Intersection guide line generation network training method and intersection guide line generation method

Abstract

The application relates to a kind of intersection guide line generation network training method and intersection guide line generation method, network model includes LSTM-E layer, pooling layer, first fully connected layer, second fully connected layer, LSTM-D layer;For each training sample, the semantic feature vector of the best passing trajectory and non-guide line trajectory is extracted using LSTM-E layer respectively, the semantic feature vector of multiple non-guide line trajectories is fused and processed using the pooling layer;The semantic feature vector of the fused non-guide line trajectory and the semantic feature vector of the best passing trajectory are used as the input of the first fully connected layer, the mean mu and the variance sigma of the distribution to which the trajectory data in the training sample is subjected are predicted;Random vector epsilon is sampled from standard normal distribution, and vector z=epsilon x mu+sigma is calculated;The semantic feature vector of the fused non-guide line trajectory and vector z are used as the input of the second fully connected layer, and the output of the second fully connected layer is used as the input of the LSTM-D layer, to predict and generate the coordinate point constituting the intersection guide line trajectory one by one.

Description

Technical Field

[0001] This invention relates to the field of high-precision map production technology, specifically to a method for training a network for generating intersection guide lines based on trajectory feature fusion and a method for generating intersection guide lines. Background Technology

[0002] In crowdsourced high-precision map creation, intersection guide lines are a key element of the map. They are a trajectory that conforms to driving behavior and is used to guide intelligent driving vehicles to pass through intersections comfortably.

[0003] There are two main types of methods for generating intersection guide lines: 1) Kinematic model-based methods. This method assumes that road traffic elements such as POIs in the map are ideal and complete. It pre-defines the objective function, designs optimization rules and constraints such as collision avoidance and speed, and generates smooth guide lines through mathematical methods. However, it does not consider the uncertainty of road element perception in high-precision map mapping, and the lack of complete real road conditions and traffic information, resulting in the practicality of guide lines generated by mathematical methods; 2) Trajectory data-driven machine learning methods. This method, mainly represented by deep learning, is primarily applied to trajectory prediction. It uses the trajectory of the previous road segment to predict the vehicle's trajectory in the next segment, generating dynamic guide lines for vehicles in motion. In practical implementation, this method needs to consider the interaction information between the autonomous vehicle and surrounding traffic entities, as well as accurate high-precision map information. However, when both of these aspects are lacking, and only historical trajectories are available, this method is difficult to implement. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a method and apparatus for generating intersection guide lines based on trajectory feature fusion, thus solving the problem of generating static guide lines at road intersections in crowdsourced high-precision map production when only historical trajectories are available.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] On one hand, this invention provides a training method for an intersection guide line generation network based on trajectory feature fusion. The network model includes an LSTM-E layer, a pooling layer, a first fully connected layer, a second fully connected layer, and an LSTM-D layer. The training method includes:

[0007] Obtain a training sample dataset, wherein each sample in the training sample dataset includes an optimal travel trajectory and multiple non-guideline trajectories in the same lane or the same virtual lane as the optimal travel trajectory;

[0008] For each training sample, the semantic feature vectors of the best passage trajectory and non-guide line trajectories are extracted using the LSTM-E layer, and the semantic feature vectors of multiple non-guide line trajectories are fused using the pooling layer.

[0009] The semantic feature vectors of the fused non-guided trajectory and the semantic feature vector of the optimal travel trajectory are used as inputs to the first fully connected layer to predict the mean μ and variance σ of the distribution of trajectory data in the training samples.

[0010] Sample a random vector ε from the standard normal distribution and calculate the vector z = ε × σ + μ;

[0011] The semantic feature vector and vector z of the fused non-guide line trajectory are used as inputs to the second fully connected layer, and the output of the second fully connected layer is used as inputs to the LSTM-D layer to predict and generate the coordinate points that constitute the guide line trajectory of the intersection one by one.

[0012] Furthermore, the method also includes performing coordinate transformation on the trajectories in the training sample dataset, converting the Earth coordinate values ​​of the coordinate points in the trajectory into local coordinate values ​​that conform to the model calculation.

[0013] Furthermore, the coordinate transformation includes:

[0014] Transform the global coordinate system into a Cartesian coordinate system with the same orientation as the global coordinate system, or

[0015] Establish a Cartesian coordinate system with the direction of vehicle entry into the intersection as the positive y-axis and the right-hand side of the y-axis as the positive x-axis. Convert the global coordinate system to a Cartesian coordinate system, or

[0016] The global coordinates of the current point are subtracted from the global coordinates of the previous point in the trajectory to obtain the local coordinates of the current point after coordinate transformation; if the current point is the first point in the trajectory, the global coordinates of the current point are subtracted from itself.

[0017] Furthermore, the pooling layer employs an AveragePooling, MaxPooling, or AttentionPooling fusion strategy to fuse the semantic feature vectors of multiple non-guided line trajectories.

[0018] On the other hand, this invention provides a method for generating intersection guide lines based on trajectory feature fusion. This method is implemented using a network structure trained using the aforementioned training method, and includes:

[0019] Historical trajectory data is acquired and used as input to the LSTM-E layer to extract semantic feature vectors from the trajectory data. Pooling layers are then used to fuse the semantic feature vectors of the historical trajectory data.

[0020] Sample a random vector ε from a standard normal distribution;

[0021] The semantic feature vector and random vector ε of the fused historical trajectory data are used as the input of the second fully connected layer, and the output of the second fully connected layer is used as the input of the LSTM-D layer to predict and generate the coordinate points that constitute the trajectory of the intersection guide line one by one.

[0022] The beneficial effects of this invention are: this invention does not require tag information, interaction information of other entities, or map information, but directly mines semantic knowledge vectors of driving behavior from historical trajectories, and can generate multiple trajectories that conform to traffic trajectories and driving behavior guidance lines. Attached Figure Description

[0023] Figure 1 A schematic diagram illustrating the training process of the guide line generation network based on trajectory feature fusion, provided in an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of the inference process of the guide line generation network based on trajectory feature fusion provided in an embodiment of the present invention. Detailed Implementation

[0025] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0026] Figure 1 This is a schematic diagram illustrating the training process of a guide line generation network based on trajectory feature fusion, provided in an embodiment of the present invention. Figure 1 As shown, the network model used in this embodiment of the invention includes an LSTM-E layer, a pooling module, a first fully connected layer FC1, a second fully connected layer FC2, and an LSTM-D layer.

[0027] In the picture Indicates the first A trajectory, represented in the form of ,in Indicates the first The longitude of the first point of the trajectory, Indicates the first The latitude of the first point of the trajectory, where n is the number of points in the trajectory. The label value represents the optimal travel trajectory for a specific lane at this intersection, expressed as: . The generated guide line trajectory is represented in the following form: .

[0028] LSTM-E stands for LSTM-based trajectory feature extraction component, used to extract temporal features of trajectories. Indicates the first The temporal feature vector of the trajectory.

[0029] The training process includes the following steps:

[0030] S1. Obtain the trajectory dataset and divide the dataset into a training sample dataset and a test sample dataset in an 8:2 ratio. Each sample in the training sample dataset includes an optimal driving trajectory and multiple non-guide line trajectories in the same lane or the same virtual lane as the optimal driving trajectory.

[0031] In this embodiment, the trajectory is represented as a sequence of trajectory points. The coordinate values ​​of each trajectory point include longitude and latitude. However, the longitude and latitude in the original trajectory are Earth coordinate values, which cannot be used as input values ​​for model calculation. Therefore, it is necessary to perform coordinate transformation on the trajectories in the training sample dataset, converting the Earth coordinate values ​​of the trajectory points in the trajectory into local coordinate values ​​that are suitable for model calculation.

[0032] The specific coordinate transformation method can be one of the following three methods:

[0033] 1) Transform the global coordinate system into a Cartesian coordinate system with the same orientation as the global coordinate system.

[0034] The formula is as follows:

[0035]

[0036] in, , These represent longitude and latitude values ​​respectively, with superscripts. , Representing local and global coordinates respectively, subscripts Indicates the first One sampling point.

[0037] 2) Establish a Cartesian coordinate system with the direction of vehicle entry into the intersection as the positive y-axis and the right-hand side of the y-axis as the positive x-axis, and transform the global coordinate system into a Cartesian coordinate system.

[0038] This method requires coordinate rotation. First, convert the global coordinates to local coordinates using the method described in 1), then rotate the coordinates. The formula is as follows:

[0039]

[0040] in, , These represent the longitude and latitude of the new local coordinates, respectively. , These represent the longitude and latitude of the old local coordinates, respectively.

[0041] 3) Subtract the global coordinates of the previous point from the global coordinates of the current point in the trajectory to obtain the local coordinates of the current point after coordinate transformation; the formula is as follows:

[0042]

[0043] If the current coordinate point is the first point in the trajectory, then subtract the current coordinate point from its own global coordinate value.

[0044] S2. For each training sample, the semantic feature vectors of the best passing trajectory and non-guided line trajectories are extracted using the LSTM-E layer, and the semantic feature vectors of multiple non-guided line trajectories are fused using the Pooling Module.

[0045] Trajectory data consists of a sequence of trajectory points and is time-series data. Considering the sequence length, an LSTM model is used to extract trajectory features. Figure 1 The LSTM-E shown can have 32, 64, 128 neurons, etc. The number of layers in the LSTM-E is a hyperparameter, which is tuned empirically. The semantic feature vector of the trajectory is finally obtained by using the neuron outputs of each iteration step of the LSTM-E. Figure 1 As shown, The semantic feature vector that guides the trajectory of the target line. For the first in the same lane The semantic feature vector of a non-guided line trajectory.

[0046] The Pooling Module employs a fusion strategy of AveragePooling, MaxPooling, or AttentionPooling to fuse semantic feature vectors from multiple non-guided line trajectories. MaxPooling calculates the maximum value in each dimension of the semantic vector. AveragePooling calculates the mean in each dimension of the semantic vector. AttentionPooling uses a self-attention mechanism to calculate the fused features. The kernel sizes for MaxPooling and AveragePooling are empirical values ​​of 3 and 5. The number of neurons in the fully connected layer of AttentionPooling is empirical values ​​of 32 and 64.

[0047] S3 takes the semantic feature vector of the fused non-guided trajectory and the semantic feature vector of the optimal travel trajectory as input to the first fully connected layer FC1 to predict the mean μ and variance σ of the distribution of trajectory data in the training samples.

[0048] S4, sample a random vector ε from the standard normal distribution and calculate z = ε × σ + μ.

[0049] To enable the model to calculate the gradient in reverse, a resampling method is used. First, a random vector ε~N(0,1) is randomly sampled from the standard normal distribution. The dimension can be 32, 64, etc., and then z=ε×σ+μ is calculated.

[0050] S5 takes the semantic feature vector of the fused non-guided line trajectory and the vector z as the input of the second fully connected layer FC2, and fuses multiple semantic feature vectors into a new feature vector to achieve feature crossing and fusion.

[0051] The output of the second fully connected layer serves as the input to the LSTM-D layer, predicting and generating coordinate points that constitute the trajectory of the intersection guide lines one by one. The generation of the next point depends on the previous point and the semantic feature vector.

[0052] The second fully connected layer FC2 can have multiple layers, with 128, 256, 32, 64, 128, etc. The number of layers in LSTM-D is a hyperparameter, and the parameters are tuned based on experience.

[0053] After training, the network model needs to be validated through inference. The inference process is as follows: Figure 2 As shown, specifically, it includes the following steps:

[0054] 1. Input the test data into the model, ensuring that each sample has the same format as the samples in the training data.

[0055] 2. Extract the semantic feature vector of the trajectory using the LSTM-E module. The trajectory feature extraction and training processes are consistent.

[0056] 3. The Pooling Module is used to fuse multiple guide line trajectories to obtain the fused semantic feature vector c.

[0057] 4. The random vector z sampled from the standard normal distribution is consistent with the training process.

[0058] 5. Combine z and c through an FC layer to extract a higher-level feature vector, which is then used as the input to LSTM-D.

[0059] 6. LSTM-D decodes the input vector into points that generate the guide line trajectory.

[0060] Based on the network structure trained using the above-described training method, this embodiment of the invention also provides a method for generating intersection guide lines based on trajectory feature fusion, comprising:

[0061] Historical trajectory data is acquired and used as input to the LSTM-E layer to extract semantic feature vectors from the trajectory data. Pooling layers are then used to fuse the semantic feature vectors of the historical trajectory data.

[0062] Sample a random vector ε from a standard normal distribution;

[0063] The semantic feature vector and random vector ε of the fused historical trajectory data are used as the input of the second fully connected layer, and the output of the second fully connected layer is used as the input of the LSTM-D layer to predict and generate the coordinate points that constitute the trajectory of the intersection guide line one by one.

[0064] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0065] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A trajectory feature fusion based intersection guide line generation network training method, characterized in that, The network model comprises an LSTM-E layer, a pooling layer, a first full connection layer, a second full connection layer, and an LSTM-D layer; the LSTM-E layer represents a trajectory feature extraction component based on an LSTM model, and is configured to extract time sequence features of a trajectory; The training method comprises: obtaining a training sample data set, wherein each sample in the training sample data set comprises one optimal passing trajectory and a plurality of non-guide-line trajectories in the same lane or the same virtual lane as the optimal passing trajectory; for each training sample, using the LSTM-E layer to extract semantic feature vectors of the optimal passing trajectory and the non-guide-line trajectories respectively, and using the pooling layer to fuse the semantic feature vectors of the plurality of non-guide-line trajectories; using the fused semantic feature vectors of the non-guide-line trajectories and the semantic feature vector of the optimal passing trajectory as inputs of the first full connection layer, to predict a mean value μ and a variance σ of a distribution to which trajectory data in the training sample conforms; sampling a random vector ε from a standard normal distribution, and calculating a vector z = ε × σ + μ; using the fused semantic feature vectors of the non-guide-line trajectories and the vector z as inputs of the second full connection layer, and using an output of the second full connection layer as an input of the LSTM-D layer, to predict and generate coordinate points constituting a guide-line trajectory of a road intersection one by one; the LSTM-D layer is configured to decode the input vector into each point of the generated guide-line trajectory.

2. The method of claim 1, wherein, Further comprising: performing coordinate conversion on the trajectories in the training sample data set, to convert earth coordinate values of coordinate points in the trajectories into local coordinate values conforming to model calculation.

3. The method of claim 2, wherein, The coordinate conversion comprises: converting a global coordinate system into a Cartesian coordinate system in the same direction as the global coordinate system, or establishing a Cartesian coordinate system with a direction in which a vehicle enters a road intersection as a positive direction of a y-axis and a right-hand side of the y-axis as a positive direction of an x-axis, and converting the global coordinate system into the Cartesian coordinate system, or subtracting a global coordinate value of a previous coordinate point from a global coordinate value of a current coordinate point in a trajectory to obtain a local coordinate value of the current coordinate point after coordinate conversion; if the current coordinate point is the first point in the trajectory, subtracting the global coordinate value of the current coordinate point from itself.

4. The method of claim 1, wherein, The pooling layer adopts an AveragePooling, MaxPooling or AttentionPooling fusion strategy to fuse the semantic feature vectors of the plurality of non-guide-line trajectories.

5. A method for intersection guide line generation based on trajectory feature fusion, characterized in that, The method is implemented by using a network structure trained by the training method according to any one of claims 1 to 4, and comprises: obtaining historical trajectory data and using the historical trajectory data as inputs of the LSTM-E layer to extract semantic feature vectors of the trajectory data, and using the pooling layer to fuse the semantic feature vectors of the historical trajectory data; sampling a random vector ε from a standard normal distribution; using the fused semantic feature vectors of the historical trajectory data and the random vector ε as inputs of the second full connection layer, and using an output of the second full connection layer as an input of the LSTM-D layer, to predict and generate coordinate points constituting a guide-line trajectory of a road intersection one by one.

Images