared with commonly used models such as the Nearest Neighbor (KNN), Support Vector Machine (SVM), and Gaussian Naive Bayes (GNB). To ensure the accuracy and objectivity of the experimental results, 10 sets of experiments were conducted for each model, and the average accuracy was used as the benchmark for model evaluation. The Random Forest model performed well overall in the three attribute recognition results, as shown in Table 3. When integrating complementary features of the road network and trajectory and considering adjacent road segments, the accuracy of the KNN and SVM models was higher than that when only considering a single feature and not considering adjacent road segments, which once again demonstrated the advantages of integrating road network and trajectory features and considering adjacent road segments. Taking into account the differences between high and low speeds and performing hierarchical processing also improved the accuracy of road speed limit recognition results, which further verified the necessity of taking into account the differences between high and low speeds. The GNB algorithm is relatively simple, with limited growth potential. It performs worst in road speed limit recognition due to its low accuracy, failing to correctly identify most speed limits, resulting in significant randomness. Integrating road network trajectory features, considering adjacent road segments, and taking into account the differences between high and low speed limits further reduces accuracy. SVM, based on the Radial Basis Function (RBF), has two important parameters: the penalty parameter C for the error term and the kernel coefficient G of the RBF. During training, a fine-grained grid search is performed on C and G. Similarly, when using KNN for training, a fine-grained grid search is also performed on two important parameters: the number of nearest neighbors k and the distance metric p.

[0083] Table 3 Evaluation Results

[0084]

[0085] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for mining road speed limit levels that takes into account information of adjacent road segments, characterized in that, Includes the following steps: Step 1, Data Preprocessing: First, perform topology checks on the road network data and quality cleaning on the trajectory data. Then, perform map matching on the road network and trajectory data to connect trajectory points with their respective road segments. Step 2, Multimodal Feature Design: Design road network features based on the road network structure, including road geometric features, road distribution features, and road topological connection features; Trajectory features are designed based on road network matching trajectory information, including one-way and two-way road characteristics, road width, road traffic flow, and road operating speed. Step 3, Road speed limit mining: Combining the multimodal features in Step 2, a random forest classification model is used as the basic classifier. The road speed limit classifier is obtained by training the model using road speed limit training samples. Then, taking into account the road network and trajectory features of the test sample road segment and its upstream and downstream adjacent road segments, the road speed limit level is identified based on the voting method. The specific implementation method of step 3 is as follows: Step 3.1, Adjacent road segment identification: Identify left and right adjacent road segments based on the road stroke construction method; Step 3.2, Feature preprocessing: Perform missing value processing and feature selection on the designed multimodal features; Step 3.3, Model Training: First, L0-level speed limit recognition is performed based on road network features and trajectory features. Then, L1-level speed limit recognition is optimized by modeling high and low speed limits based on trajectory features. Based on road network features, a random forest model is used to perform L1-level speed limit recognition. Finally, considering the current road segment and the left and right adjacent road segments, the following six results are voted on to obtain the final L1-level speed limit recognition result: L1-level speed limit recognition result of the current road segment based on trajectory features, L1-level speed limit recognition result of the current road segment based on road network features, L1-level speed limit recognition result of the left adjacent road segment based on trajectory features, L1-level speed limit recognition result of the left adjacent road segment based on road network features, L1-level speed limit recognition result of the right adjacent road segment based on trajectory features, and L1-level speed limit recognition result of the right adjacent road segment based on road network features. The specific implementation method of step 3.3 is as follows: (3.3.1) Speed ​​limit L0 and L1 classification: First, the speed limit is roughly divided into four categories: ultra-low speed, low speed, high speed and ultra-high speed, which is the L0 level speed limit information. Then, each category is further refined to obtain the L1 level speed limit information. (3.3.2) L0 speed limit recognition: p% of the samples are randomly selected as the training set and (100-p)% are selected as the test set; the training samples in the training set are divided into L0 levels to form new training samples. Based on the trajectory features, a random forest classifier is trained to form an L0 speed limit recognition classifier; considering the influence of the left and right adjacent road segments on the current road segment speed limit, based on the trajectory features, the L0 speed limit recognition classifier is used to perform L0 speed limit recognition on the current road segment and the left and right adjacent road segments in the test set respectively. Then, based on the voting method, the three results are fused to realize the L0 speed limit recognition of the test set. (3.3.3) L1 speed limit recognition: When recognizing road speed limits based on trajectory features, it is difficult to accurately identify ultra-low speed limits and ultra-high speed limits. Only the above two types of speed limits are recognized to the L0 level. For the low-speed and high-speed road sections after L0 classification, a random forest classification model is used to model low classifiers and high classifiers respectively to carry out L1 speed limit recognition optimization. When identifying road speed limits based on road network features, a random forest classifier is directly used to train the labeled L1-level speed limit samples to obtain a speed limit classification model, and then the test samples are tested. Considering the influence of adjacent road segments on the left and right, the identification results are optimized based on the trajectory characteristics of the current road segment and the adjacent road segments on the left and right, as well as the road network characteristics identification results of the current road segment and the adjacent road segments on the left and right. The voting method is used to fuse the six results and finally obtain the L1 level speed limit identification result.

2. The road speed limit hierarchy mining method considering adjacent road segment information as described in claim 1, characterized in that: The road geometric features in step 2 include road geometric length and road curvature; Road distribution characteristics include the area of ​​adjacent meshes on the road, the average density of n nodes on the road, and the average line density of n-1 road segments on the road; The road topology features include node clustering coefficient, degree centrality, proximity centrality, and betweenness centrality as measured from the perspective of the constituent objects (nodes), and node clustering coefficient, degree centrality, proximity centrality, and betweenness centrality as measured from the perspective of the constituent objects (links).

3. The road speed limit hierarchy mining method considering adjacent road segment information as described in claim 1, characterized in that: Step 2 includes one-way and two-way road features, which include one-way and two-way road information. Road traffic flow characteristics include the number of trajectories matched to the road; Road width features include the mean, mode, standard deviation, median, and maximum of the absolute values ​​of the distances from the road matching trajectory points to the road centerline, and the 5th to 90th percentiles of the absolute values ​​of the distances from the road matching trajectory points to the road centerline; Road speed characteristics include the mean, mode, standard deviation, median, minimum, and maximum speeds of the road matching trajectory points, the 5th to 90th percentiles of the road matching trajectory point speeds, the average instantaneous speeds of the six largest time periods within 24 hours, and the mean, mode, standard deviation, median, minimum, and maximum speed characteristics of the high-speed features.

4. The road speed limit hierarchy mining method considering adjacent road segment information as described in claim 1, characterized in that: The specific implementation method of step 3.1 is as follows: Based on the road stroke construction method, approximate upstream and downstream adjacent road segments are obtained. The road segment connected to the starting point with the smallest deflection angle of the road segment under study and less than a certain angle is identified as the upstream adjacent road segment, and the road segment connected to the corresponding ending point is identified as the downstream adjacent road segment. For the case where the road network is represented by a double line and there are short connecting road segments, the identification of upstream and downstream adjacent road segments is optimized. That is, if the adjacent road segment is identified as a short connecting road segment, the search continues; if the search finds a long connecting road segment, it is identified as an adjacent road segment; otherwise, the current road segment is identified as an adjacent road segment.

5. The road speed limit hierarchy mining method considering adjacent road segment information as described in claim 1, characterized in that: The specific implementation method of step 3.2 is as follows: (3.2.1) Missing value handling: When calculating the relevant features for road attribute identification, the road segment features with missing values ​​based on trajectory data are filled with 0; (3.2.2) Feature selection: First, Spearman correlation coefficient is used to analyze the correlation of design features, then important features are extracted based on this, and finally recursive feature elimination method is used for final feature selection; The formula for calculating the Spearman correlation coefficient is as follows: P =Cov(r X ,r Y ) / s rX s rY formula (1) Where X and Y are the original variables, r X r Y The order of the original variables after sorting them is called the rank variable, Cov(r). X ,r Y ) is the covariance of the rank variable, σ rX ,σ rY It is the standard deviation of the rank variable.

6. The road speed limit hierarchy mining method considering adjacent road segment information as described in claim 1, characterized in that: In the random forest model, the maximum feature determination method, impurity, maximum tree depth, minimum number of samples contained in a leaf node, and class weights are set to None, gini, None, 1, and balanced, respectively, while the number of weak classifiers is searched based on a fine grid.

7. A road speed limit hierarchy discovery system that takes into account information of adjacent road segments, characterized in that, Includes the following modules: The preprocessing module is used for data preprocessing: first, it performs topology checks on the road network data and quality cleaning on the trajectory data; then, it performs map matching of the road network and trajectory data to connect the trajectory points with their respective road segments. The feature design module is used for multimodal feature design: designing road network features based on the road network structure, including road geometric features, road distribution features, and road topological connection features; Trajectory features are designed based on road network matching trajectory information, including one-way and two-way road characteristics, road width, road traffic flow, and road operating speed. The speed limit level mining module is used for road speed limit mining: it combines multimodal features, uses a random forest classification model as the basic classifier, and trains the model using road speed limit training samples to obtain the road speed limit classifier; then, taking into account the road network and trajectory features of the test sample road segment and its upstream and downstream adjacent road segments, it identifies the road speed limit level based on the voting method. The specific implementation method of the speed limit level mining module is as follows: Step 3.1, Adjacent road segment identification: Identify left and right adjacent road segments based on the road stroke construction method; Step 3.2, Feature preprocessing: Perform missing value processing and feature selection on the designed multimodal features; Step 3.3, Model Training: First, L0-level speed limit recognition is performed based on road network features and trajectory features. Then, L1-level speed limit recognition is optimized by modeling high and low speed limits based on trajectory features. Based on road network features, a random forest model is used to perform L1-level speed limit recognition. Finally, considering the current road segment and the left and right adjacent road segments, the following six results are voted on to obtain the final L1-level speed limit recognition result: L1-level speed limit recognition result of the current road segment based on trajectory features, L1-level speed limit recognition result of the current road segment based on road network features, L1-level speed limit recognition result of the left adjacent road segment based on trajectory features, L1-level speed limit recognition result of the left adjacent road segment based on road network features, L1-level speed limit recognition result of the right adjacent road segment based on trajectory features, and L1-level speed limit recognition result of the right adjacent road segment based on road network features. The specific implementation method of step 3.3 is as follows: (3.3.1) Speed ​​limit L0 and L1 classification: First, the speed limit is roughly divided into four categories: ultra-low speed, low speed, high speed and ultra-high speed, which is the L0 level speed limit information. Then, each category is further refined to obtain the L1 level speed limit information. (3.3.2) L0 speed limit recognition: p% of the samples are randomly selected as the training set and (100-p)% are selected as the test set; the training samples in the training set are divided into L0 levels to form new training samples. Based on the trajectory features, a random forest classifier is trained to form an L0 speed limit recognition classifier; considering the influence of the left and right adjacent road segments on the current road segment speed limit, based on the trajectory features, the L0 speed limit recognition classifier is used to perform L0 speed limit recognition on the current road segment and the left and right adjacent road segments in the test set respectively. Then, based on the voting method, the three results are fused to realize the L0 speed limit recognition of the test set. (3.3.3) L1 speed limit recognition: When recognizing road speed limits based on trajectory features, it is difficult to accurately identify ultra-low speed limits and ultra-high speed limits. Only the above two types of speed limits are recognized to the L0 level. For the low-speed and high-speed road sections after L0 classification, a random forest classification model is used to model low classifiers and high classifiers respectively to carry out L1 speed limit recognition optimization. When identifying road speed limits based on road network features, a random forest classifier is directly used to train the labeled L1-level speed limit samples to obtain a speed limit classification model, and then the test samples are tested. Considering the influence of adjacent road segments on the left and right, the identification results are optimized based on the trajectory characteristics of the current road segment and the adjacent road segments on the left and right, as well as the road network characteristics identification results of the current road segment and the adjacent road segments on the left and right. The voting method is used to fuse the six results and finally obtain the L1 level speed limit identification result.

Images