Road segment trajectory-based lane geometry information extraction method, system, device and medium

Abstract

The application discloses a lane geometry information extraction method based on a road section track, performs track cluster analysis on a track subset passing through the same road, obtains a track cluster belonging to the same lane, extracts a track geometry center of the track subset in the same cluster in the form of a non-parametric curve, and fits to obtain a candidate lane center line; a lane change lane line identification method based on maximum group search, an intersection relationship graph is established, the lane change lane line is identified and removed through maximum group analysis, a set of lane center lines which do not intersect with each other is obtained as basic lane center lines, the number of lanes at both ends of the road section is reviewed through a Gaussian mixture model, and the candidate lane center line is extracted as a newly-added lane center line based on the newly-added Gaussian component and parameters. The application effectively extracts accurate lane geometry information based on track data, the acquisition method is simple and effective, easy to implement, has important significance for lane-level mapping application, and provides bottom layer data support for automatic driving application.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent traffic information technology, and particularly relates to a lane geometry information extraction method and system based on road section trajectories. BACKGROUND

[0002] High-precision maps provide important underlying data support for autonomous driving. At present, the road mapping technology providing data basis for high-precision maps gradually changes from previous manual measurement to semi-automated production based on various sensors. The current road mapping technology mainly concentrates on collecting roads based on satellite remote sensing images, unmanned aerial vehicle orthographic images and professional mobile vehicle collection systems. Although these professional equipment centralized collection methods have high precision, they have the disadvantages of low update frequency and high collection cost.

[0003] Since crowd-sourced trajectory data is derived from travel vehicles equipped with GPS receiving modules, it has the advantages of wide coverage and high timeliness in urban road areas, providing a high time-freshness and low-cost production method for road mapping. Existing lane geometry information extraction methods based on crowd-sourced trajectory data can be divided into lane geometry information extraction methods based on local distribution of trajectory points and lane estimation methods based on trajectory clustering algorithms according to the processing objects.

[0004] Publication No. CN202010140103 discloses lane geometry information extraction based on local distribution of trajectory points: when a vehicle travels along a lane centerline, the trajectory points calculated by the vehicle-mounted GPS receiver module follow a Gaussian distribution perpendicular to the lane centerline, with the mean center being the lane centerline position. Therefore, on the road cross section, the distribution of trajectory points can be regarded as a Gaussian mixture distribution composed of multiple Gaussian components, and the solution of lane geometry information can be regarded as the solution of parameters in the Gaussian mixture model.

[0005] The specific algorithm process is as follows: a sliding window is used to select trajectory point data in the window and project it to the vertical lane direction; a known lane width is introduced as a known parameter, the number of lanes is iterated, the lane center position, weight, etc. are solved by fitting a Gaussian mixture model with trajectory point distribution; the confidence of the solution under different lane conditions is evaluated by a risk assessment model; and the number and position of the lanes are determined according to the confidence.

[0006] Publication No. CN202111043369, Lane Estimation Method Based on Trajectory Clustering Algorithm: Compared with the lane geometry information extraction method only for local GPS point distribution, this method uses complete trajectory or sub-trajectory as the input of clustering, emphasizing the order of trajectory observation points and the connection relationship of trajectory. By trajectory clustering, the trajectory cluster passing through the same lane is extracted, and each group of clustered trajectory cluster is used for lane center line estimation. The initial clustering center line is determined according to the distribution state of the target trajectory data; the target trajectory data is clustered based on the initial clustering center line to obtain the target clustering center line after clustering; and the target lane to which the vehicle to be detected belongs is determined based on the target clustering center line. In this way, the distribution state of the collected preset number of target trajectory data can be used to quickly determine the initial clustering center line; the target trajectory data is clustered based on the initial clustering center line to obtain an accurate target clustering center line, and then the target lane to which the vehicle to be detected belongs is quickly determined according to the target clustering center line.

[0007] Trajectory data can reflect the details of urban road network, such as the position, width and driving direction of the lane. In urban road design, the entrance and exit lanes of the flat intersection, the harbor-type parking station and the distribution lanes of the high-speed expressway are typical real lane number change scenarios. The existing lane geometry information extraction method based on trajectory data, such as K-Means clustering, kernel density estimation and Gaussian mixture model, essentially models the distribution characteristics of trajectory points within a certain range to obtain the number and position of the lane. However, this method relies on the statistical characteristics of trajectory data, and in the process of identifying the number of lanes and the position of the lane center, the Gaussian mixture model is sensitive to the distribution of trajectory data, the parameters often need to be adjusted manually based on prior information, and the accuracy is not high. It cannot extract the lane geometry and topological connectivity information in the common diverging and merging area of urban road network, obtain the lane change situation and connection relationship in the actual road network. The lane estimation method based on trajectory clustering algorithm relies on trajectory data, and when there are lane-changing trajectories generated by driving behavior in the trajectory data, the lane generated from such trajectories will be inconsistent with the actual lane in terms of position and number.

[0008] Urban road network often has a more complex lane-level topological structure, such as complex lane number change scenarios. It is still a very difficult task to accurately represent the complete geometry information and topological structure of the lane network under the variable lane scenario based on trajectory data. Therefore, it is difficult to use the above method to construct a truly complete structured lane geometry information for urban road network. SUMMARY

[0009] The present application aims at the above-mentioned problems existing in the prior art, and provides a method and system for extracting fine lane geometry information based on trajectory data. Through trajectory clustering, candidate lane lines are fitted for the same lane trajectory, and complex traffic scenarios where the number of lanes changes are identified and processed, so as to extract basic lane lines and newly added lane lines, and realize complete fine lane center line geometry information extraction.

[0010] Therefore, based on one aspect of the present application, a lane geometry information extraction method based on trajectories is provided. A non-parametric curve fitting method based on iterative optimization is used to perform trajectory clustering analysis on a trajectory subset passing through the same road, so as to obtain a trajectory cluster belonging to the same lane. The trajectory geometry center of the trajectory subset in the same cluster is extracted in the form of a non-parametric curve, and a candidate lane center line is fitted. The lane center line is regarded as a node in a graph, and an intersection relationship graph between lane lines is established. The maximum group is used to identify a lane change line, and a set of non-intersecting lane center lines is obtained as basic lane lines. The candidate lane center line is extracted as a newly added lane center line based on the newly added Gaussian components at both ends of the road section, and the lane change connection position is determined according to the basic lane line and the newly added lane center line, so as to construct a lane geometry information and topological relationship graph.

[0011] Further preferably, the basic lane line is obtained by regarding the lane line as a node in a graph, recording the geometric characteristics of the lane line as the weight of the node, recording the intersection relationship characteristics between the lane lines as the weight of the edge, constructing a candidate lane line intersection relationship graph by using the fitted lane center lines of the road sections, defining each candidate lane line as a node of the graph, and representing the intersection relationship between the nodes by the edges. If two lane lines do not intersect, the edges are connected, so as to construct an undirected graph, and a set of non-intersecting lane center lines is extracted as the basic lane line through the undirected graph.

[0012] Further preferably, the lane center line fitting of the trajectory cluster of the same lane further includes: according to the time sequence of the trajectory observation points in the trajectory cluster of the same lane, a non-parametric curve fitting method based on iterative optimization is used to fit the lane center line.

[0013] Further preferably, the lane center line fitting includes: a trajectory line is randomly selected as an initial lane line, the cross section line of each point of the lane line is calculated; the geometric center of the intersection point of the cross section line and the input trajectory cluster is calculated as the center point of the lane, and each trajectory cluster intersection point is connected as the fitted lane center line.

[0014] Further preferably, in the road section where the number of lanes changes, the number of lanes at the increasing end is taken as the total number of lanes of the road section, the basic lane line and the Gaussian component at the increasing end of the number of lanes are corresponded, and the newly added Gaussian component is calculated to determine and extract the newly added lane line.

[0015] Further preferably, the clustering of the trajectories in the lane comprises: initializing each trajectory segment as a cluster, and calculating the Hausdorff distance between each pair of trajectories, merging the clusters with the Hausdorff distance less than a distance threshold into a trajectory cluster of the same lane, calculating the inter-cluster distance between all trajectory clusters, merging the clusters with the inter-cluster distance less than a distance threshold into a trajectory cluster of the same lane, until all inter-cluster distances are greater than the distance threshold, and completing the lane trajectory clustering.

[0016] Further preferably, the clustering of the trajectories in the lane comprises: obtaining the trajectory clusters belonging to each road segment based on a density-based clustering algorithm, and collecting all trajectories belonging to the same road segment as a basic unit of lane extraction.

[0017] Further preferably, the maximum clique algorithm based on a graph is used to obtain the basic lane lines, a candidate lane line intersection relationship graph is constructed using the lane center lines fitted for each road segment, each candidate lane line is defined as a node of the graph, the edges between the nodes represent the intersection relationship between the lanes, if two lane lines do not intersect, the edges are connected, a non-directed graph is constructed, and a set of lane center lines that do not intersect each other is extracted from the non-directed graph as the basic lane lines.

[0018] Further preferably, each candidate lane line is defined as a vertex set V = {1,..., n} of the graph, and the edge set E between the nodes represents the intersection relationship between the lanes, if two lane lines do not intersect, the edges between the nodes represented by the two lane lines are connected, and G = (V, E) is constructed as a non-directed graph with the vertex set V = {1,..., n} and the edge set If a subset U ∈ V and for any two vertices u, v ∈ U, (u, v) ∈ E, then U is a complete subgraph of G, and the maximum complete subgraph is the maximum clique of the graph.

[0019] Further preferably, when there are multiple maximum cliques, the set with the maximum sum of the weights of the lane line nodes is selected as the basic lane line of the non-changing lane, and the node weight weight is calculated according to the formula:

[0020]

[0021] The node weight weight is calculated according to the formula:

[0022] Further preferably, the prior Gaussian mixture model used to identify the number of lanes at both ends of the road section comprises: identifying the main direction of the road and the road cross-section direction perpendicular to the main direction through the motion direction of the trajectories within the window, projecting all GPS trajectory points in the road cross-section direction to obtain the projection distribution of the data in the road cross-section direction, and obtaining the position of each lane center line as the peak value μ of the lane Gaussian component j , the distribution of the trajectory observation points on each lane corresponds to the standard deviation σ of each Gaussian component data distribution j , the weight of each Gaussian component distribution is determined according to the proportion of the number of trajectories in each lane, and the probability density function of the Gaussian mixture model is constructed:

[0023]

[0024] In the formula, k is the number of Gaussian components, j = 1, 2, …, k, w j is the weight of the jth Gaussian component distribution, wherein w1+w2+…+w k =1.

[0025] Further preferably, the evaluation function for optimizing the selection of the number of lanes is designed as

[0026] The fitting of the Gaussian mixture model is evaluated, overfitting is prevented, the lane width value is constrained, the distance between the peak value distributions of each Gaussian component is controlled, and the penalty term for preventing overfitting as the number of mixed components increases is taken as the likelihood function, and the Gaussian mixture model with the minimum evaluation function value is selected as the optimal model, wherein λ is a regularization parameter, and Δh i represents the average distance between the center lines of adjacent lanes.

[0027] According to another aspect of the present application, a trajectory-based lane geometry information extraction system is provided, comprising: a lane trajectory clustering module, a lane center line fitting module, a basic lane line extraction module, a lane change scene detection module, a new lane line extraction module, and a complete lane line construction module. The lane trajectory clustering module is configured to perform trajectory clustering analysis based on a trajectory subset of a same road segment, obtain a trajectory cluster of a same lane, and divide the road segment into trajectory clusters of different lanes. The lane center line fitting module is configured to extract a trajectory geometric center fitting candidate lane center line based on a trajectory subset in a same trajectory cluster. The basic lane line extraction module is configured to obtain mutually disjoint basic lane lines based on the candidate lane lines. The lane change scene detection module is configured to determine whether the road segment belongs to a lane change scene based on trajectory data of the road segment. The new lane line extraction module is configured to extract a candidate lane center line as a new lane center line based on a newly added Gaussian component at both ends of the road segment. The complete lane line construction module is configured to construct lane geometry information and topological relationship in a lane change scene based on the extracted basic lane lines and new lane lines.

[0028] Further preferably, the basic lane line extraction module obtains the basic lane lines based on a maximum clique algorithm of a graph, constructs a candidate lane line intersection relationship graph using the fitted lane center lines of each road segment, defines each candidate lane line as a node of the graph, and connects the nodes to represent intersection relationship between lanes. If two lane lines do not intersect, the nodes representing the two lane lines are connected by an edge. A non-directional graph is constructed, and a set of mutually disjoint lane center lines is extracted as the basic lane lines through the non-directional graph.

[0029] Further preferably, each candidate lane line is defined as a vertex set V = {1, …, n} of the graph, and an edge set E between the nodes is If two lane lines do not intersect, the nodes representing the two lane lines are connected by an edge. A non-directional graph G = (V, E) is constructed with the vertex set V = {1, …, n} and the edge set E = { (u, v) | u, v ∈ V}. If a subset U ∈ V and for any two vertices u, v ∈ U, (u, v) ∈ E, then U is a complete subgraph of G. The maximum complete subgraph is the maximum clique of the graph.

[0030] Further preferably, when there are multiple maximum cliques, the set with the maximum sum of lane line node weights is selected as the basic lane line of a non-lane change. According to the formula:

[0031]

[0032] ​The weight of the computing node weight is transformed by a negative logarithmic function, so that the smaller the average change of the trajectory yaw angle is, the greater the weight of the lane line is, and the greater the average change of the trajectory yaw angle is, the smaller the weight of the trajectory is, wherein, Δyaw represents the average change of the trajectory yaw angle.

[0033] Further preferably, the prior Gaussian mixture model used to identify the number of lanes at both ends of the road section comprises: identifying the main direction of the road and the road cross-section direction perpendicular thereto through the motion direction of the trajectories in the window, projecting all GPS trajectory points in the road cross-section direction to obtain the projection distribution of the data in the road cross-section direction, and obtaining the position of each lane center line as the peak value μ of the Gaussian component of the lane j The distribution of the trajectory observation points on each lane corresponds to the standard deviation σ of the data distribution of each Gaussian component j The weight of each Gaussian component distribution is determined according to the proportion of the number of trajectories in each lane, and a probability density function of the Gaussian mixture model is constructed:

[0034]

[0035] In the formula, k is the number of Gaussian components, j = 1, 2, …, k, w j is the weight of the jth Gaussian component distribution, wherein,

[0036] w1+w2+…+W k = 1.

[0037] Further preferably, the evaluation function for optimizing the design of the number of lanes is selected,

[0038] The fitting of the Gaussian mixture model is evaluated, and overfitting is prevented. The lane width value is constrained, and the distance between the peak values of each Gaussian component distribution is controlled. As a penalty term for preventing overfitting when the number of mixed components increases, the Gaussian mixture model with the minimum evaluation function value is selected as the optimal model, wherein λ is a regularization parameter, and Δh i represents the average distance between the center lines of adjacent lanes.

[0039] According to another aspect of the present application, an electronic device is provided, comprising: a processor; and a memory storing a program, wherein the program includes instructions that, when executed by the processor, cause the processor to perform the trajectory-based lane geometric information extraction method described above.

[0040] According to another aspect of the present application, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause the computer to perform the trajectory-based lane geometric information extraction method described above

[0041] The application provides a fine lane geometry information extraction method and system based on trajectory data, extracts the geometric shape information of the lane center line through trajectory clustering and a non-parametric fitting method of trajectory iterative optimization, effectively solves the interference influence of the vehicle lane changing trajectory based on a maximum group lane changing identification method, considers the scene test of lane number change and the extraction of new lane information, and realizes the construction of lane connection relationship in the scene where the number of lanes changes. The method reduces the number of trajectory requirements and the amount of parameter solving calculation, reduces the dependence on prior information such as lane width and lane curvature, and is more suitable for lane information extraction in the lane changing scene, such as divergence and flow area. The application reduces the cost of obtaining detailed lane geometry information of urban roads, can effectively extract accurate lane geometry information based on trajectory data, the obtaining method is simple and effective, easy to implement, and has important significance for lane-level mapping application, and provides bottom layer data support for high-precision map of future upper-layer automatic driving application. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 The lane geometry information extraction system based on trajectory data in the example embodiment;

[0043] Figure 2 The lane geometry information extraction method flowchart based on trajectory data in the example embodiment;

[0044] Figure 3 The lane clustering effect schematic diagram obtained by the method in the embodiment;

[0045] Figure 4 The lane center line iterative optimization fitting schematic diagram in the example embodiment;

[0046] Figure 5 The lane changing identification and basic lane line extraction method schematic diagram based on the maximum group in the example embodiment;

[0047] Figure 6 The new lane line extraction schematic diagram based on the Gaussian mixture model in the example embodiment;

[0048] Figure 7 The complete lane line construction display schematic diagram in the lane changing scene based on the embodiment;

[0049] Figure 8 The principle schematic diagram of the electronic device in the embodiment.

[0050] Among them, 100-electronic device, 101-processor, 102-bus, 103-memory. DETAILED DESCRIPTION

[0051] Embodiments of the present application will be described below in greater detail with reference to the accompanying drawings. While certain embodiments of the present application are shown in the drawings, it is understood that the present application can be embodied in various forms and should not be interpreted as being limited to the embodiments set forth herein; rather, these embodiments are provided so that the present application will be thoroughly and completely understood. It is to be understood that the drawings of the present application are only for illustrative purposes and are not intended to limit the scope of the present application.

[0052] It is to be understood that each of the steps recited in the method embodiments of the present application can be performed in different orders and / or in parallel. In addition, the method embodiments can include additional steps and / or omit performing the steps shown. The scope of the present application is not limited in this regard.

[0053] As used herein, the term "comprising" and its variants are open-ended, that is "including, but not limited to." The term "based on" is "based, at least in part, on."

[0054] The present application extracts candidate lane lines by lane trajectory clustering and lane center line fitting, and designs recognition and processing methods for complex traffic scenarios where the number of lanes changes, automatically extracts basic lane lines and newly added lane change lines, and realizes complete and fine lane center line geometric information extraction. It includes: obtaining trajectory data set of each road section. Extracting road section sub-trajectory according to GPS points collected continuously at a specified time interval on the driving road section, clustering the start-end pairs of sub-trajectories to obtain trajectory clusters belonging to each road section.

[0055] The extraction of trajectory clusters belonging to the same lane completes the lane trajectory clustering. The following methods can be used to realize it:

[0056] Each trajectory segment is initialized as a cluster as an original cluster, clusters with a distance between two original clusters less than a distance threshold are merged into a new cluster, and the distance between the new cluster and the original cluster is further calculated, clusters with a distance less than the distance threshold are merged, until the distance between all merged clusters is greater than the distance threshold, and lane trajectory clustering is completed; for the lane center line fitting of the trajectory cluster belonging to the same lane, according to the time sequence of the trajectory observation points in the trajectory cluster of the same lane, an iterative optimization non-parametric curve fitting method is used to fit the output lane center line; the candidate lane center line extracted for each road section is used to construct a lane line intersection relationship graph, a undirected graph is constructed, and a set of lane center lines that do not intersect with each other is extracted, so that the basic lane line extraction is realized; the number of lanes at both ends of the road section is identified, and the lane change scene of the road section is determined; in the road section where the lane changes, the number of lanes at the increasing end of the lane number is taken as the total number of lanes of the road section, the extracted basic lane line data is corresponded with the Gaussian component at the increasing end of the lane number, the newly added Gaussian component is calculated to determine the newly added lane line, and the newly added lane line is extracted; the connection relationship between the newly added lane line and the basic lane line is constructed, the common subsequence part between the newly added lane line and the basic lane line is deleted, and the lane change connection position is recorded, so as to realize the construction of the lane geometry and topological relationship in the lane change scene.

[0057] The application adopts a modular design to design a corresponding practical system, and provides a lane geometry information extraction system based on trajectory data, which comprises:

[0058] The lane trajectory clustering module, the lane center line fitting module, the basic lane line extraction module, the lane change scene detection module, the newly added lane line extraction module and the complete lane line construction module.

[0059] The lane trajectory clustering module is used for dividing the trajectory data in a road section into trajectory clusters of different lanes, and obtaining the trajectory clusters of different lanes according to the trajectory data of the same road section; the lane center line fitting module is used for iterative optimization fitting of the lane center line, and obtains the geometric representation fitting candidate lane line of the lane center line according to the trajectory cluster of the same lane; the basic lane line extraction module is used for extracting the basic lane lines that do not intersect with each other, and obtains the basic lane lines that do not intersect with each other according to the candidate lane line fitted by the lane center line fitting module; the lane change scene detection module is used for judging whether a road section belongs to a scene where the number of lanes changes, and judges whether the road section belongs to a lane change scene according to the trajectory data of the same road section; the newly added lane line extraction module is used for extracting the newly added lane line of a road section, and determines the newly added lane line according to the trajectory data of the same road section and the extracted basic lane line; and the complete lane line construction module is used for constructing the complete lane line structure of a road section, and constructs the final lane line extraction result of the road section according to the extracted basic lane line and the newly added lane line.

[0060] The application will be further described below in combination with the drawings and embodiments.

[0061] As shown in Figure 1 The lane trajectory clustering module 1 is used to divide the trajectory data in the road section into trajectory clusters of different lanes. The lane center line fitting module 2 is used for iterative optimization fitting of the lane center line. The basic lane line extraction module 3 is used for extraction of the mutually disjoint basic lane lines. The lane change scene detection module 4 is used to determine whether the road section belongs to a scene in which the number of lanes changes. The new lane line extraction module 5 is used for extraction of the new lane line of the road section. The complete lane line construction module 6 is used for construction of the complete lane line structure of the road section.

[0062] As shown in Figure 2 The lane trajectory clustering module 1 is used to divide the trajectory data in the road section into trajectory clusters of different lanes. The lane center line fitting module 2 is used for iterative optimization fitting of the lane center line. The basic lane line extraction module 3 is used for extraction of the mutually disjoint basic lane lines. The lane change scene detection module 4 is used to determine whether the road section belongs to a scene in which the number of lanes changes. The new lane line extraction module 5 is used for extraction of the new lane line of the road section. The complete lane line construction module 6 is used for construction of the complete lane line structure of the road section.

[0063] The trajectory data of each road section is collected.

[0064] In this embodiment, the trajectory sequence composed of the GPS points collected at fixed time intervals on the driving road section is used, and the long time sequence trajectory is broken into sub-trajectories by taking the road intersection as the interval. The sub-trajectory origin-destination point pair clustering can be used based on the start and end positions of the sub-trajectory to divide all the sub-trajectories into trajectory sets of different road sections. The start and end positions of the sub-trajectory can quickly and effectively determine the spatial position of the trajectory, and the trajectory clustering (DBSCAN clustering can be used) is performed by calculating the distance sum between the start-end point pairs of the sub-trajectories, so that the sub-trajectories with similar spatial positions and relatively dense distribution are clustered into a cluster. Finally, the trajectory sets belonging to different road sections are obtained. The trajectory sets belonging to the same road section will be used as the input for the subsequent lane geometry information extraction.

[0065] Further, the lane trajectory clustering is performed. In this embodiment, the trajectory hierarchical clustering algorithm is used for lane trajectory clustering by taking the sub-trajectory set belonging to the same road section as the input, so as to further divide the sub-trajectory set of the same road section into different lanes, thereby extracting the trajectory clusters belonging to each lane.

[0066] The following method can be specifically used:

[0067] First, each trajectory segment in the road section is initialized as a cluster, and the Hausdorff distance between each pair of trajectories is calculated. The calculation of the distance metric considers the complete distribution of the trajectory sequence. According to the one-way Hausdorff distance d(A, B) from trajectory A to trajectory B and the one-way Hausdorff distance d(B, A) from trajectory B to trajectory A, the formula is called:

[0068] D h (A,B) = max{d(A,B),d(B,A)}

[0069]

[0070] Computing the bidirectional Hausdorff distance D between trajectory A and trajectory B h (A,B), where A and B are trajectories containing a series of ordered GPS observation points, b is the point in trajectory B closest to each point a in trajectory A, the distance between the two points is calculated, and the maximum value of the distance is taken as the value of d(A,B). Similarly, the one-way Hausdorff distance d(B,A) of trajectory B to trajectory A can be obtained. Finally, the bidirectional Hausdorff distance D h (A,B) takes the maximum value of d(A,B) and d(B,A).

[0071] Trajectories with a distance less than a distance threshold (which is preferably the lane width) are merged into a new trajectory cluster.

[0072] The merged cluster will be a new cluster, and the distance between the new cluster and other clusters will be recalculated. The distance between clusters is the maximum distance between two trajectories, and the clustering of lane trajectories is completed until the distance between clusters is greater than the set distance threshold.

[0073] As Figure 3 Fig. (b) shows the lane clustering result according to the original trajectory data in Fig. (a).

[0074] Iterative optimization fitting of lane center line. The lane center line is fitted by the trajectory clusters belonging to the same lane.

[0075] This embodiment is aimed at the time sequence of trajectory observation points, and the traditional principal curve algorithm for point set data fitting is improved. A non-parametric curve fitting method suitable for iterative optimization of trajectory data is designed. The geometric center of the trajectory is sequentially extracted in the form of a non-parametric curve to fit the lane center line within the same cluster. The lane center line fitting module outputs the lane center line according to the trajectory cluster of the same lane.

[0076] As Figure 4 shown in the lane center line iterative optimization fitting diagram in the example embodiment of the present application, in the figure,

[0077] (a) randomly select a trajectory as the initial lane line; (b) calculate the cross section line for each point of the initial lane line; (c) calculate the geometric center of the intersection point of the cross section line and the input trajectory cluster, take the geometric center as the center point of the lane, connect the intersection points, and update the fitted lane center line; (d) iterate steps (a) (b) until the position of the center point is stable and unchanged as the convergence condition, and obtain the final fitting result.

[0078] The basic lane line is extracted based on the lane change recognition method of the maximum clique.

[0079] Based on the feature that the actual lanes usually do not intersect, the basic lane center line selection problem of non-lane change is converted into solving the maximum clique problem in graph theory.

[0080] As shown in Figure 5 The maximum clique-based lane change recognition and basic lane line extraction method in the present exemplary embodiment is shown in the schematic diagram, in which:

[0081] (a) is the candidate lane center line, A, B, C, D, E, F, and G in the figure represent the numbers of each candidate lane center line, and the lane line intersection relationship graph is constructed based on the extracted candidate lane center line;

[0082] (b) is the lane line intersection relationship graph, each candidate lane line is defined as a node V = {A, B, C, D, E, F, G} in the graph, and the edge between the nodes represents the intersection relationship between the lanes. If two lane lines do not intersect, the edges are connected, thereby constructing an undirected graph;

[0083] (c) is the basic lane center line recognition, and the set {A, B, C, D} of mutually non-intersecting lane center lines is extracted as the basic lane line result;

[0084] (d) is the corresponding maximum clique recognition result, and the maximum clique result of the graph is in the box, containing the mutually non-intersecting basic lane line nodes A, B, C, and D. The remaining nodes E, F, and G in the graph are regarded as lane change lines and are excluded. The mutually non-intersecting basic lane center lines are extracted by identifying the maximum clique in the undirected graph, and the lane change lines are efficiently excluded.

[0085] In the present exemplary embodiment, the following maximum clique-based lane change recognition algorithm is used to construct an undirected graph, based on lane center line recognition, and extract a set of mutually non-intersecting lane center lines.

[0086] Each candidate lane line is defined as a vertex set V = {1, …, n} of the graph, and the edge set E between the nodes represents the intersection relationship between the lanes. If two lane lines do not intersect, the edges between the nodes represented by the two lane lines are connected. G = (V, E) is the vertex set V = {1, …, n} and the edge set An undirected graph is called a complete subgraph of G if a subset U ∈ V such that for any two vertices u, v ∈ U, (u, v) ∈ E. Figure 5 As shown, in (a), candidate lane lines A, B, C, D, E, F, and G are defined as the vertex set V = {A, B, C, D, E, F, G} of the graph. Taking lane line A as an example, lane line A does not intersect with lane lines B, C, D, F, and G, so the corresponding undirected edges (A, B), (A, C), (A, D), (A, F), and (A, G) are constructed respectively. This process continues, resulting in the lane intersection graph shown in (b). The maximum clique of a graph refers to the largest complete subgraph in the graph. Solving for the maximum clique yields the set of mutually disjoint basic lane lines {A, B, C, D}, which are retained as non-lane-changing basic lane lines as shown in (c). In (d), the area within the box represents the maximum clique in the graph, and the remaining nodes E, F, and G are considered lane-changing lines and are removed.

[0087] Considering the more complex situation where multiple maximum cliques exist simultaneously, this embodiment incorporates more trajectory geometric features for judgment. The likelihood of lane changes is assessed based on the average change in trajectory yaw angle, and node weight attributes are introduced. A negative logarithmic function transformation is used to assign higher weights to lanes with smaller average yaw angle changes and lower weights to lanes with larger average yaw angle changes, according to the formula:

[0088]

[0089] Calculate the node weight. In the formula, n represents the total number of trajectory points on the lane line represented by the node, and Δyaw represents the average change in the trajectory yaw angle.

[0090] This invention transforms the problem into a maximum weight clique problem by introducing node weight attributes. When multiple maximum cliques exist, the set with the largest sum of weights of lane line nodes is selected as the basic lane line for non-lane change.

[0091] Lane change scenario detection.

[0092] This embodiment uses a priori Gaussian mixture model for lane change scene detection (other models can also be used for identification). It identifies the number of lanes at both ends of a road segment to determine if the segment belongs to a lane change scenario. If the number of lanes at both ends of the road segment is the same, it indicates that there is no lane change in that segment, and the number of lanes in the segment is equal to the previously extracted basic lane lines. The extracted basic lane lines are then merged to complete the final lane line structure integration and construction.

[0093] If the detection results show that the number of lanes at both ends of a road segment changes, then the road segment is determined to be a scenario where lanes change.

[0094] Further, the prior Gaussian mixture model adopted in the example embodiment specifically includes the following steps for identifying the number of lanes at both ends of a road section: in combination with the distribution law of GPS trajectory data, the projection distribution of trajectory points on the road cross section has the characteristic of being gathered around the center line of each lane, and then gradually sparse towards both ends of the lane. Starting from the beginning of the road section, a rectangular moving window is constructed, with the maximum coverage width of the trajectory on the road as the side length of the window, the trajectory observation points in the window are detected, and the window is moved step by step along the main direction of the road until the end of the road section. The main direction of the road and the direction perpendicular to the road cross section are identified through the moving direction of the trajectory in the window, all GPS trajectory points in the window are projected on the road cross section direction to obtain the projection distribution of the data on the road cross section direction, and the Gaussian mixture model is modeled to obtain the position of the center line of each lane as the peak value μ j of each Gaussian component in the Gaussian mixture model, the standard deviation σ j of the data distribution corresponding to each Gaussian component according to the distribution of the trajectory observation points on each lane, and the weight of each component distribution in the Gaussian mixture model is determined according to the proportion of the number of trajectories in each lane to construct the probability density function of the Gaussian mixture model:

[0095] The density function p(x) is used to represent the comprehensive probability of the Gaussian mixture model. In the formula, x represents the trajectory point sample projected to the road cross section direction to be calculated, k is the number of Gaussian mixture components, j = 1, 2, …, k, μ j represents the average value of the sample belonging to the jth Gaussian component, σ j represents the standard deviation of the sample in the jth Gaussian component, w j is the weight of the jth Gaussian component distribution, and w1+w2+…+w k =1.

[0096] The classic Gaussian mixture model needs to manually set the number of mixtures, i.e. the number of lanes, and the Akaike information criterion and the Bayesian information criterion are usually used to determine the number of components in the mixture model and determine the best fitting model. However, for the application scenario of lane number identification, the information criterion judgment in the past may appear overfitting, and the fitting effect seems to be better with the increase of the number of components. In order to solve the problem of too many mixed components caused by overfitting, the prior condition of variable lane width is introduced on the basis of the traditional Gaussian mixture model in the embodiment, the evaluation function for selecting the number of lanes is optimized and designed, and the applicability of the algorithm in lane number identification is improved. The evaluation function for selecting the number of lanes is as follows:

[0097]

[0098] In the formula, for the trajectory point n represents the total number of track points of the lane line represented by the node, k is the number of Gaussian mixture components, p(x i |θ k ) represents the sample value x i The Gaussian probability value under the condition of the Gaussian mixture model parameter θ k The model parameter θ k may be represented as θ k (w k ,μ k ,σ). The first term on the right side of the equation is the log-likelihood function for evaluating the model fitting, and the second term is a regularization term for preventing overfitting, where λ is the regularization parameter, the default value is 1, and Δh i represents the average distance between adjacent lane center lines. A sigmoid function is introduced in the regularization term for lane width information constraint. Since the lane width value is usually within a certain range, such as 3.25m-3.75m, in actual urban road conditions, the sigmoid function can map real values to the interval (0, 1), and the lane width value in the model is constrained within a certain range, thereby controlling the distance between the peak distributions of the Gaussian components, as a penalty term for preventing overfitting as the number of mixture components increases. The final model selects the Gaussian mixture model with the minimum function evaluation value as the optimal model, determines the number of mixed components as the number of lanes, and thus automatically identifies the number of lanes. For the lane change scene, new lane lines are extracted. As shown in Figure 6 , it is a schematic diagram of new lane line extraction based on the Gaussian mixture model in the present exemplary embodiment.

[0099] In the road section where the lane changes, new lane lines other than the basic lane lines are extracted. New lane lines are identified by adding new Gaussian components to the basic lane lines, and the track data is projected onto the road cross section after adding the new lane lines.

[0100] The number of lanes at the end of the increased lane is taken as the total number of lanes after the actual lane increase. The extracted basic lane line data is matched with the Gaussian components at the end of the increased lane, and the difference set of all Gaussian components at the end of the increased lane and the corresponding Gaussian components of the basic lane line is taken as the new Gaussian component. The new lane line is extracted by extracting the corresponding lane center line of the new Gaussian component.

[0101] Complete lane line construction in the lane change scene. For the scene where the number of lanes changes, the obtained disjoint basic lane lines and the new lane lines are retained. When the new lane line partially overlaps with the basic lane line, the common subsequence part between the new lane line and the basic lane line is deleted, the connection relationship between the new lane line and the basic lane line is retained, and the lane change connection position is recorded, so as to realize the construction of lane geometry and topological relationship in the lane change scene.

[0102] Figure 7 A complete schematic diagram of complete lane line construction in the lane change scene is shown. The finally extracted lane line is composed of disjoint basic lane lines and new lane line parts reflecting the real change of the road network, and the task of lane line construction is completed. Among them, (a) original trajectory data, (b) trajectory clustering, (c) lane center line fitting, (d) basic lane line extraction, (e) new lane line extraction.

[0103] As Figure 8 shown, the electronic device 100 includes a processor 101 and a memory 103. Among them, the processor 101 and the memory 103 are connected, such as connected through a bus 102.

[0104] The structure of the electronic device 100 does not constitute a limitation on the embodiments of the present application.

[0105] The processor 101 can be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It can realize or execute various exemplary logical blocks, modules and circuits described in combination with the disclosure content of the present application. The processor 101 can also be a combination of computing functions, such as one or more microprocessor combinations, combinations of DSP and microprocessor, etc.

[0106] The bus 102 can include a channel for transmitting information between the above-mentioned components. The bus 102 can be a PCI bus or an EISA bus, etc. The bus 102 can be divided into an address bus, a data bus, a control bus, etc. For the sake of expression, Figure 3 only one thick line is used in the figure, but it does not mean that there is only one bus or only one type of bus.

[0107] The memory 103 can be a ROM, or other type of static storage device that can store static information and instructions; a RAM, or other type of dynamic storage device that can store information and instructions; a EEPROM, CD-ROM or other optical disk storage, a magnetic disk storage or other magnetic storage devices, or any other medium capable of storing desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited to this.

[0108] The applicant makes detailed description and illustration of the embodiments of the present application in combination with the drawings of the present application. However, those skilled in the art should understand that the above embodiments are only the preferred embodiments of the present application, and the detailed description is only for better understanding of the spirit of the present application, and is not a limitation on the protection scope of the present application. Those skilled in the art can make various modifications or supplements to the described specific embodiments or adopt similar ways to replace them, but will not deviate from the spirit of the present application or exceed the scope defined by the appended claims.

Claims

1. A method for extracting lane geometry information based on road segment trajectory, characterized in that, Trajectory clustering analysis is performed on subsets of trajectories passing through the same road segment to divide the sub-trajectory sets of the same road segment into different lanes, and trajectory clusters belonging to each lane are extracted. The geometric centers of the trajectory subsets within the same cluster are extracted in the form of non-parametric curves, and candidate lane centerlines are fitted. The lane centerlines are regarded as nodes in the graph to establish an intersection relationship graph between lane lines. The lane change lane lines are identified by the maximum clique, and a set of non-intersecting lane centerlines is obtained as the basic lane lines. Based on the newly added Gaussian components at both ends of the road segment, candidate lane centerlines are extracted as new lane centerlines. The lane change connection positions are determined according to the basic lane lines and the new lane centerlines, and a lane geometry information and topology relationship graph is constructed. Obtaining the basic lane lines involves treating lane lines as nodes in a graph, with node weights recording the geometric features of the lane lines and edge weights recording the intersection relationship features between lane lines. A candidate lane line intersection relationship graph is constructed using the lane center lines fitted to each road segment. Each candidate lane line is defined as a node in the graph, and the edges between nodes represent the intersection relationship between lanes. If two lane lines do not intersect, the edges are connected. An undirected graph is constructed, and a set of non-intersecting lane center lines is extracted from the undirected graph as the basic lane lines. Each candidate lane line is defined as the set of vertices V = {1, ..., n} of the graph, and the set of edges between nodes. This represents the intersection relationship between lanes. If two lane lines do not intersect, then the edges between the nodes represented by these two lane lines are connected. Construct G = (V, E) as the vertex set V = {1, ..., n} and the edge set... For an undirected graph, if a subset U∈V and for any two vertices u, v∈U, (u, v)∈E, then U is a complete subgraph of G, and the largest complete subgraph is the largest clique of the graph.

2. The method according to claim 1, characterized in that, Lane centerline fitting includes: fitting the lane centerline using an iterative optimization non-parametric curve fitting method based on the temporal sequence of trajectory observation points in the trajectory cluster of the same lane; randomly selecting a trajectory line as the initial lane line, calculating the cross-sectional line of each point of the lane line, calculating the geometric center of the intersection of the cross-sectional line and the input trajectory cluster as the center point of the lane, and connecting the intersection points of each trajectory cluster as the fitted lane centerline.

3. The method according to claim 1, characterized in that, In road sections where lane numbers change, the number of lanes at the end where the number of lanes increases is taken as the total number of lanes in the road section. Based on the Gaussian components corresponding to the basic lane lines and the ends where the number of lanes increases, the newly added Gaussian components are calculated to determine and extract the new lane lines.

4. The method according to any one of claims 1-3, characterized in that, Lane trajectory clustering includes: initializing each trajectory segment as a cluster and calculating the Hausdorff distance between each pair of trajectories; merging clusters with Hausdorff distances less than a distance threshold into a new trajectory cluster; calculating the inter-cluster distance between the new trajectory cluster and other trajectory clusters; merging clusters with inter-cluster distances less than a distance threshold into trajectory clusters of the same lane; and so on until all inter-cluster distances are greater than the distance threshold, thus completing lane trajectory clustering.

5. The method according to any one of claims 1-3, characterized in that, Lane trajectory clustering includes: density-based clustering algorithms to obtain trajectory clusters belonging to each road segment lane, and using the set of all trajectories belonging to the same road segment as the basic unit for lane extraction.

6. The method according to claim 1, characterized in that, When multiple maximum cliques exist, the set with the largest sum of lane line node weights will be selected as the basic lane line for non-lane-changing, according to the formula: The node weights are calculated by transforming them using a negative logarithmic function. Lanes with smaller average changes in trajectory yaw angle have larger weights, while trajectories with larger average changes in trajectory yaw angle have smaller weights. Here, Δyaw represents the average change in trajectory yaw angle.

7. The method according to any one of claims 1-3, 6, characterized in that, Identifying the number of lanes at both ends of a road segment includes: using the maximum coverage width of the trajectory on the road surface as the side length of the window, detecting trajectory observation points within the window; starting from the beginning of the road segment, moving gradually along the main direction of the road until the end of the road segment; identifying the main direction of the road and the road cross-sectional direction perpendicular to it by the movement direction of the trajectory within the window; projecting all GPS trajectory points within the window onto the road cross-sectional direction to obtain the projection distribution of data on the road cross-sectional direction; and obtaining the position of the centerline of each lane as the peak value μ of the Gaussian component of that lane. j The distribution of trajectory observation points on each lane corresponds to the standard deviation σ of the distribution of each Gaussian component data. j The weights of each Gaussian component distribution are determined based on the proportion of trajectories in each lane, and the probability density function of the Gaussian mixture model is constructed as follows: In the formula, k is the number of Gaussian components, j = 1, 2, ..., k, w j Let be the weight of the j-th Gaussian component distribution, where w1+w2+…+w k =1。 8. The method according to claim 7, characterized in that, The evaluation function for optimizing lane number selection is based on the formula: To evaluate the fit of the Gaussian mixture model and prevent overfitting, constraints are imposed on the lane width values. The distance between the peak distributions of each Gaussian component is controlled as a penalty term in the likelihood function to prevent overfitting as the number of mixture components increases. The Gaussian mixture model with the smallest evaluation function value is selected as the optimal model, where λ is the regularization parameter and Δh... i This indicates the average distance between the center lines of adjacent lanes.

9. A lane geometry information extraction system based on road segment trajectory, characterized in that, include: The system includes a lane trajectory clustering module, a lane centerline fitting module, a basic lane line extraction module, a lane change scene detection module, a new lane line extraction module, and a complete lane line construction module. The lane trajectory clustering module is used to perform trajectory clustering analysis based on the trajectory subsets of the same road segment, divide the sub-trajectory set of the same road segment into different lanes, and extract the trajectory clusters belonging to each lane. The lane centerline fitting module extracts the geometric center of the trajectory from a subset of trajectories within the same cluster in the form of a non-parametric curve, and fits the candidate lane centerlines. The basic lane line extraction module obtains non-intersecting basic lane lines based on candidate lane lines; The lane change scenario detection module determines whether a road segment belongs to a lane change scenario based on the trajectory data of the road segment; the new lane line extraction module extracts candidate lane center lines based on the newly added Gaussian components at both ends of the road segment as the new lane center lines; the complete lane line construction module constructs lane geometry information and topological relationships under the lane change scenario based on the extracted basic lane lines and the new lane lines. The basic lane line extraction module obtains basic lane lines based on the maximum clique algorithm of the graph. The lane center line is regarded as a node in the graph. At the same time, the node weight records the geometric features of the lane line, and the edge weight records the intersection relationship features between lane lines. The candidate lane line intersection relationship graph is constructed using the lane center lines fitted by each road segment. Each candidate lane line is defined as a node of the graph, and the edge between the nodes represents the intersection relationship between lanes. If two lane lines do not intersect, the edge is connected. An undirected graph is constructed, and the set of non-intersecting lane center lines is extracted as the basic lane lines through the undirected graph. Each candidate lane line is defined as the set of vertices V = {1, ..., n} of the graph, and the set of edges between nodes. This represents the intersection relationship between lanes. If two lane lines do not intersect, then the edges between the nodes represented by these two lane lines are connected. Construct G = (V, E) as the vertex set V = {1, ..., n} and the edge set... For an undirected graph, if a subset U∈V and for any two vertices u, v∈U, (u, v)∈E, then U is a complete subgraph of G, and the largest complete subgraph is the largest clique of the graph.

10. The system according to claim 9, characterized in that, When multiple maximum cliques exist, the set with the largest sum of lane line node weights will be selected as the basic lane line for non-lane-changing, according to the formula: The node weights are calculated by transforming them using a negative logarithmic function. Lanes with smaller average changes in trajectory yaw angle have larger weights, while trajectories with larger average changes in trajectory yaw angle have smaller weights. Here, Δyaw represents the average change in trajectory yaw angle.

11. The system according to any one of claims 9-10, characterized in that, A prior Gaussian mixture model is used to identify the number of lanes at both ends of a road segment. The maximum coverage width of the trajectory on the road surface is used as the side length of the window. Starting from the beginning of the road segment, trajectory observation points within the window are detected and gradually moved along the main direction of the road until the end of the segment. The main direction of the road and the road cross-sectional direction perpendicular to it are identified by the movement direction of the trajectory within the window. All GPS trajectory points within the window are projected onto the road cross-sectional direction to obtain the projection distribution of the data in the road cross-sectional direction. The position of the centerline of each lane is obtained as the peak value μ of the Gaussian component of that lane. j The distribution of trajectory observation points on each lane corresponds to the standard deviation σ of the distribution of each Gaussian component data. j The weights of each Gaussian component distribution are determined based on the proportion of trajectories in each lane, and the probability density function of the Gaussian mixture model is constructed as follows: In the formula, k is the number of Gaussian components, j = 1, 2, ..., k, w j Let w1 + w2 + ... + w be the weights of the j-th Gaussian component distribution, where w1 + w2 + ... + w are the weights of the distribution. k =1.

12. The system according to claim 11, characterized in that, The evaluation function for optimizing the selection of the number of lanes in the design. To evaluate the fit of the Gaussian mixture model and prevent overfitting, constraints are imposed on the lane width values. The distance between the peak distributions of each Gaussian component is controlled as a penalty term in the likelihood function to prevent overfitting as the number of mixture components increases. The Gaussian mixture model with the smallest evaluation function value is selected as the optimal model, where λ is the regularization parameter and Δh... i This indicates the average distance between the center lines of adjacent lanes.

13. An electronic device, comprising: processor; And a memory for storing programs, characterized in that the programs include instructions that, when executed by the processor, cause the processor to perform the lane geometry information extraction method based on road segment trajectory according to any one of claims 1-8.

14. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, in, The computer instructions are used to cause the computer to execute the lane geometry information extraction method based on road segment trajectory according to any one of claims 1-8.

Images