Local road topology online construction method, system and product

Abstract

The invention provides a local road topology on-line construction method, system and product, and the method comprises the steps: predicting a lane center line through a training qualified lane center line prediction network based on lane line data, and obtaining a target lane center line set, the lane center line being composed of a series of ordered data points; under the condition that a new lane center line is obtained through prediction, transverse truncation is carried out on all the lane center lines through the end points of the new lane center line; calculating the geometric distance between the terminal point of the lane center line before the transverse cut-off position and the starting point of the lane center line after the transverse cut-off position to obtain a corresponding calculation result; and determining the branch and confluence road topology between the lane center lines according to the calculation result. The method aims to predict road topology in an end-to-end mode, so that the calculation power is saved, and the real-time performance of track and intention prediction is improved at the same time.

Description

Technical Field

[0001] The present application relates to the field of point cloud processing technology, and specifically to a method, system and product for online construction of local road topology. Background Art

[0002] In the field of autonomous driving, the intention and trajectory prediction of traffic participants plays a vital role in downstream decision-making and planning. Whether the intention and trajectory of traffic participants can be accurately predicted in this process is highly dependent on the road topology structure near the vehicle and the traffic participants (including: lane centerlines, relative position relationships of lane centerlines, whether there are divergent flows, etc.).

[0003] In recent years, the online construction of map elements around the vehicle and the target through surround view images and perception algorithms has become increasingly mature. However, existing online methods only generate map elements. In order to construct the road topology structure required for prediction, an additional model (prediction environment model) is required to reconstruct the road topology structure. This method not only adds additional computing power, but also increases the delay between modules in real-time operation, ultimately reducing the prediction accuracy of the target trajectory and intention. Summary of the invention

[0004] In view of this, the present application provides a method, system and product for online construction of local road topology, aiming to predict road topology in an end-to-end manner, saving computing power while improving the real-time performance of trajectory and intention prediction.

[0005] The first aspect of the present application provides a method for online construction of local road topology, the method comprising:

[0006] Based on the lane line data, the lane center line is predicted by training a qualified lane center line prediction network to obtain a target lane center line set. The lane center line consists of a series of ordered data points.

[0007] When a new lane centerline is predicted, all lane centerlines are transversely cut off using the endpoints of the new lane centerline;

[0008] Calculating a geometric distance between an end point of the lane centerline before the lateral cutoff position and a start point of the lane centerline after the lateral cutoff position to obtain a corresponding calculation result;

[0009] According to the calculation results, the diverging and merging road topology between the lane center lines is determined.

[0010] Optionally, when a new lane center line is predicted, before all lane center lines are transversely truncated with the endpoint of the new lane center line, the method further includes:

[0011] Different multi-layer perceptrons are used to encode the centerline of the lane where the vehicle is located and other lane centerlines in the target lane centerline set and map them to a high-dimensional semantic space to obtain the multi-dimensional vector features of each lane centerline.

[0012] Perform dimensionality reduction processing on multi-dimensional vector features and extract the sequence information of data points to obtain low-dimensional vector features;

[0013] The inner product of the low-dimensional vector features of the center line of the lane where the vehicle is located is calculated with the low-dimensional vector features of the center lines of each other lane to determine the positional relationship between the center lines of each other lane and the center line of the lane where the vehicle is located;

[0014] After determining the diverging flow road topology between lane center lines according to the calculation results, the method further includes: determining the positional relationship between the new lane center line and the center line of the lane where the vehicle is located according to the diverging flow road topology and the positional relationship.

[0015] Optionally, before predicting the lane centerline based on the lane line data by using a trained lane centerline prediction network to obtain the target lane centerline set, the method further includes:

[0016] Constructing a training data set, wherein the sample data in the training data set includes sample lane line data and a set of real lane center lines corresponding to the sample lane line data;

[0017] Based on the sample lane line data in the training data set, the lane center line is predicted by the initial lane center line prediction network to obtain a lane center line set;

[0018] Performing category matching on the lane centerline set and the corresponding real lane centerline set to obtain a target category matching result;

[0019] Performing point set matching on the lane center lines that match each other in the target category matching result and the real lane center line to obtain a target point set matching result;

[0020] The target category matching result and the target point set matching result are calculated by using a loss function to obtain a loss value;

[0021] When the loss value satisfies a set condition, determining that the initial lane centerline prediction network is trained to be qualified, and obtaining a trained lane centerline prediction network;

[0022] When the loss value does not meet the set conditions, the model parameters of the initial lane centerline prediction network are updated and the model training is performed again.

[0023] Optionally, performing category matching on the lane centerline set and a corresponding real lane centerline set to obtain a target category matching result includes:

[0024] The lane centerline set and the corresponding real lane centerline set are matched by category, and the matching cost is calculated for the obtained matching result by using the category matching cost function to obtain the matching cost value corresponding to the matching result. The category matching cost function is: in Indicates the prediction under the corresponding matching result and real v i The pairwise matching cost between v i ={c i ,P i}, c i , P i Represent the category label and point set of the real lane centerline respectively, and Respectively represent the corresponding matching results and the real v i Matching predictions The predicted category scores and predicted point sets of

[0025] According to the matching cost values ​​corresponding to the obtained matching results, the matching result with the smallest matching cost value is determined as the target category matching result.

[0026] Optionally, define the matching cost as in, is the cost term for category matching; Represents the position matching loss term of the point set.

[0027] Optionally, a geometric distance between an end point of the lane centerline before the lateral cutoff position and a start point of the lane centerline after the lateral cutoff position is calculated to obtain a corresponding calculation result, including:

[0028] The geometric distance between the end point of the lane centerline before the lateral truncation position and the starting point of the lane centerline after the lateral truncation position is calculated by the target algorithm to obtain a geometric distance matrix between the end point of the lane centerline before the lateral truncation position and the starting point of the lane centerline after the lateral truncation position. The target algorithm is: D = {d ij |i,j=0,…n-1}, where D is the geometric distance matrix between lane centerlines, l i Indicates the lane centerline before the lateral cutoff position i, l j Indicates the lane centerline j after the lateral cutoff position, d ijIndicates the lane centerline l i and l j The geometric distance between Indicates the lane centerline l i The end point, Indicates the lane centerline l j Starting point;

[0029] The geometric distance matrix is ​​calculated by a mapping function to obtain a corresponding calculation result. The mapping function is: where x = d ij , σ is the standard deviation of the geometric distance matrix D, and σ and λ are learnable parameters.

[0030] Optionally, determining the diverging and merging road topology between lane centerlines according to the calculation result includes:

[0031] In the calculation result, when the mapping values ​​between the end point of a lane centerline before the lateral truncation position and the starting points of at least two lane centerlines after the lateral truncation position meet the first condition, the corresponding position is determined to be a diverging road topology;

[0032] In the calculation result, when the mapping value between the starting point of a lane centerline after the lateral truncation position and the end points of at least two lane centerlines before the lateral truncation position satisfies the first condition, the corresponding position is determined to be a converging road topology.

[0033] Optionally, the method further includes:

[0034] Constructing a first loss function related to lane centerline prediction, constructing a second loss function related to determining a positional relationship between lane centerlines, and constructing a third loss function related to determining a diverging and converging road topology between lane centerlines;

[0035] Determine a target loss function based on the constructed first loss function, second loss function, and third loss function;

[0036] The target loss function is used to guide the training of a target neural network, which is a neural network used in the process of predicting lane center lines, determining the positional relationship between lane center lines, and determining the diverging and merging road topology between lane center lines.

[0037] Optionally, the method further includes:

[0038] Construct lane detection loss function;

[0039] The determining of the target loss function based on the constructed first loss function, second loss function and third loss function includes:

[0040] Based on the constructed first loss function, second loss function, third loss function and the lane line detection loss function, a target loss function is determined, and the target function is: in, is the first loss function, is the second loss function, is the third loss function, is the lane detection loss function, λ1, λ2, and λ3 are learnable parameters;

[0041] A second aspect of the present application provides a local road topology online construction system, the system comprising:

[0042] The lane centerline prediction module is used to predict the lane centerline based on the lane line data through a trained lane centerline prediction network to obtain a target lane centerline set. The lane centerline consists of a series of ordered data points.

[0043] A truncation module, configured to perform transverse truncation of all lane centerlines using the endpoints of the new lane centerline when a new lane centerline is predicted;

[0044] A geometric distance calculation module is used to calculate the geometric distance between the end point of the lane centerline before the transverse cutoff position and the start point of the lane centerline after the transverse cutoff position to obtain a corresponding calculation result;

[0045] The road topology determination module is used to determine the diverging and merging road topology between lane center lines according to the calculation results.

[0046] The third aspect of the present application provides an electronic device, comprising: a processor, a memory, and a computer program stored in the memory and running on the processor, wherein when the computer program is executed by the processor, the steps in the method for online construction of a local road topology as described in the first aspect of the present application are implemented.

[0047] The fourth aspect of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps in the method for online construction of local road topology as described in the first aspect of the present application are implemented.

[0048] The local road topology online construction method provided by this application has the following advantages:

[0049] An embodiment of the present application provides an online construction method for local road topology. First, the vehicle's existing lane line detection module detects the lane line in front of the vehicle on the surround view image collected by the lane to obtain all lane line instances in front of the vehicle in the current scene (that is, all lane lines). Then, for the lane line data composed of the obtained lane line instances, a qualified lane centerline prediction network is trained to predict the lane centerline to obtain a target lane centerline set, where the lane centerline consists of a series of ordered data points. In the continuous prediction process, if a new lane centerline is predicted, all lane centerlines are transversely truncated with the endpoint of the new lane centerline. The geometric distance between the end point of the lane centerline before the transverse truncation position and the starting point of the lane centerline after the transverse truncation position is calculated to obtain a corresponding calculation result. According to the calculation result, the diverging and merging road topology between the lane centerlines is determined. Therefore, the present application predicts the lane centerline and calculates the lane centerline branch flow road topology by adding the corresponding lane centerline prediction generation module and the lane centerline branch flow road topology calculation module to the existing lane line detection module, thereby achieving the goal of using the surround view image as input and outputting the predicted road topology in an end-to-end mode. This method does not require the construction of a separate model (prediction environment model) to reconstruct the road topology structure, thereby reducing computing power and improving processing efficiency, and ultimately improving the real-time performance of trajectory and intention prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments of the present application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor.

[0051] Figure 1 A flowchart of a method for online construction of local road topology is shown as an embodiment of the present application;

[0052] Figure 2 A schematic diagram of lateral truncation in a method for online construction of a local road topology according to an embodiment of the present application;

[0053] Figure 3 Another schematic diagram of lateral truncation in a method for online construction of a local road topology according to an embodiment of the present application;

[0054] Figure 4 A schematic diagram of traffic diversion in a method for online construction of a local road topology is shown as an embodiment of the present application;

[0055] Figure 5A schematic diagram of a processing procedure of a method for online construction of a local road topology according to an embodiment of the present application;

[0056] Figure 6 Another flow chart of a method for online construction of local road topology shown in one embodiment of the present application;

[0057] Figure 7 A schematic diagram of a local road topology online construction system shown in one embodiment of the present application. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0059] refer to Figure 1 , Figure 1 This is a schematic diagram of a method for online construction of local road topology according to an embodiment of the present application. Figure 1 As shown, the method includes:

[0060] Step S1: Based on the lane line data, the lane center line is predicted by a trained qualified lane center line prediction network to obtain a target lane center line set, where the lane center line consists of a series of ordered data points.

[0061] In this embodiment, the lane line detection module of the vehicle is used to perform lane line detection on the surround view image collected by the camera of the vehicle to obtain all lane line instances in front of the vehicle, that is, all lane lines. These lane line instances constitute lane line data. The obtained lane line data is processed by a trained lane center line prediction network to obtain the lane center lines between all lane lines in the lane line data. All the obtained lane center lines constitute a target lane center line set, which consists of a series of ordered data points.

[0062] Step S2: When a new lane center line is predicted, all lane center lines are transversely cut using the endpoints of the new lane center line.

[0063] In this embodiment, the lane centerline prediction network will continuously predict the lane centerline based on the new lane line data. After predicting the new lane centerline at the next moment, since the lane centerline is composed of a series of ordered data points, the endpoint position of the new lane centerline can be determined accordingly. Then, all lane centerlines at the next moment are transversely truncated with the endpoint of the new lane centerline, and the data points of each lane centerline at the transverse truncation position are analyzed to determine whether the position belongs to the diverging road topology or the converging road topology. Figure 2 As shown, Figure 2 The vehicle in the figure is the ego vehicle. During the driving process of the ego vehicle, a new lane centerline of a diverging flow road will appear at a corresponding moment (it should be understood that the lane centerline of the diverging flow lane may be the lane centerline of the diverging flow road in the lane where the ego vehicle is located, or it may be the lane centerline of the diverging flow road that appears in the lane next to the lane where the ego vehicle is located. Figure 2 For example, only the lane centerline of the diverging road of the lane where the vehicle is located is shown. Based on a series of ordered data points of the newly appeared lane centerline, its endpoint position can be determined, and then all lane centerlines are transversely truncated at the endpoint position.

[0064] Step S3: Calculate the geometric distance between the end point of the lane centerline before the lateral cutoff position and the start point of the lane centerline after the lateral cutoff position to obtain a corresponding calculation result.

[0065] In this embodiment, after all lane center lines are transversely cut off in step S2, the geometric distances between the end points of the lane center lines before the transverse cut-off position and the start points of the lane center lines after the transverse cut-off position are calculated to obtain corresponding calculation results. Figure 3 As shown, Figure 3 A schematic diagram of lateral truncation of a lane centerline and corresponding calculation is shown exemplarily. For example, the lane centerlines before the lateral truncation position include L3 and L4, and the lane centerlines after the lateral truncation position include L1, L2 and L5. The geometric distances between the end points of the lane centerlines before the lateral truncation position and the starting points of the lane centerlines after the lateral truncation position are calculated. It can be determined that the geometric distances between the end points of the lane centerline L3 before the lateral truncation position and the starting points of the lane centerlines L1, L2 and L5 after the lateral truncation position are zero (or approximately zero), non-zero and non-zero, respectively. It can be determined that the geometric distances between the end points of the lane centerline L4 before the lateral truncation position and the starting points of the lane centerlines L1, L2 and L5 after the lateral truncation position are non-zero, zero (or approximately zero) and zero (or approximately zero), respectively.

[0066] Step S4: Determine the diverging and merging road topology between lane centerlines based on the calculation results.

[0067] In this embodiment, after the corresponding calculation results are obtained through step S3, the road topology of the diverging flows between the lane center lines is determined based on the calculation results. For example, the example in the above step S3 is continued to be used. Since only one of the geometric distances between the end point of the lane center line L3 before the transverse cutoff position and the starting point of the lane center lines L1, L2 and L5 after the transverse cutoff position is zero (or approximately zero), there is no diverging flow after the lane center line L3. Since two of the geometric distances between the end point of the lane center line L4 before the transverse cutoff position and the starting point of the lane center lines L1, L2 and L5 after the transverse cutoff position are zero (or approximately zero), a road topology relationship of diverging flows appears after the lane center line L4. The determination of the corresponding converging road topological relationship is similar. For example, if the geometric distances between the starting points of a lane centerline (such as L6) after the transverse cutoff position and the end points of two lane centerlines greater than or equal to the transverse cutoff position are zero (or approximately zero), it is determined that a converging road topological relationship occurs before the lane centerline (L6).

[0068] An embodiment of the present application provides an online construction method for local road topology. First, the vehicle's existing lane line detection module detects the lane line in front of the vehicle on the surround view image collected by the lane to obtain all lane line instances in front of the vehicle in the current scene (that is, all lane lines). Then, for the lane line data composed of the obtained lane line instances, a qualified lane centerline prediction network is trained to predict the lane centerline to obtain a target lane centerline set, where the lane centerline consists of a series of ordered data points. In the continuous prediction process, if a new lane centerline is predicted, all lane centerlines are transversely truncated with the endpoint of the new lane centerline. The geometric distance between the end point of the lane centerline before the transverse truncation position and the starting point of the lane centerline after the transverse truncation position is calculated to obtain a corresponding calculation result. According to the calculation result, the diverging and merging road topology between the lane centerlines is determined. Therefore, the present application predicts the lane centerline and calculates the lane centerline branch flow road topology by adding the corresponding lane centerline prediction generation module and the lane centerline branch flow road topology calculation module to the existing lane line detection module, thereby achieving the goal of using the surround view image as input and outputting the predicted road topology in an end-to-end mode. This method does not require the construction of a separate model (prediction environment model) to reconstruct the road topology structure, thereby reducing computing power and improving processing efficiency, and ultimately improving the real-time performance of trajectory and intention prediction.

[0069] In combination with the above embodiments, in one implementation, the embodiment of the present application further provides a method for online construction of local road topology. In the method for online construction of local road topology, before step S2, the method further includes steps S01 to S03:

[0070] Step S01: The center line of the lane where the vehicle is located and the center lines of other lanes in the target lane center line set are respectively encoded and mapped to a high-dimensional semantic space through different multi-layer perceptrons to obtain the multi-dimensional vector features of each lane center line.

[0071] In this embodiment, in the road topology construction, the relative position relationship between lanes is also very important. Therefore, the present application will also determine the relative position relationship between lanes in the road topology construction. Specifically: first, the lane centerline of the lane where the ego vehicle is located in the target lane centerline set is determined, and the distance between the coordinates of the ego vehicle and all lane centerlines is used to determine which lane centerline the ego vehicle's coordinates are closest to. Correspondingly, the lane corresponding to the closest lane centerline is determined as the lane where the ego vehicle is located. Then, a multi-layer perceptron (MLP) is used to encode the features of the lane centerline of the lane where the ego vehicle is located in the target lane centerline set. In order to enhance the feature expression ability and the generalization ability of the multi-layer perceptron, the encoded features are mapped to a high-dimensional semantic space to obtain a multi-dimensional feature vector of the lane centerline of the lane where the ego vehicle is located. At the same time, another multi-layer perceptron (MLP) is used to encode the features of the lane centerlines of the target lane centerline set except the lane centerline of the lane where the vehicle is located. In order to enhance the feature expression ability and the generalization ability of the multi-layer perceptron, the encoded features are mapped to a high-dimensional semantic space to obtain the multi-dimensional feature vectors of the lane centerlines of the lanes other than the lane centerline of the vehicle's lane. An optional specific implementation method is: first, two multi-layer perceptrons (MLPs) are used to encode the features of the lane centerline set of the target lane centerline, the lane centerline of the lane where the vehicle is located, and the lane centerline of the lane where the vehicle is located. and other lane center lines To enhance the expressive power of features and the generalization ability of multi-layer perceptrons, the encoded features are mapped to a high-dimensional semantic space, represented as L sem =(n l ,n p ,128), where n l Indicates the number of lane centerlines, n pThe number of data points representing the center line of each lane is preferably set to 150. If the number of data points for a single lane center line is less than 150, it is padded with 0. If the number of data points exceeds 150, it is truncated from the front to the back. The last dimension represents the x and y coordinates of each data point. Before high-dimensional semantic feature mapping is performed, the dimension is 2, and after mapping, it is preferably 128.

[0072] Step S02: Perform dimensionality reduction processing on the multi-dimensional vector features and extract sequence information of the data points therein to obtain low-dimensional vector features.

[0073] In this embodiment, after the multidimensional feature vectors of the center lines of each lane are obtained in step S01, the obtained multidimensional vector features are further reduced in dimension and the sequence information of the data points therein is extracted, thereby obtaining the low-dimensional vector features corresponding to each multidimensional feature vector. An optional specific implementation method is: the multidimensional feature vector L mapped to the high-dimensional semantic space is mapped by an LSTM network (Long Short-Term Memory). sem The sequence information of the data points is extracted. In order to improve the processing efficiency, the dimensionality reduction processing is performed. The dimensionality reduction processing here is to integrate the data of the number dimension of the lane centerline data points and the sequence information of the extracted data points into the dimension representing the coordinate information of the data points, thereby obtaining the low-dimensional feature vector of the multi-dimensional feature vector, which is expressed as L dim =(n l ,128).

[0074] Step S03: perform inner product calculations on the low-dimensional vector features of the center line of the lane where the vehicle is located and the low-dimensional vector features of the center lines of each other lane to determine the positional relationship between the center lines of each other lane and the center line of the lane where the vehicle is located.

[0075] In this embodiment, after the dimensionality reduction processing is performed on each multi-dimensional feature vector obtained in step S02 to obtain the corresponding low-dimensional feature vectors, the inner product calculation is performed on the low-dimensional vector features of the center line of the lane where the vehicle is located and the low-dimensional vector features of the center lines of each other lane, and the position relationship between each of the center lines of the lane where the vehicle is located and the center line of the lane where the vehicle is located is obtained based on the calculation results. An optional specific implementation method is: first, the inner product between the low-dimensional vector features of the center line of the lane where the vehicle is located and the low-dimensional vector features of the center lines of each other lane is calculated to represent the spatial position between the lane center lines, which is expressed as I mul Then, the obtained spatial positions between each other are mapped to relative position relationships through the softmax multi-classification function, which is expressed as L rel The whole process from step S01 to step S03 is as follows:

[0076]

[0077] L rel =softmax(I mul )

[0078] In this embodiment, after the relative positional relationship between the lane centerline of the lane where the vehicle is located and the lane centerlines of other lanes is determined through the above steps S01 to S03, the determined lane centerline of the lane where the vehicle is located will be assigned a value of zero), the lane centerline of the lane on the left side of the vehicle will be assigned a negative number, specifically, the value becomes smaller as it goes to the left, and the lane centerline of the lane on the right side of the vehicle will be assigned a positive number, specifically, the value becomes larger as it goes to the right. In the present application, when the method further includes steps S01 to S03, after step S4, the method further includes: determining the positional relationship between the new lane centerline and the centerline of the lane where the vehicle is located based on the diverging flow road topology and the positional relationship.

[0079] In this embodiment, after the relative position relationship between each lane centerline in the target lane centerline set is obtained through steps S01 to S03, and the diverging and merging road topology between the predicted new lane centerline and the lane centerline in the target lane centerline set is obtained through steps S1 to S3, based on the obtained diverging and merging road topology, it is determined which lane centerline (such as L6) and the target lane centerline set (such as L7) form the diverging road topology (or merging road topology), and at this time, the lane centerline value of the lane centerline (L7) is assigned to the new lane centerline (L6), that is, the new lane centerline (L6) and the lane centerline (L7) will have the same lane centerline value. Figure 4 As shown, a diverging road topology appears in front of the lane where the vehicle is located, and accordingly the new lane centerline in the diverging road topology is assigned the same value of zero as the centerline of the lane where the vehicle is located.

[0080] In combination with the above embodiments, in one implementation, the embodiment of the present application further provides a method for online construction of local road topology. In the method for online construction of local road topology, before step S1, the method further includes steps S101 to S107:

[0081] Step S101: constructing a training data set, wherein the sample data in the training data set includes sample lane line data and a set of real lane center lines corresponding to the sample lane line data.

[0082] In this embodiment, before step S1, an initial lane centerline prediction network is first constructed and trained to obtain a qualified lane centerline prediction network. Specifically: first, a training data set is constructed, in which a sample data in the training data set includes sample lane line data and a set of real lane center lines corresponding to the sample lane line data.

[0083] Step S102: Based on the sample lane line data in the training data set, the lane center line is predicted by the initial lane center line prediction network to obtain a lane center line set.

[0084] In this embodiment, the sample lane line data in the training data set is input into the constructed initial lane centerline prediction network to perform lane centerline prediction, and the lane centerlines between all lane lines in the sample lane line data are obtained, and all the obtained lane centerlines constitute a lane centerline set.

[0085] Step S103: performing category matching on the lane centerline set and the corresponding real lane centerline set to obtain a target category matching result.

[0086] In this embodiment, each lane centerline in the lane centerline set is matched with each real lane centerline in the real lane centerline set in the sample data where the sample lane line data used to obtain the lane centerline set is located, and multiple matching results are obtained. Then, the matching result with the highest accuracy is determined by corresponding evaluation means, and the matching result with the highest accuracy is determined as the target category matching result. For example, the obtained lane centerline set includes lane centerlines L1 and L2, and the corresponding real lane centerline set includes real lane centerline sets Ln1 and Ln2. At this time, L1 is matched with Ln1 and L2 is matched with Ln2 to obtain a matching result 1, and L1 is matched with Ln2 and L2 is matched with Ln1 to obtain a matching result 2. Then, the matching result 2 is determined to have the highest accuracy by corresponding evaluation means, that is, the predicted lane centerline L1 corresponds to the real lane centerline Ln2 and the predicted lane centerline L2 corresponds to the real lane centerline Ln1. The possibility is the highest, and the matching result 2 is correspondingly determined as the target category matching result.

[0087] Step S104: performing point set matching on the lane center lines that match each other in the target category matching result and the real lane center line to obtain a target point set matching result.

[0088] In this embodiment, after the target category matching result is obtained through step S103, the mutually matching lane centerlines and the real lane centerlines are point set matched through corresponding evaluation means, that is, through corresponding evaluation means, the data points of the mutually matching lane centerlines and the real lane centerlines are matched with each other, and it is determined which data point in the real lane centerline each data point in the mutually matching lane centerlines corresponds to, so as to obtain the final target point set matching result. For example, continuing with the above example, the target category matching results are that L1 matches Ln2 and L2 matches Ln1, the lane centerline L1 includes data points p11 to p13, the lane centerline L2 includes data points p21 to p23, the lane centerline Ln1 includes data points pn11 to pn13, and the lane centerline Ln2 includes data points pn21 to pn23. Corresponding evaluation means are used to determine which data point in the lane centerline Ln2 the data point p11 in the lane centerline L1 corresponds to, and to determine which data point in the lane centerline Ln2 the data point p12 in the lane centerline L1 corresponds to, and to determine which data point in the lane centerline Ln2 the data point p13 in the lane centerline L1 corresponds to. At the same time, corresponding evaluation means are used to determine which data point in the lane center line Ln1 the data point p21 in the lane center line L2 corresponds to, and to determine which data point in the lane center line Ln1 the data point p22 in the lane center line L2 corresponds to, and to determine which data point in the lane center line Ln1 the data point p23 in the lane center line L2 corresponds to, and the matching results of these data points constitute the target point set matching results.

[0089] Step S105: Calculate the target category matching result and the target point set matching result obtained by using a loss function to obtain a loss value.

[0090] In this embodiment, after obtaining the target category matching result and the target point set matching result corresponding to the sample data through steps S101 to S104, the obtained target category matching result and the target point set matching result are calculated through the loss function to obtain the corresponding loss value, that is, the deviation between the target category matching result and the real category matching, and the deviation between the target point set matching result and the real point set matching are evaluated through the loss function. If the deviation between the target category matching result and the real category matching and the deviation between the target point set matching result and the real point set matching are relatively small, the corresponding loss value obtained will also be relatively small, which indicates that the prediction result of the lane centerline prediction network is relatively accurate.

[0091] Step S106: When the loss value satisfies a set condition, it is determined that the initial lane centerline prediction network is trained to be qualified, and a qualified lane centerline prediction network is obtained.

[0092] In this embodiment, the present application presets a loss threshold in advance. When the loss value obtained through step S105 is less than or equal to the loss threshold, it is determined that the loss value meets the set conditions, and the initial lane centerline prediction network is determined to be qualified for training, and a qualified lane centerline prediction network is obtained.

[0093] Step S107: When the loss value does not meet the set conditions, the model parameters of the initial lane centerline prediction network are updated and the model training is performed again.

[0094] In this embodiment, the present application presets a loss threshold in advance. When the loss value obtained through step S105 is greater than the loss threshold and it is determined that the loss value does not meet the set conditions, the model parameters of the initial lane centerline prediction network are adjusted and the model is retrained until the loss value after training meets the set conditions.

[0095] In combination with the above embodiments, in one implementation, the embodiment of the present application further provides a method for online construction of local road topology. In the method for online construction of local road topology, step S103 may include: class matching the lane centerline set and the corresponding real lane centerline set, and performing matching cost calculation on the obtained matching result through a class matching cost function to obtain a matching cost value corresponding to the matching result, and the class matching cost function is: in Indicates the prediction under the corresponding matching result and real v i The pairwise matching cost between v i ={c i ,P i}, c i , P i Represent the category label and point set of the real lane centerline respectively, and Respectively represent the corresponding matching results and the real v i Matching predictions The predicted category score and predicted point set are obtained; according to the matching cost values ​​corresponding to the obtained matching results, the matching result with the smallest matching cost value is determined as the target category matching result.

[0096] In this embodiment, each lane centerline in the predicted lane centerline set is matched one-to-one with each real lane centerline in the real lane centerline set in the sample data where the sample lane line data used to obtain the lane centerline set is located, and multiple matching results are obtained. The matching cost is calculated for each matching result obtained by the category matching cost function to obtain the matching cost value corresponding to each matching result. Then, the matching result with the smallest matching cost value is determined as the target category matching result. Among them, the category matching cost function is:

[0097] in Indicates the prediction of the lane centerline obtained under the corresponding matching result and the real v of the real lane centerline i The pairwise matching cost between

[0098] v i ={c i ,P i}, c i , P i Represent the category label of the real lane centerline and the point set of data points, respectively. and Respectively represent the corresponding matching results and the real v i Matching predictions The predicted class scores and the point set of predicted data points.

[0099] In combination with the above embodiments, in one implementation, the present application embodiment further provides a method for online construction of local road topology. In the method for online construction of local road topology, the matching cost is defined as in, is the cost term for category matching; Represents the position matching loss term of the point set.

[0100] In this embodiment, in order to improve the matching effect, the present application considers both the category matching and the data point matching in the category matching cost function. Specifically: the present application defines the matching cost in the category matching cost function as for in, is the cost term for category matching; Represents the position matching loss term of the point set.

[0101] In combination with the above embodiments, in one implementation, the embodiment of the present application further provides a method for online construction of local road topology. In the method for online construction of local road topology, step S3 may include: calculating the geometric distance between the end point of the lane centerline before the lateral truncation position and the starting point of the lane centerline after the lateral truncation position by a target algorithm, and obtaining a geometric distance matrix between the end point of the lane centerline before the lateral truncation position and the starting point of the lane centerline after the lateral truncation position, wherein the target algorithm is; D = {d ij |i,j=0,…n-1}, where D is the geometric distance matrix between lane centerlines, l i Indicates the lane centerline before the lateral cutoff position i, l j Indicates the lane centerline j after the lateral cutoff position, d ij Indicates the lane centerline l i and l j The geometric distance between Indicates the lane centerline l i The end point, Indicates the lane centerline l j The starting point is: the geometric distance matrix is ​​calculated by a mapping function to obtain the corresponding calculation result, and the mapping function is

[0102] where x = d ij , σ is the standard deviation of the geometric distance matrix D, and σ and λ are learnable parameters.

[0103] In this embodiment, the geometric distances between the end points of the lane centerline before each lateral truncation position and the starting points of the lane centerline after each lateral truncation position are calculated by the target algorithm to obtain the geometric distance matrix between the end points of the lane centerline before the lateral truncation position and the starting points of the lane centerline after the lateral truncation position. The target algorithm is: D = {d ij |i,j=0,…n-1}, where D is the geometric distance matrix between lane centerlines, l i Indicates the lane centerline before the lateral cutoff position i, l j Indicates the lane centerline j after the lateral cutoff position, d ij Indicates the lane centerline l i and l j The geometric distance between Indicates the lane centerline l i The data point at the end point, Indicates the lane centerline l jAfter obtaining the geometric distance matrix between the center lines of the lanes before and after the lateral truncation position, the mapping function is used to calculate each geometric distance in the geometric distance matrix to obtain the corresponding calculation result. The calculation result includes the mapping value corresponding to each geometric distance in the geometric distance matrix. When the input geometric distance value is 0 or close to 0, the corresponding mapping value is 1 or close to 1, indicating that the two lane center lines corresponding to the geometric distance are connected end to end. Among them, the mapping function is where x = d ij , σ is the standard deviation of the geometric distance matrix D, and σ and λ are learnable parameters.

[0104] In combination with the above embodiments, in one implementation, the embodiment of the present application further provides a method for online construction of a local road topology. In the method for online construction of a local road topology, according to the calculation result, the diverging and converging road topology between the lane center lines is determined, including: in the calculation result, when the mapping value between the end point of a lane center line before the transverse cutoff position and the starting point of at least two lane center lines after the transverse cutoff position satisfies the first condition, the corresponding position is determined to be a diverging road topology; in the calculation result, when the mapping value between the starting point of a lane center line after the transverse cutoff position and the end point of at least two lane center lines before the transverse cutoff position satisfies the first condition, the corresponding position is determined to be a converging road topology.

[0105] In this embodiment, based on the obtained calculation results, it is determined whether there is a mapping value of 1 or approximately approaching 1 between the end point of a lane centerline (such as L01) before the transverse truncation position and the starting point of at least two lane centerlines after the transverse truncation position. If so, it is determined that the first condition is satisfied, and it is determined that there is a diverging road topology after the lane centerline (L01). Based on the obtained calculation results, it is determined whether there is a mapping value of 1 or approximately approaching 1 between the starting point of a lane centerline (such as L02) after the transverse truncation position and the end point of at least two lane centerlines before the transverse truncation position. If so, it is determined that the first condition is satisfied, and it is determined that there is a converging road topology before the lane centerline (L02).

[0106] In combination with the above embodiments, in one implementation, the embodiment of the present application also provides a method for online construction of local road topology. In the method for online construction of local road topology, the method also includes: constructing a first loss function related to lane centerline prediction, and constructing a second loss function related to the determination of the positional relationship between lane centerlines, and constructing a third loss function related to the determination of the diverging and converging road topology between lane centerlines; determining a target loss function based on the constructed first loss function, second loss function, and third loss function; guiding the training of a target neural network through the target loss function, the target neural network is a neural network used in the process of lane centerline prediction, determination of the positional relationship between lane centerlines, and determination of the diverging and converging road topology between lane centerlines.

[0107] In this embodiment, for the deep neural networks involved in the process of lane centerline prediction, positional relationship determination between lane centerlines, and diverging and converging road topology determination between lane centerlines in a local road topology online construction method provided by this application, a total target loss function is constructed to guide the training of each deep neural network involved. Specifically: for lane centerline prediction, a first loss function related to lane centerline prediction is constructed; for positional relationship determination between lane centerlines, a second loss function related to positional relationship determination between lane centerlines is constructed; for diverging and converging road topology determination between lane centerlines, a third loss function related to diverging and converging road topology determination between lane centerlines is constructed; then the constructed first loss function, second loss function, and third loss function are combined into a target loss function; the target loss function is used to guide the training of the target neural network, which is the neural network used in the process of lane centerline prediction, positional relationship determination between lane centerlines, and diverging and converging road topology determination between lane centerlines.

[0108] In combination with the above embodiments, in one implementation, the embodiment of the present application also provides a method for online construction of local road topology. In the method for online construction of local road topology, the method further includes: constructing a lane line detection loss function; the target loss function is determined based on the constructed first loss function, second loss function, and third loss function, including: determining the target loss function based on the constructed first loss function, second loss function, third loss function, and the lane line detection loss function, wherein the target function is in, is the first loss function, is the second loss function, is the third loss function, is the lane detection loss function, λ1, λ2, and λ3 are learnable parameters;

[0109] In this embodiment, since the corresponding neural network is also involved in determining the lane line instance, the present application also constructs a lane line detection loss function, which is combined with the constructed first loss function, second loss function, and third loss function to form a target loss function. The target function is expressed as in, is the first loss function, is the second loss function, is the third loss function, is the lane detection loss function, λ1, λ2, and λ3 are learnable parameters;

[0110]

[0111] In this embodiment, if Figure 5 and Figure 6 As shown, first, the lane line in front of the vehicle is detected by the surround view image collected by the vehicle's existing lane line detection module. Specifically, the collected surround view image is converted to the BEV perspective to obtain the feature map under the BEV perspective, and then the lane line detection module of the vehicle performs lane line detection processing on the feature map under the BEV perspective to obtain the lane line. Then, the lane center line is predicted by the trained qualified lane center line prediction network to obtain all lane center lines in the current scene. Based on the obtained lane center lines, the relative position relationship between the lane center lines and the geometric distance between the starting point and the end point are calculated, so as to obtain the final diverging and converging road topology between the lane center lines.

[0112] Based on the same inventive concept, an embodiment of the present application provides a local road topology online construction system, such as Figure 7 As shown, the system 700 includes:

[0113] A lane centerline prediction module 701 is used to predict the lane centerline based on the lane line data by using a trained lane centerline prediction network to obtain a target lane centerline set, where the lane centerline consists of a series of ordered data points;

[0114] A truncation module 702 is used to perform transverse truncation of all lane centerlines using the endpoints of the new lane centerline when a new lane centerline is predicted;

[0115] A geometric distance calculation module 703 is used to calculate the geometric distance between the end point of the lane centerline before the transverse cutoff position and the start point of the lane centerline after the transverse cutoff position to obtain a corresponding calculation result;

[0116] The road topology determination module 704 is used to determine the diverging and merging road topology between lane center lines according to the calculation results.

[0117] Optionally, the system 700 further includes:

[0118] The encoding module is used to encode the center line of the lane where the vehicle is located and other lane center lines in the target lane center line set through different multi-layer perceptrons and map them to a high-dimensional semantic space to obtain the multi-dimensional vector features of each lane center line;

[0119] Dimensionality reduction module, used to reduce the dimensionality of multi-dimensional vector features and extract the sequence information of data points therein to obtain low-dimensional vector features;

[0120] An inner product calculation module is used to perform inner product calculations on the low-dimensional vector features of the center line of the lane where the vehicle is located and the low-dimensional vector features of the center lines of each other lane, respectively, to determine the positional relationship between the center lines of each other lane and the center line of the lane where the vehicle is located;

[0121] The position relationship determination module is used to determine the position relationship between the center line of the new lane and the center line of the lane where the vehicle is located according to the diverging and merging road topology and the position relationship.

[0122] Optionally, the system 700 further includes:

[0123] A data set building module, used to build a training data set, wherein the sample data in the training data set includes sample lane line data and a set of real lane center lines corresponding to the sample lane line data;

[0124] A prediction module, used to predict lane center lines through an initial lane center line prediction network based on sample lane line data in a training data set, and obtain a lane center line set;

[0125] A category matching module, used to perform category matching on the lane centerline set and the corresponding real lane centerline set to obtain a target category matching result;

[0126] A point set matching module is used to perform point set matching on the lane center lines that match each other in the target category matching result and the real lane center line to obtain a target point set matching result;

[0127] A loss calculation module, used to calculate the obtained target category matching result and the obtained target point set matching result through a loss function to obtain a loss value;

[0128] A first result determination module is used to determine that the initial lane centerline prediction network is qualified when the loss value meets the set conditions, and obtain a qualified lane centerline prediction network;

[0129] The second result determination module is used to update the model parameters of the initial lane centerline prediction network and re-train the model when the loss value does not meet the set conditions.

[0130] Optional, category matching module, including:

[0131] The first category matching module is used to perform category matching on the lane centerline set and the corresponding real lane centerline set, and calculate the matching cost of the obtained matching result through a category matching cost function to obtain a matching cost value corresponding to the matching result, wherein the category matching cost function is: in Indicates the prediction under the corresponding matching result and real v i The pairwise matching cost between v i ={c i ,P i},c i , P i Represent the category label and point set of the real lane centerline respectively, and Respectively represent the corresponding matching results and the real v i Matching predictions The predicted category scores and predicted point sets of

[0132] The category matching result determination module is used to determine the matching result with the smallest matching cost value as the target category matching result according to the matching cost values ​​corresponding to the obtained matching results.

[0133] Optionally, the matching cost in the category matching cost function in the system 700 is defined as in, is the cost term for category matching; Represents the position matching loss term of the point set.

[0134] Optional, geometric distance calculation module, including:

[0135] The first calculation module is used to calculate the geometric distance between the end point of the lane centerline before the lateral truncation position and the starting point of the lane centerline after the lateral truncation position by a target algorithm to obtain a geometric distance matrix between the end point of the lane centerline before the lateral truncation position and the starting point of the lane centerline after the lateral truncation position, wherein the target algorithm is:

[0136] D={d ij |i,j=0,…n-1}, where D is the geometric distance matrix between lane centerlines, l i Indicates the lane centerline before the lateral cutoff position i, l j Indicates the lane centerline j after the lateral cutoff position, d ij Indicates the lane centerline l i and l j The geometric distance between Indicates the lane centerline l i The end point, Indicates the lane centerline l j Starting point;

[0137] A mapping module is used to calculate the geometric distance matrix through a mapping function to obtain a corresponding calculation result, wherein the mapping function is where x = d ij , σ is the standard deviation of the geometric distance matrix D, and σ and λ are learnable parameters.

[0138] Optionally, a road topology determination module includes:

[0139] A first road topology determination module is used to determine that the corresponding position is a diverging road topology when, in the calculation result, a mapping value between an end point of a lane centerline before the transverse truncation position and a starting point of at least two lane centerlines after the transverse truncation position satisfies a first condition;

[0140] The second road topology determination module is used to determine that the corresponding position is a converging road topology when, in the calculation result, a mapping value between a starting point of a lane centerline after the lateral truncation position and an end point of at least two lane centerlines before the lateral truncation position satisfies a first condition.

[0141] Optionally, the system 700 further includes:

[0142] A loss function determination module, for constructing a first loss function related to lane centerline prediction, a second loss function related to positional relationship determination between lane centerlines, and a third loss function related to diverging and converging road topology determination between lane centerlines;

[0143] A target loss function determination module, used to determine the target loss function based on the constructed first loss function, second loss function and third loss function;

[0144] A training module is used to guide the training of a target neural network through the target loss function, wherein the target neural network is a neural network used in the process of predicting lane center lines, determining the positional relationship between lane center lines, and determining the diverging and merging road topology between lane center lines.

[0145] Optionally, the system 700 further includes:

[0146] Loss function construction module, used to construct lane detection loss function;

[0147] The target loss function determination module is used to determine the target loss function based on the constructed first loss function, the second loss function, the third loss function and the lane line detection loss function, wherein the target function is in, is the first loss function, is the second loss function, is the third loss function, is the lane detection loss function, λ1, λ2, and λ3 are learnable parameters;

[0148] Based on the same inventive concept, an embodiment of the present application provides an electronic device, including: a processor, a memory, and a computer program stored in the memory and running on the processor. When the computer program is executed by the processor, it implements the steps in a method for online construction of a local road topology as described in the first aspect of the present application.

[0149] Based on the same inventive concept, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps in a method for online construction of a local road topology as described in the first aspect of the present application are implemented.

[0150] As for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0151] It should be noted that, for the method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present application are not limited by the described order of actions, because according to the embodiments of the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present application.

[0152] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0153] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the embodiments of the present application may adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the embodiments of the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0154] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0155] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0156] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0157] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present application.

[0158] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.

[0159] The above is a detailed introduction to a local road topology online construction method, system and product provided by the present application. This article uses specific examples to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for online construction of local road topology, characterized in that: The method comprises: Based on the lane line data, the lane center line is predicted by training a qualified lane center line prediction network to obtain a target lane center line set. The lane center line consists of a series of ordered data points. When a new lane centerline is predicted, all lane centerlines are transversely cut off using the endpoints of the new lane centerline; Calculating a geometric distance between an end point of the lane centerline before the lateral cutoff position and a start point of the lane centerline after the lateral cutoff position to obtain a corresponding calculation result; According to the calculation results, the diverging and merging road topology between the lane center lines is determined.

2. A method for online construction of local road topology according to claim 1, characterized in that: In the case where a new lane center line is predicted, before all lane center lines are transversely truncated with the endpoint of the new lane center line, the method further includes: Different multi-layer perceptrons are used to encode the centerline of the lane where the vehicle is located and other lane centerlines in the target lane centerline set and map them to a high-dimensional semantic space to obtain the multi-dimensional vector features of each lane centerline. Perform dimensionality reduction processing on multi-dimensional vector features and extract the sequence information of data points to obtain low-dimensional vector features; The inner product of the low-dimensional vector features of the center line of the lane where the vehicle is located is calculated with the low-dimensional vector features of the center lines of each other lane to determine the positional relationship between the center lines of each other lane and the center line of the lane where the vehicle is located; After determining the diverging flow road topology between lane center lines according to the calculation results, the method further includes: determining the positional relationship between the new lane center line and the center line of the lane where the vehicle is located according to the diverging flow road topology and the positional relationship.

3. A method for online construction of local road topology according to claim 1, characterized in that: Before predicting the lane centerline based on the lane line data by training a qualified lane centerline prediction network to obtain a target lane centerline set, the method further includes: Constructing a training data set, wherein the sample data in the training data set includes sample lane line data and a set of real lane center lines corresponding to the sample lane line data; Based on the sample lane line data in the training data set, the lane center line is predicted by the initial lane center line prediction network to obtain a lane center line set; Performing category matching on the lane centerline set and the corresponding real lane centerline set to obtain a target category matching result; Performing point set matching on the lane center lines that match each other in the target category matching result and the real lane center line to obtain a target point set matching result; The target category matching result and the target point set matching result are calculated by using a loss function to obtain a loss value; When the loss value satisfies a set condition, determining that the initial lane centerline prediction network is trained to be qualified, and obtaining a trained lane centerline prediction network; When the loss value does not meet the set conditions, the model parameters of the initial lane centerline prediction network are updated and the model training is performed again.

4. A method for online construction of local road topology according to claim 3, characterized in that: Performing category matching on the lane centerline set and the corresponding real lane centerline set to obtain a target category matching result, including: The lane centerline set and the corresponding real lane centerline set are matched by category, and the matching cost is calculated for the obtained matching result by using the category matching cost function to obtain the matching cost value corresponding to the matching result. The category matching cost function is: in Indicates the prediction under the corresponding matching result and real v i The pairwise matching cost between v i ={c i ,P i }, c i , P i Represent the category label and point set of the real lane centerline respectively, and Respectively represent the corresponding matching results and the real v i Matching predictions The predicted category scores and predicted point sets of According to the matching cost values ​​corresponding to the obtained matching results, the matching result with the smallest matching cost value is determined as the target category matching result.

5. A method for online construction of local road topology according to claim 4, characterized in that: Define the matching cost as in, is the cost term for category matching; Represents the position matching loss term of the point set.

6. A method for online construction of local road topology according to claim 3, characterized in that: The geometric distance between the end point of the lane centerline before the transverse cutoff position and the start point of the lane centerline after the transverse cutoff position is calculated to obtain corresponding calculation results, including: The geometric distance between the end point of the lane centerline before the lateral truncation position and the starting point of the lane centerline after the lateral truncation position is calculated by the target algorithm to obtain a geometric distance matrix between the end point of the lane centerline before the lateral truncation position and the starting point of the lane centerline after the lateral truncation position. The target algorithm is: D = {d ij |i,j=0,…n-1}, where D is the geometric distance matrix between lane centerlines, l i Indicates the lane centerline before the lateral cutoff position i, l j Indicates the lane centerline j after the lateral cutoff position, d ij Indicates the lane centerline l i and l j The geometric distance between Indicates the lane centerline l i The end point, Indicates the lane centerline l j Starting point; The geometric distance matrix is ​​calculated by a mapping function to obtain a corresponding calculation result. The mapping function is: where x = d ij , σ is the standard deviation of the geometric distance matrix D, and σ and λ are learnable parameters.

7. A method for online construction of local road topology according to claim 6, characterized in that: According to the calculation results, the diverging and merging road topology between the lane center lines is determined, including: In the calculation result, when the mapping values ​​between the end point of a lane centerline before the lateral truncation position and the starting points of at least two lane centerlines after the lateral truncation position meet the first condition, the corresponding position is determined to be a diverging road topology; In the calculation result, when the mapping value between the starting point of a lane centerline after the lateral truncation position and the end points of at least two lane centerlines before the lateral truncation position satisfies the first condition, the corresponding position is determined to be a merging road topology.

8. A method for online construction of local road topology according to claim 7, characterized in that: The method further comprises: Constructing a first loss function related to lane centerline prediction, constructing a second loss function related to determining a positional relationship between lane centerlines, and constructing a third loss function related to determining a diverging and converging road topology between lane centerlines; Determine a target loss function based on the constructed first loss function, second loss function, and third loss function; The target loss function is used to guide the training of a target neural network, which is a neural network used in the process of predicting lane center lines, determining the positional relationship between lane center lines, and determining the diverging and merging road topology between lane center lines.

9. A method for online construction of local road topology according to claim 7, characterized in that: The method further comprises: Construct lane detection loss function; The determining of the target loss function based on the constructed first loss function, second loss function and third loss function includes: Based on the constructed first loss function, second loss function, third loss function and the lane line detection loss function, a target loss function is determined, and the target function is: in, is the first loss function, is the second loss function, is the third loss function, is the lane detection loss function, λ1, λ2, and λ3 are learnable parameters; 10. A local road topology online construction system, characterized in that: The system comprises: The lane centerline prediction module is used to predict the lane centerline based on the lane line data through a trained lane centerline prediction network to obtain a target lane centerline set. The lane centerline consists of a series of ordered data points. A truncation module, configured to perform transverse truncation of all lane centerlines using the endpoints of the new lane centerline when a new lane centerline is predicted; A geometric distance calculation module is used to calculate the geometric distance between the end point of the lane centerline before the transverse cutoff position and the start point of the lane centerline after the transverse cutoff position to obtain a corresponding calculation result; The road topology determination module is used to determine the diverging and merging road topology between lane center lines according to the calculation results.

11. An electronic device, characterized in that: include: A processor, a memory, and a computer program stored in the memory and running on the processor, wherein when the computer program is executed by the processor, the steps in the method for online construction of a local road topology as described in claims 1 to 9 are implemented.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the method for online construction of local road topology as claimed in claims 1 to 9 are implemented.

Images