Bicycle road intersection recognition method based on neighborhood region behavior difference

Abstract

The application discloses a kind of based on near neighbor area behavior difference's cycling road intersection identification method, first utilize quadtree model to divide detection area into different size units, extract the behavior characteristics in each grid unit by cycling trajectory data, according to the behavior characteristic difference of each grid and its near neighbor grid, the correlation characteristics of grid and near neighbor grid are calculated, subsequently using XGBoost classifier identifies the unit containing cycling road intersection, for all turning trajectory points in the unit containing cycling road intersection, the center position of cycling road intersection is identified using adaptive bandwidth Mean-shift algorithm based on flow perception, finally using annular trajectory point density variation algorithm determines the coverage of cycling road intersection.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of trajectory mining, and particularly relates to a riding road intersection identification method based on behavior difference in adjacent regions. BACKGROUND

[0002] With the large-scale increase in the number of non-motor vehicles, non-motor vehicle-related location services have entered a period of rapid development, and residents' daily travel increasingly considers non-motor vehicles as a means of transportation, thereby facilitating and expediting arrival at a destination. In actual scenarios, riders usually choose some narrow roads in parks, squares and other places to save travel time, and the roads they converge form small-scale road intersections that are adjacent to each other and complex. Due to the lack of professional and accurate non-motor vehicle riding navigation maps, riders are prone to getting lost or even driving into the wrong route when passing through these intersections, which affects the riding experience. As an important part of the riding map, the riding road intersection is of great significance to the riding map navigation service. In the real-time navigation scenario, the position and range of the riding road intersection need to be determined in advance for turning announcement, and the lack of and inaccurate identification of the riding road intersection will affect the rationality of path planning.

[0003] In recent years, a large number of documents have been devoted to the detection of intersection recognition in motor vehicle road networks. Currently, the algorithm for generating road intersections based on vehicle trajectories can be divided into two basic types. The first type of algorithm simply regards the map nodes generated by the map construction algorithm as intersections. This method will misidentify road endpoints as intersections and cannot determine the range of the intersection, which is not suitable for navigation scenarios. The second type of algorithm locates the intersection area according to the multi-directional and low-speed behavior characteristics of trajectories at the intersection, and then clusters the trajectory points with turning behaviors to determine the intersection position. However, in the riding scenario, it is difficult to identify small-scale intersections that are adjacent to each other and have sparse passing trajectories by only considering the above behavior characteristics. SUMMARY

[0004] The technical problem to be solved by the present application is that the small-scale intersection in the riding scenario is difficult to detect, and the existing intersection detection method based on motor vehicle trajectory data cannot be directly used for riding intersection detection. A riding road intersection identification method based on behavior difference in adjacent regions is provided, which can effectively and accurately locate the riding road intersection using riding trajectory data, and provide effective decision support for navigation services and path planning.

[0005] To achieve the above technical purposes, the technical solution adopted by the present application is as follows:

[0006] A riding road intersection identification method based on behavior difference in adjacent regions, comprising the following steps:

[0007] S1: using a quadtree-based unit division method to divide the detection area into various size grids.

[0008] S2: based on the riding trajectory data, extracting the behavior characteristics in each grid unit, and calculating the correlation characteristics of the grid and the adjacent grid according to the difference in the behavior characteristics of the grid and the adjacent grid.

[0009] S3: using an XGBoost classification model to obtain the grid with a riding road intersection.

[0010] S4: based on the grid with an intersection obtained in S3, screening out the riding turning point with a turning behavior.

[0011] S5: using a flow-aware adaptive bandwidth Mean-Shift clustering algorithm to determine the center position of the intersection.

[0012] S6: based on the intersection center position obtained in S5, determining the coverage range of the intersection by a ring turning point density change detection algorithm.

[0013] To optimize the above technical solutions, the specific measures taken also include:

[0014] The above step S1 specifically includes:

[0015] Considering that the scale of the intersection accessible to the riding driver is different, the maximum diameter is usually not more than 100m, and the minimum diameter is usually only 6-10m, the observation area is divided into various size grids by using a quadtree-based unit division method, the minimum side length of the quadtree is set to 12.5m, and the road intersection grid is searched by using a hierarchical traversal from 100m (i.e., from the fourth layer of the quadtree).

[0016] The above step S2 specifically includes:

[0017] Some small-scale riding intersections are adjacent to each other, and only a small amount of trajectories pass through in a unit time. It is difficult to detect a grid containing only a small-scale intersection by simply identifying the road intersection grid according to the average speed, direction distribution and other behavior characteristics of the trajectory points in the grid. Therefore, when detecting the road intersection grid, not only the turning distribution and average speed of the trajectory points in the grid, but also the difference between the current grid and its adjacent grid in these characteristics are considered. Specifically, two types of features are extracted:

[0018] Behavior characteristics: considering the multi-direction and low-speed characteristics of the trajectory in the intersection area, the riding behavior characteristics of the driver in each grid are extracted according to the riding trajectory, including: (1) turning frequency f tp: represents the number of turning points in the grid, and a turning point is defined as: if the orientation of the current trajectory point is greater than 35 degrees from the orientation of the previous trajectory point, then the current trajectory point is determined to be a turning point.(2) Direction distribution : represents the direction distribution of the trajectory points in the grid: the orientation of each trajectory point is mapped into 8 normalized and equal-size direction histograms (0°-45°, 45°-90°, 90°-135°, 135°-180°, 180°-225°, 225°-270°, 270°-315°, 315°-360°) to obtain the eight-direction distribution corresponding to the grid.(3) Average speed f sp : represents the average speed of the trajectory points in the grid, and the speed of the current trajectory point is calculated by the ratio of the time difference and the distance between the current trajectory point and the previous trajectory point.

[0019] Contextual features: considering the sparsity of the riding trajectory in the small intersection area, the contextual features corresponding to the grid are extracted according to the difference in behavior characteristics between the grid and its adjacent grid, wherein the adjacent grid represents the grid directly adjacent to the current grid in the east, west, south and north directions, and the contextual features specifically include: (1) Speed difference f sd : considering that the driving speed of the riding driver at the intersection is usually lower than the driving speed on the road, f is represented by calculating the ratio of the average speed of the current grid to the average speed of all trajectory points in its adjacent grid. sd (2) Point density difference f pd : considering that the number of trajectories in the intersection in a unit time is greater than the number of trajectories on the associated road segment, f is represented by calculating the ratio of the number of trajectory points in the current grid to the number of trajectory points in its adjacent grid. pd (3) Turning density difference f td : considering that the turning behavior occurs more frequently at the intersection than on the road segment, f is represented by calculating the ratio of the number of turning points in the current grid to the number of turning points in its adjacent grid. td .

[0020] The above step S3 specifically includes:

[0021] Splicing the grid behavior features and contextual features obtained based on S2 to obtain a 13-dimensional feature vector of the grid By sequentially traversing the quadtree, an XGBoost classifier is used for each layer of the quadtree to identify the grids containing intersections in the layer.

[0022] The above step S4 specifically includes:

[0023] Based on the grids containing intersections obtained in S3, all turning points contained in these grids are extracted according to the definition of the turning points.

[0024] The step S5 specifically includes:

[0025] Based on the turning points obtained in S4, the Mean-Shift clustering algorithm is used to cluster the turning points to obtain the center of the intersection.

[0026] Mean-shift is a non-parametric point clustering method, which aims to locate the extreme points of the clustering point density function, and iteratively calculates the centroid of the clustering point until convergence through the following formula.

[0027] c i+1 =c i +M(c i ),

[0028] Where c i represents the center position of the clustering cluster at the i-th iteration, M(c i ) represents the drift vector that helps c i locate the density maximum, and its specific value is determined by the bandwidth b of the standard Gaussian kernel function K(·), which represents the search distance of the adjacent trajectory points p j of the clustering cluster center position c i , and x represents the independent variable of the Gaussian kernel function.

[0029] The bandwidth b, as the only parameter of the algorithm, is crucial to the clustering result. Considering the different sizes of the intersection, it is unreasonable to simply set the bandwidth b as a constant. For example, if the value of b is set too large, the algorithm will mistakenly merge multiple small cycling intersections into one intersection, and if the value of b is set too small, the algorithm will identify a large-scale intersection as multiple small-scale intersections. Considering that the size of the cycling intersection is positively linearly correlated with the number of trajectory points within the intersection per unit time (referred to as traffic flow), the bandwidth b is defined as a variable that changes adaptively within a certain range of the clustering centroid c i , and is defined as follows:

[0030] σ=max({d(p i )|p i ∈P}), d(c i )=∑ t∈P I({t|dis(t,c i )<δ})

[0031] Where u and v represent the minimum and maximum intersection diameters (set to 10m and 80m respectively in the present application) to ensure that the trajectory points within any size intersection will participate in the weight calculation of the kernel function, and d(c i ) represents ci The number of trajectory points in the range delta (delta is set to 80m in the application, which is the same as v), sigma represents the maximum value of the number of adjacent trajectory points in all trajectory points, to ensure that the finally calculated clustering bandwidth b is in the [u, v] interval; P represents the set of all trajectory points, and I represents an indication function, which is equal to 1 when the condition is met, and 0 otherwise.

[0032] The cluster center position after the convergence of the clustering algorithm is regarded as the center position of the cycling intersection.

[0033] The above step S6 specifically includes:

[0034] Based on the cycling intersection center position obtained in S5, as the intersection range radius increases, the number of turning points in the growing ring will gradually decrease, and when the number of turning points is less than a given threshold, the road intersection range is obtained. The radius increment (denoted as r delta ) is set to 5 meters. The minimum radius (denoted as r min = 10) is set to avoid extreme cases, and the maximum radius (denoted as r max = 80) is set to ensure that the coverage radius does not overlap with adjacent road intersections.

[0035] Based on the above method, the application also provides a cycling road intersection identification system based on the behavior difference of the adjacent region, comprising a memory and a processor; the memory stores a computer program, and when the computer program is executed by the processor, the above-mentioned method is realized.

[0036] The application also provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to realize the above-mentioned method of the application.

[0037] The application has the following beneficial effects:

[0038] 1. Based on cycling trajectory data, this study first combines the multi-feature adjacent behavior difference of the region to accurately detect the grid containing the road intersection.

[0039] 2. Considering the traffic difference of different scale cycling intersections, this study designs a Mean-shift clustering algorithm based on adaptive bandwidth, which can accurately identify cycling intersections of different scales.

[0040] 3. This study first proposes a center position and range detection method for cycling intersections. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 A cycling road intersection identification method based on the behavior difference of the adjacent region.

[0042] Figure 2A road intersection detection effect diagram after the CITT intersection detection method in the embodiment of the present application.

[0043] Figure 3 A road intersection detection effect diagram after the intersection detection method of the present application in the embodiment of the present application. DETAILED DESCRIPTION

[0044] The application will be further described in conjunction with the following specific embodiments and drawings. The process, conditions, experimental methods, etc. for implementing the present application are the general knowledge and common sense in the art, and the present application does not have special restrictions.

[0045] The application discloses a kind of based on near neighbor area behavior difference's cycling road intersection identification method, first using quadtree model in the present application Detection area is divided into different size units, the behavior characteristics in each grid unit are extracted by cycling trajectory data, the correlation characteristics of grid and near neighbor grid are calculated according to the behavior characteristic difference of each grid and near neighbor grid, subsequently using XGBoost classifier identifies the unit containing cycling road intersection, the center position of cycling road intersection is identified using adaptive bandwidth Mean-shift algorithm based on flow perception to all turning trajectory points in the unit containing cycling road intersection, finally using ring trajectory point density variation algorithm determines the coverage of cycling road intersection.The application discloses a kind of based on near neighbor area behavior difference's cycling road intersection identification method, as shown in figure Figure 1 The application discloses a kind of based on near neighbor area behavior difference's cycling road intersection identification method, as shown in figure

[0046] As shown in figure Figure 1 The application discloses a kind of based on near neighbor area behavior difference's cycling road intersection identification method, as shown in figure

[0047] S1: using the unit division method based on quadtree divides detection area into various size grids.

[0048] In the embodiment, step S1 specifically includes:

[0049] Considering that the sizes of the crossroads accessible to the riding driver are various, the maximum diameter is usually not more than 100 m, and the minimum diameter is usually only 6-10 m, the observation area is segmented into various size grids by using a unit division method based on a quadtree, the minimum side length of the quadtree is set to 12.5 m, and the grid containing the road intersection is searched by using a hierarchical traversal starting from 100 m (that is, starting from the fourth layer of the quadtree).

[0050] S2: extracting the behavior features in each grid unit based on the riding trajectory data, and calculating the correlation features of the grid and the neighboring grid according to the difference in the behavior features of the grid and the neighboring grid.

[0051] In the embodiment, step S2 specifically includes:

[0052] Some small-scale riding crossroads are adjacent to each other, and only a small amount of trajectories pass through in a unit time. It is difficult to detect the grid containing only a small-scale crossroad by simply identifying the road intersection grid according to the average speed, direction distribution and other behavior features of the trajectory points in the grid. Therefore, when detecting the road intersection grid, not only the behavior features such as the turning distribution and average speed of the trajectory points in the grid are considered, but also the difference in these features between the current grid and its neighboring grid is considered. Specifically, two types of features are extracted:

[0053] 1. Behavior features: considering the multi-direction and low-speed characteristics of the trajectory in the intersection area, the riding behavior features of the driver in each grid are extracted according to the riding trajectory, specifically including: (1) turning frequency f tp : representing the number of turning points in the grid, the definition of the turning point is that if the direction of the current trajectory point is greater than 35 degrees from the direction of the previous trajectory point, the current trajectory point is determined as a turning point. (2) Direction distribution : representing the direction distribution of the trajectory points in the grid: the direction of each trajectory point is mapped into eight standardized and equal-size direction histograms (0°-45°, 45°-90°, 90°-135°, 135°-180°, 180°-225°, 225°-270°, 270°-315°, 315°-360°) to obtain the eight-direction distribution corresponding to the grid. (3) Average speed f sp : representing the average speed of the trajectory points in the grid, the speed of the current trajectory point is calculated by the ratio of the time difference and the distance between the current trajectory point and the previous trajectory point.

[0054] 2. Context features: considering the sparsity of the riding trajectory in the small intersection area, the context features corresponding to the grid are extracted according to the difference in the behavior features between the grid and its neighboring grid, wherein the neighboring grid represents the grid directly adjacent to the current grid in the east, west, south and north directions, and the context features specifically include: (1) speed difference f sdConsidering that the driving speed of the cyclist at the intersection is usually lower than the driving speed on the road, f is represented by calculating the ratio of the average speed of the current grid to the average speed of all trajectory points in its neighboring grids sd (2) Point density difference f pd Considering that the number of trajectories in the intersection is greater than the number of trajectories on its associated road segment in a unit of time, f is represented by calculating the ratio of the number of trajectory points of the current grid to the number of trajectory points in its neighboring grids pd (3) Turning density difference f td Considering that the turning behavior occurs more frequently at the intersection than on the road segment, f is represented by calculating the ratio of the number of turning points of the current grid to the number of turning points in its neighboring grids td .

[0055] S3: Obtain the grid where the intersection of the cycling road exists using the XGBoost classification model.

[0056] In the embodiment, step S3 specifically includes:

[0057] Concatenate the grid behavior features and context features obtained based on S2 to obtain a 13-dimensional feature vector of the grid By sequentially traversing the quadtree, an XGBoost classifier is used for each layer of the quadtree to identify the grid containing the intersection in the layer.

[0058] S4: Filter out the turning points of the cycling turning behavior based on the grid containing the intersection obtained in S3.

[0059] In the embodiment, step S4 specifically includes:

[0060] Based on the grid containing the intersection obtained in S3, all turning points contained in these grids are extracted according to the definition of the turning point.

[0061] S5: Cluster the turning points using the adaptive bandwidth Mean-Shift clustering algorithm based on traffic perception to determine the center position of the intersection.

[0062] In the embodiment, step S5 specifically includes:

[0063] Based on the turning points obtained in S4, the Mean-Shift clustering algorithm is used to cluster them to generate a cluster center representing the center position of the road intersection.

[0064] Mean-Shift is a non-parametric point clustering method, which aims to locate the extreme points of the clustering point density function, and iteratively calculates the centroid of the clustering point until convergence through the following formula.

[0065]

[0066] Where c i M(c) represents the center position of the cluster in the i-th iteration. i ) indicates help c i The drift vector for locating the density maxima is specifically determined by the bandwidth b of the standard Gaussian kernel function K(·), where bandwidth b represents the cluster center position c. i Searching for neighboring trajectory points p j The search distance.

[0067] Bandwidth *b*, as the sole parameter of this algorithm, is crucial to the clustering results. Considering the varying sizes of cycling intersections, simply setting *b* to a constant is unreasonable. For example, setting *b* too large might cause the algorithm to mistakenly merge multiple small cycling intersections into a single intersection, while setting it too small might cause the algorithm to identify a large intersection as multiple smaller intersections. Given that the size of a cycling intersection is positively linearly correlated with the number of trajectory points within the intersection per unit time (referred to as traffic flow), bandwidth *b* is defined as a value that varies with the cluster centroid *c* during the clustering process. i The variable whose number of trajectory points changes adaptively within a certain range is defined as follows:

[0068]

[0069] Where u and v represent the minimum and maximum intersection diameters (set to 10m and 80m respectively in this invention) to ensure that trajectory points within intersections of any size participate in the weight calculation of the kernel function, d(c i ) represents c i The number of trajectory points within the range δ (in this invention, δ is set to 80m, the same as v), σ represents the maximum number of neighboring trajectory points among all trajectory points, to ensure that the final calculated clustering bandwidth b is within the interval [u, v].

[0070] The center of the cluster after the clustering algorithm converges is regarded as the center of the cycling intersection.

[0071] S6: Based on the intersection center position obtained from S5, the intersection coverage area is determined by the circular turning point density change detection algorithm.

[0072] In this embodiment, step S6 specifically includes:

[0073] Based on the center position of the cycling intersection obtained from S5, as the intersection radius increases, the number of turning points within the increasing loop gradually decreases. The intersection radius is determined when the number of turning points falls below a given threshold. The radius increment (denoted as r) is then calculated. delta Set the radius to 5 meters. Set the minimum radius (denoted as r).min =10) to avoid extreme cases, maximum radius (denoted as r) max =80) Ensure that the coverage radius does not overlap with adjacent road intersections.

[0074] Finally, the center position of the intersection and its corresponding range contained in the area to be detected are obtained and output.

[0075] To verify the effectiveness of this invention, real trajectory data was selected for intersection detection on cycling routes. Based on a publicly available cycling road network, the results were compared and evaluated with existing intersection detection methods. The selected methods included Huang, CITT, Wang, and Qing, such as... Figure Two The image shown is a diagram illustrating the intersection recognition performance of CITT. Figure Three The image shows the intersection recognition effect based on the intersection detection method of this invention. Experiments show that this invention can accurately identify large-scale intersections marked by rectangular boxes and small-scale intersections marked by elliptical boxes that are nearest neighbors. The evaluation metrics for intersection recognition—Precision, Recall, and F1-score—are selected for quantitative evaluation. The true location is obtained from OSM road network data. truth L represents the actual number of intersections. detect L represents the number of intersections detected. corr This represents the number of correctly identified intersections, and d represents the matching distance between the detected intersection and the baseline intersection. A higher F1-score indicates better performance. Precision, recall, and F1-score are defined as follows:

[0076]

[0077]

[0078]

[0079] The comparative evaluation results of all intersections are shown in Table 2. Using this invention to identify cycling intersections improves precision, recall, and F1-score to a certain extent. Considering that small-scale intersections only involve single-vehicle road segments with a length of no more than 5m, this invention considers intersections with all associated road segment widths less than 5m as small-scale intersections. Table 2 shows the comparative evaluation results of small-scale intersections. Using this invention to identify small-scale cycling intersections significantly improves recall.

[0080] Table 1: Comparison of Intersection Detection Results

[0081]

[0082] Table two small-scale intersection detection effect comparison table

[0083] Methods Recall Huang 0.4583 CITT 0.5385 Wang 0.3854 Qing 0.4271 Ours 0.7802

[0084] The application comprehensively adopts technologies such as a quadtree model, an XGBoost classifier and a Mean-Shift clustering technology, forms a cycling road intersection recognition method based on a near neighbor region behavior difference, can accurately locate a cycling road intersection center position and a coverage range through trajectory data mining, and thus provides effective help for navigation broadcasting and path planning.

[0085] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for ordinary skilled persons in the art, some improvements without departing from the principle of the present application shall be considered as the protection scope of the present application.

Claims

1.A method for identifying a road intersection for cycling based on differences in behaviors of nearby areas, characterized by, The method comprises the following steps: S1: using a quadtree-based cell division method to divide the detection area into various size grids; S2: extracting behavior features in each grid unit based on the riding trajectory data, and calculating the correlation features between the grid and its neighbor grids according to the difference of the behavior features of the grid and the neighbor grids; the behavior features include (1) turning frequency , (2) direction distribution , (3) average speed ; S3: using an XGBoost classification model to obtain grids with cycling road intersections; S4: based on the grids with intersections obtained in S3, screening out cycling turning points with turning behaviors; S5: using a flow-aware adaptive bandwidth Mean-Shift clustering algorithm to determine the intersection center position; the step S5 specifically comprises: Based on the turning points obtained in S4, the Mean-Shift clustering algorithm is used to cluster them to generate a cluster center representing the road intersection center position; Mean-Shift is a non-parametric point clustering method, which aims to locate the extreme points of the clustering point density function, and iteratively calculates the cluster point centroid until convergence through the following formula: , , ， in, Indicates the first The center position of the cluster in the next iteration To show help The drift vector that locates the density maxima is specifically determined by the standard Gaussian kernel function. bandwidth Decision, bandwidth Indicates the location of the cluster center Searching for adjacent trajectory points The search distance, where x represents the independent variable of the Gaussian kernel function; bandwidth. The bandwidth is defined as the only parameter of this algorithm. To follow the cluster centroid during the clustering process The variable that adaptively changes the number of trajectory points within the range is defined as follows: ， ， ， wherein, and denote the minimum and maximum intersection diameters to ensure that trajectory points within any size intersection will participate in the weight computation of the kernel function, denote the number of trajectory points within the range , denote the maximum number of neighboring trajectory points among all trajectory points to ensure that the final computed cluster bandwidth is within the interval; P denotes the set of all trajectory points, and I denotes the indicator function which equals 1 when the condition is met and 0 otherwise. The cluster center position after the convergence of the clustering algorithm is regarded as the center position of the cycling intersection; S6: based on the intersection center position obtained in S5, the intersection coverage range is determined through a ring turning point density change detection algorithm. 2.The intersection recognition method based on the difference in the behavior of the adjacent area according to claim 1, wherein, The step S1 specifically comprises: The observation area is divided into various size grids using a quadtree-based cell division method, the minimum edge length of the quadtree is set to 12.5m, and the layer sequence traversal search is started from 100m to include the road intersection grid. 3.The intersection recognition method based on the difference of behaviors in the adjacent area of claim 1, wherein, The step S2 specifically comprises: When detecting the road intersection grid, not only the turning distribution and average speed behavior characteristics of the trajectory points in the grid are considered, but also the differences in these characteristics between the current grid and its adjacent grids are considered; two types of features are extracted: Behavioral features: According to the riding trajectory, the driver's riding behavior characteristics in each grid are extracted, including: (1) turning frequency : represents the number of turning points in the grid, and the turning point is defined as: if the orientation of the current trajectory point is greater than 35 degrees from the orientation of the previous trajectory point, then the trajectory point is determined to be a turning point; (2) direction distribution : represents the direction distribution of the trajectory points in the grid: the orientation of each trajectory point is mapped into 8 standardized equal-size direction histograms to obtain the eight-direction distribution corresponding to the grid; (3) average speed : represents the average speed of the trajectory points in the grid, and the speed of the current trajectory point is calculated by the ratio of the time difference and the distance between it and its previous trajectory point; Contextual features: Extract the contextual features corresponding to the grid according to the difference of the grid and its adjacent grid in the behavior characteristics, wherein the adjacent grid represents the grid directly adjacent to the current grid in the east, west, south and north directions, and the contextual features include: (1) Velocity difference : Represented by calculating the ratio of the average speed of the current grid to the average speed of all trajectory points in its adjacent grid ; (2) Point density difference : Represented by calculating the ratio of the number of trajectory points in the current grid to the number of trajectory points in its adjacent grid ; (3) Steering density difference : Represented by calculating the ratio of the number of steering points in the current grid to the number of steering points in its adjacent grid . 4.The intersection recognition method based on the difference of behaviors in the adjacent area according to claim 3, wherein, The step S3 specifically comprises: Splice the grid behavior features and context features based on S2 to obtain a 13-dimensional feature vector of the grid By layer sequence traversal of the quadtree, an XGBoost classifier is used for each layer of the quadtree to identify the grid containing the intersection in the layer. 5.The intersection recognition method based on the difference of behaviors in the adjacent area of a cycling road according to claim 1, wherein, The step S4 specifically comprises: Based on the intersection-containing grids obtained in S3, all turning points contained in these grids are extracted according to the definition of the turning point. 6.The intersection recognition method based on the difference of behaviors in the adjacent area of a cycling road according to claim 1, wherein, The step S6 specifically comprises: Based on the intersection center position of the intersection obtained in S5, as the intersection range radius increases, the number of turning points in the growing ring will gradually decrease, and when the number of turning points is less than a given threshold, the road intersection range is obtained; the radius increment is set to 5 meters; the minimum radius =10 is set to avoid extreme cases, and the maximum radius =80 is set to ensure that the coverage radius does not overlap with adjacent road intersections; finally, the intersection center position and its corresponding range contained in the to-be-detected region are obtained and output.

Images