A vehicle trajectory prediction method and system considering multiple target points

Abstract

The present invention discloses a vehicle trajectory prediction method and system considering multiple target points, including a vehicle feature and lane centerline feature extraction network, a feature fusion network, a target point prediction network and a feature decoding network, comprising the following steps: S1. extracting the historical trajectory features and lane centerline features of all vehicles in the current traffic scene respectively; S2. propagating and updating the lane centerline features as environmental constraints, adopting a spatial cross-attention mechanism to fuse them with the vehicle historical trajectory features, and then adopting a spatial self-attention mechanism to realize constrained vehicle-to-vehicle interaction; S3. predicting the target points of each vehicle respectively, including the midpoint and multiple end points, and assigning the lane environment features near the midpoint and the end point with the highest confidence to the vehicle features; S4. decoding and outputting the predicted multiple trajectories and confidences, and performing smoothing processing; the present invention increases the accuracy of the trajectory prediction results and retains multiple possible predictions for future predictions.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and more particularly to a vehicle trajectory prediction method and system considering multiple target points. Background Art

[0002] Currently, autonomous driving systems include modules such as perception, decision-making and planning, and control. Direct decision-making and planning require extremely high real-time processing capabilities, requiring timely adjustments to driving strategies based on current environmental conditions. However, in reality, using perception as input, followed by data processing and control, introduces time lags. Furthermore, without considering the plans of other vehicles, each vehicle focuses solely on its own local optimal solution. This non-globally optimal decision-making and planning does not improve traffic conditions. Furthermore, because all traffic participants are dynamic and uncertain, it is difficult for the vehicle to plan a safe trajectory. An effective approach to addressing this problem is trajectory prediction, which predicts the future behavior of surrounding traffic participants. This allows for continuous optimization of driving strategies and better trajectory planning.

[0003] At present, trajectory prediction includes physics-based models, such as constant speed, constant acceleration, constant angle and rotation speed, but they are only suitable for short-term predictions; intention-based models, such as prototype trajectory clustering, hidden Markov intent prediction, dynamic Bayesian networks, etc., are more accurate than the former in long-term predictions, but do not consider the mutual influence between surrounding vehicles; interactive behavior models are more in line with reality. By estimating potential vehicle-to-vehicle interactions, they can provide more accurate long-term predictions for all vehicles. This model also usually adopts deep learning methods, such as LSTM-based recurrent neural networks, social pool-based convolutional neural networks, and graph neural networks based on the topological structure of vehicle positions. However, most of them only extract features from historical time series and then directly predict trajectories. They do not take into account various possibilities in the future, resulting in a high average error.

[0004] Therefore, how to improve the accuracy of long-term vehicle trajectory prediction is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a vehicle trajectory prediction method and system considering multiple target points to solve the problems mentioned in the background technology.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A vehicle trajectory prediction method considering multiple target points includes the following steps:

[0008] S1. Extract the vehicle historical trajectory features and lane centerline environmental features of all vehicles in the current traffic scene respectively;

[0009] S2. Propagate and update the lane centerline environmental features. Using the updated lane centerline environmental features as constraints, a spatial cross-attention mechanism is used to fuse them with the vehicle's historical trajectory features. The fused features of each vehicle are then subjected to a constrained inter-vehicle feature interaction using a spatial self-attention mechanism.

[0010] S3. Predict the target point of each vehicle, including the midpoint and multiple endpoints, and assign the lane environment features near the midpoint and the endpoint with the highest confidence to the vehicle features;

[0011] S4. Decode and output multiple predicted trajectories and confidence levels, and smooth the predicted trajectories using a filtering algorithm.

[0012] Preferably, the specific content of extracting the vehicle historical features in step S1 includes:

[0013] S11. Obtain the historical trajectories of all vehicles in the current traffic scene, select the target vehicle to establish a local coordinate system, and obtain the historical trajectory input information of the target vehicle and surrounding vehicles. The historical trajectory input information includes the relative distance, average speed, direction information at adjacent moments, and a flag indicating whether a trajectory point exists;

[0014] S12. Based on the historical trajectory input information, the residual causal convolutional network under the FPN structure is used to perform hierarchical feature extraction and multi-layer feature fusion to obtain historical trajectory features.

[0015] Preferably, the specific content of extracting the lane centerline environmental features in step S1 includes:

[0016] S13. Construct a lane centerline road topology map for the current traffic scenario, and integrate the location and shape information of lane centerline node features through MLP.

[0017] S14. Using dilated convolution, we merge the features of nodes at different distances to obtain potential directional constraints. We then use learnable weights on the features of the left and right adjacent nodes to obtain lateral lane line constraints. At the same time, we add the features of the current node to obtain the lane centerline environmental features.

[0018] Preferably, the specific contents of step S2 include:

[0019] S21. Use dilated convolution to propagate and update the aggregated lane center environmental features.

[0020] S22. Filter road nodes near the current vehicle position based on relative distance. Use the distance from the node to the vehicle position as the position code. Combine this with the vehicle and lane centerline features, and then perform MLP fully connected layer learning. This spatial attention is used to achieve feature interaction between lane lines and vehicles.

[0021] S23. Filter vehicles with potential interactions near each vehicle based on relative distance, and use the distance between vehicles as the position encoding. Through this spatial attention mechanism, feature interaction between vehicles is achieved.

[0022] Preferably, the specific contents of step S3 include:

[0023] S31. Use the residual fully connected layer to predict the midpoint of the future trajectory and assign the lane line environment features near the midpoint to each vehicle through the spatial attention mechanism;

[0024] S32. Use a composite residual fully connected layer to predict the locations of multiple future trajectory endpoints. Concatenate the relative distances between the start and end points as position codes, then decode them using a residual fully connected layer. Softmax is used to obtain the confidence score for each endpoint.

[0025] S33. The lane centerline environmental features whose relative distance from the endpoint with the highest confidence is within a preset threshold range are assigned to each vehicle through the attention mechanism, and the target vehicle features are output.

[0026] Preferably, the target vehicle features include vehicle historical trajectory information, environmental information, potential interaction information, and constraints on the target point position.

[0027] Preferably, the specific contents of step S4 include:

[0028] S41. Use multiple parallel residual fully connected layers to output multiple complete trajectory points. Use the relative distance between the start and end points as the position code for concatenation and decode using a residual fully connected layer. Softmax is used to obtain the confidence score for each predicted trajectory.

[0029] S42. Trajectory smoothing, based on the data smoothing method of local least squares polynomial approximation, performs n-order polynomial fitting on the data points within a certain length window, and finally outputs the smoothed predicted trajectory.

[0030] Preferably, the loss function included in target point prediction and trajectory prediction is specifically:

[0031] The midpoint forecast loss is shown as:

[0032]

[0033] Among them, p mid，nrepresents the predicted midpoint position of the nth vehicle, p gtm，n Indicates the true midpoint position of the nth vehicle;

[0034] Multi-endpoint prediction loss includes classification and regression losses:

[0035] L multi =L multi，cls +L multi,reg

[0036] Among them, the classification loss adopts the maximum margin loss:

[0037]

[0038] Among them, c multi，k Indicates the confidence value corresponding to the k modes of multi-endpoint prediction, ε represents the defined margin, and max indicates that if other confidences are less than the selected highest confidence, they are recorded as 0. If they are larger, a loss value will be generated, requiring backpropagation optimization to ensure that the subsequent endpoint with the highest confidence is the currently selected optimal endpoint;

[0039] Regression loss:

[0040]

[0041] Among them, k represents the endpoint with the highest confidence, p gt，n Indicates the actual destination position of the nth vehicle;

[0042] Trajectory prediction loss, including classification loss and regression loss:

[0043] L traj =L traj，cls +L traj，reg .

[0044] in:

[0045]

[0046]

[0047] Finally, the total loss is:

[0048] L=L mid +L multi +L traj .

[0049] A vehicle trajectory prediction system considering multiple target points, based on the vehicle trajectory prediction method considering multiple target points, includes a vehicle feature and lane centerline feature extraction network, a feature fusion network, a target point prediction network and a feature decoding network;

[0050] Vehicle feature and lane centerline feature extraction network, used to extract the vehicle historical trajectory features and lane centerline environmental features of all vehicles in the current traffic scene respectively;

[0051] The feature fusion network is used to propagate and update the new environmental features of the lane center. The updated lane centerline environmental features are used as constraints and a spatial cross-attention mechanism is used to assign constraints to vehicle features. The fused features of each vehicle are then used to perform constrained inter-vehicle feature interaction using a spatial self-attention mechanism.

[0052] The target point prediction network is used to predict each target point separately. The target points include the midpoint and multiple end points, and the lane environment features near the midpoint and the end point with the highest confidence are assigned to the vehicle features.

[0053] The feature decoding network is used to decode and output multiple predicted trajectories and confidence levels, and smooth the predicted trajectories through a filtering algorithm.

[0054] It can be seen from the above technical solution that compared with the prior art, the present invention discloses a vehicle trajectory prediction method and system that considers multiple target points, uses causal convolution to extract vehicle historical features, uses empty lane convolution to learn environmental features, and uses a spatial cross-attention mechanism to fuse environmental information and vehicle-to-vehicle interaction information; considers the target points of the future trajectory including the midpoint and multiple possible end points for prediction, to directly limit the average error of the predicted trajectory, and finally predicts multiple complete trajectories and corresponding confidence levels, thereby increasing the accuracy of the trajectory prediction results and retaining multiple possible predictions for future predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0056] Figure 1 The accompanying drawing is an overall schematic diagram of the vehicle trajectory prediction method considering multiple target points provided by the present invention;

[0057] Figure 2 The accompanying drawing is a schematic diagram of vehicle feature extraction provided by the present invention;

[0058] Figure 3 The accompanying figure is a schematic diagram of the spatial attention mechanism structure provided by the present invention;

[0059] Figure 4 The accompanying drawing is a schematic diagram of the multi-endpoint prediction structure provided by the present invention;

[0060] Figure 5 The accompanying drawing is a schematic diagram of the prediction results provided by the present invention; DETAILED DESCRIPTION

[0061] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0062] The embodiment of the present invention discloses a vehicle trajectory prediction method considering multiple target points, such as Figure 1 , including the following steps:

[0063] S1. Extract the vehicle historical trajectory features and lane centerline environmental features of all vehicles in the current traffic scene respectively;

[0064] S2. Propagate and update the lane centerline environmental features. Using the updated lane centerline environmental features as constraints, a spatial cross-attention mechanism is used to fuse them with the vehicle's historical trajectory features. The fused features of each vehicle are then subjected to a constrained inter-vehicle feature interaction using a spatial self-attention mechanism.

[0065] S3. Predict the target point of each vehicle, including the midpoint and multiple endpoints, and assign the lane environment features near the midpoint and the endpoint with the highest confidence to the vehicle features;

[0066] S4. Decode and output multiple predicted trajectories and confidence levels, and smooth the predicted trajectories using a filtering algorithm.

[0067] In order to further implement the above technical solution, the specific content of extracting vehicle historical features in step S1 includes:

[0068] S11. Obtain the historical trajectories of all vehicles in the current traffic scene, select the target vehicle to establish a local coordinate system, and obtain the historical trajectory input information of the target vehicle and surrounding vehicles. The historical trajectory input information includes the relative distance, average speed, direction information at adjacent moments, and a flag indicating whether a trajectory point exists;

[0069] Specifically:

[0070] The target detection and tracking algorithm is used to obtain the historical trajectory of each vehicle in the scene, including the position information of the most recent time period before the current moment (for example, sampling at 10Hz within 2 seconds). The target vehicle is selected, and a local coordinate system is established with the forward direction of its lane as the positive y-axis direction and the direction perpendicular to the positive y-axis and to the right as the x-axis direction. The observation positions of the surrounding vehicles are projected onto this coordinate system. The input set of historical trajectories is:

[0071]

[0072] Among them, h t Represents the input at time t, including the target vehicle T and all other vehicles S:

[0073]

[0074] Where i represents the i-th vehicle. The vehicle information inputs of T and S are the same, which are the relative distance between adjacent moments, the average speed, the direction information represented by its sine and cosine, and the flag indicating whether the trajectory point exists:

[0075]

[0076] S12. Based on the historical trajectory input information, a residual causal convolutional network under the FPN structure is used to obtain historical trajectory features;

[0077] Specifically, such as Figure 2 :

[0078] Causal convolution is used as the basic processing unit, and a residual structure is constructed to extract historical trajectory features in three layers. Each layer is a ResBlock residual module, where CausalConv is a causal convolution. In its parameter setting, 3 is the convolution kernel size, C is the output channel, S is the convolution step, GN is the group normalization, and ReLU is the activation function. The convolution step of the first layer is set to 1, which is mainly used to change the number of feature channels. When the input and output sizes are consistent, there is no need to add an extra layer of convolution to the input to change the size of the splicing. Only the branch structure is retained. The last two layers are used to compress the size and expand the dimension, continuously improving the receptive field, thereby aggregating longer historical trajectories. Therefore, the number of output channels is set to 64 and 128 respectively, and the convolution step is set to 2.

[0079] In order to fuse multi-scale features, the FPN feature pyramid structure is adopted, such as Figure 2As shown on the right, the three layers of local and global features are fused. The ResBlock module parameters in each layer define the number of output channels and the convolution step size. Conv uses conventional convolution to keep the number of multi-scale channels consistent. Upsample is upsampling to keep the sizes of adjacent layers consistent so that addition can be performed. The purpose of adding the ResBlock module again in the last part is to learn to compress the time series length into a unit value. That is, each vehicle only uses a 128-dimensional feature vector to represent its historical trajectory characteristics.

[0080] To further implement the above technical solution, the specific contents of extracting lane centerline environmental features in step S1 include:

[0081] S13. Construct a lane centerline road topology map for the current traffic scenario. Lane centerline node features are integrated with position and shape information using an MLP.

[0082] S14 uses dilated convolution to combine features of nodes at different distances to obtain potential directional constraints. It then uses learnable weights on features of adjacent nodes to obtain lateral lane constraints. It also adds features of the current node to obtain lane centerline environmental features.

[0083] Specifically:

[0084] The nodes of the lane centerline are represented by the position information (the midpoint v of the two end points) i ) and shape information (the vector formed by subtracting the endpoints ), so the feature of each waypoint is also composed of the sum of these two parts:

[0085]

[0086] MLP consists of two linear fully connected layers and is processed by group normalization and activation function ReLU. The two parts of features are added together and then processed through ReLU to obtain the waypoint features;

[0087] In order to strengthen the association information of the lane centerline, it is necessary to aggregate it in a larger range according to the graph structure. Specifically, the lane graph structure can be easily divided into three parts: the waypoint features of the current position, the left and right waypoint features related to the horizontal direction, and the front and rear waypoint features related to the longitudinal direction. The horizontal direction usually represents a soft constraint that can be crossed, while the longitudinal direction represents the direction guidance that can be maintained, so different weight coefficients are used to represent it. Along the longitudinal direction of the lane line, it will advance to a relatively far position, so the concept of void convolution is introduced to achieve the aggregation of longitudinal waypoint features at different distances:

[0088]

[0089] The first term is the feature of the waypoint itself, the second term is the aggregation of the left and right adjacent waypoints, and the third term is the aggregation of the forward and backward nodes, which represents the transmission range. X is the waypoint feature matrix, and W is the corresponding weight matrix. represents the c-th power of the forward adjacency matrix, which means that the feature information of the waypoint is forwarded c times, thereby aggregating the features of the waypoints that are c distances away. Similarly, represents the c-th power of the backward adjacency matrix;

[0090] After the dilated lane convolution, the image is fused through a fully connected layer with 128 input and output channels. The input and output are then summed using a residual structure, and the final lane encoding is obtained through an activation function.

[0091] In order to further implement the above technical solutions, Figure 3 , the specific contents of step S2 include:

[0092] S21. Use dilated convolution to propagate and update the aggregated lane center environmental features.

[0093] S22. Filter road nodes near the current vehicle location based on relative distance, use the distance from the node to the vehicle location as the position code, and concatenate the result with the target vector and the environment information to perform MLP fully connected layer learning.

[0094] The target vector in S22 corresponds to the vehicle feature, and the environmental information corresponds to the lane feature. Based on the Euclidean distance to the target vehicle's last observed position, the lane centerline-related waypoints within the target range are filtered out, and those outside the range are masked. A fully connected layer is added to achieve dimensionality unification. The filtered relative distance is encoded using position embedding, and then concatenated with the vehicle feature for full-connected layer learning. The lane centerline feature is assigned to the vehicle feature, and the vehicle feature itself is connected to ensure at least identity mapping to obtain the final updated vehicle feature:

[0095]

[0096] Among them, x i is the feature of the i-th target point, x j is the feature of the j-th waypoint, Δ i，j Represents the distance feature after position embedding, which is obtained by two linear layers and ReLU. W is the weight matrix, which is implemented by the Linear linear layer. Concat is the MLP fully connected layer, which includes four layers of Linear linear layers. Normalization and nonlinear activation are performed between each layer.

[0097] S23. Filter vehicles near each vehicle that could potentially interact based on relative distance, and use the distance between vehicles as the position code. The concatenation of the three results is then used to learn the feature interaction between vehicles in an MLP fully connected layer.

[0098] Specifically, the interaction features between vehicles are transmitted to each other and realized by the spatial self-attention mechanism. At this time, the target vector is the feature of each vehicle, and the environmental information also corresponds to the features of the surrounding vehicles. At this time, this module can be regarded as a spatial self-attention mechanism. In the corresponding previous formula, x i It represents the characteristics of the i-th vehicle, x j is the feature representing the j-th vehicle.

[0099] In order to further implement the above technical solution, the specific contents of step S3 include:

[0100] Target point prediction is divided into two processes: midpoint prediction and multi-endpoint prediction. The former only predicts one midpoint of the future trajectory, while the latter predicts six end points and their corresponding confidence levels.

[0101] Specifically:

[0102] S31. Use the residual fully connected layer to predict the midpoint position of the future trajectory and assign the lane line environment features near the midpoint to each vehicle through the spatial attention mechanism:

[0103] Through Figure 4 The fully connected layer decoding under the ResLinear residual structure shown on the left outputs the midpoint position coordinates. The spatial cross-attention method is used to select lane center points whose relative distance from the midpoint is within a threshold range, thereby assigning the lane constraint at this location to the vehicle feature.

[0104] S32. Use the composite residual fully connected layer to predict multiple future trajectory endpoint positions and corresponding confidence levels;

[0105] Predict 6 endpoints, such as Figure 4 As shown in the figure, for each endpoint, the same decoding subnetwork as the midpoint prediction is used for a total of 6 times. The confidence of each endpoint is calculated by the latter two parts. In addition to the vehicle features, the relative distance between the starting and ending points needs to be encoded. After splicing the two, the decoding output is used using the ResLinear structure. Finally, the confidence of each endpoint is obtained through Softmax.

[0106] S33. For the selected endpoint with the highest confidence, use the spatial attention mechanism to filter the lane centerline environmental features within the preset threshold range and assign them to each vehicle.

[0107] In order to further implement the above technical solution, the target vehicle characteristics include vehicle historical trajectory information, environmental information, potential interaction information, and constraints on the target point position.

[0108] In order to further implement the above technical solution, the specific contents of step S4 include:

[0109] S41. Use multiple parallel residual fully connected layers to output multiple complete trajectory points. Use the relative distance between the start and end points as the position code for concatenation and decode using a residual fully connected layer. Softmax is used to obtain the confidence score for each predicted trajectory.

[0110] In this embodiment, the structure of the feature decoding output is as follows: Figure 4 Change the multi-end point to a complete trajectory output, with a total of 30×(x, y) trajectory points. 6 trajectories need to be decoded 6 times, and the final prediction value of the (6, 30, 2) dimension is obtained.

[0111] S52 In this embodiment, the predicted trajectory is smoothed using the Savitzky-Golay filtering algorithm. This algorithm is based on a local least squares polynomial, performs nth-order polynomial fitting on the data points within a window of a certain length, and finally outputs the smoothed predicted trajectory.

[0112] In order to further implement the above technical solution, the loss functions included in target point prediction and trajectory prediction are specifically as follows:

[0113] Midpoint prediction loss, only predicting a midpoint position for each vehicle, and finding the smoothL1 loss relative to the Euclidean distance:

[0114]

[0115] Among them, p mid，n represents the predicted midpoint position of the nth vehicle, p gtm，n Indicates the true midpoint position of the nth vehicle;

[0116] End point prediction loss, including classification and regression losses:

[0117] L multi =L multi，cls +L multi，r x g

[0118] Among them, the classification loss adopts the maximum margin loss, because the confidence level expected to be selected is the highest confidence level, so the initial definition is Mark this value, and the confidence corresponding to other endpoints will not exceed the selected optimal endpoint, that is, calculate the values ​​of other k modes and The difference between , to maximize the marginalized expected confidence:

[0119]

[0120] Among them, c multi，k Represents the confidence value corresponding to the k modes of multi-endpoint prediction, ε represents the defined margin, and max indicates that if other confidences are smaller than the selected highest confidence, they are recorded as 0. Otherwise, a loss value will be generated and backpropagation optimization is required to ensure that the subsequent endpoint with the highest confidence is the currently selected optimal endpoint;

[0121] Regression loss:

[0122]

[0123] in, Indicates the selection of the endpoint with the highest confidence, p gt，n Indicates the actual destination position of the nth vehicle;

[0124] Trajectory prediction loss is for the entire trajectory of multimodal prediction, including classification loss and regression loss:

[0125] L traj =L traj，cls +L traj，reg .

[0126]

[0127]

[0128] The final total loss is:

[0129] L=L mid +L multi +L traj .

[0130] A vehicle trajectory prediction system considering multiple target points is based on a vehicle trajectory prediction method considering multiple target points, including a vehicle feature and lane centerline feature extraction network, a feature fusion network, a target point prediction network and a feature decoding network;

[0131] Vehicle feature and lane centerline feature extraction network, used to extract the vehicle historical trajectory features and lane centerline environmental features of all vehicles in the current traffic scene respectively;

[0132] The feature fusion network is used to propagate and update the new environmental features of the lane center. The updated lane centerline environmental features are used as constraints and are fused with the vehicle's historical trajectory features using a spatial cross-attention mechanism. The fused features of each vehicle are then used to perform constrained inter-vehicle feature interaction using a spatial self-attention mechanism.

[0133] The target point prediction network is used to predict each target point separately. The target points include the midpoint and multiple end points, and the lane environment features near the midpoint and the end point with the highest confidence are assigned to the vehicle features.

[0134] The feature decoding network is used to decode and output multiple predicted trajectories and confidence levels, and smooth the predicted trajectories through a filtering algorithm.

[0135] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0136] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A vehicle trajectory prediction method considering multiple target points, characterized in that: The following steps are involved: S1. Extract the historical trajectory features and lane centerline environmental features of all vehicles in the current traffic scene respectively; S2. Propagate and update lane centerline environmental features. Using the updated lane centerline environmental features as constraints, the spatial cross-attention mechanism is used to fuse them with vehicle historical trajectory features. The fused features of each vehicle are then used to perform constrained inter-vehicle feature interaction using the spatial self-attention mechanism. S3. Predict the target point of each vehicle, including the midpoint and multiple endpoints, and assign the lane environment features near the midpoint and the endpoint with the highest confidence to the vehicle features; S4. Decode and output multiple predicted trajectories and confidence levels, and smooth the predicted trajectories using a filtering algorithm; The specific contents of step S2 include: S21. Use dilated convolution to propagate and update the aggregated lane centerline environmental features again; S22. Filter road nodes near the current vehicle position based on relative distance. Use the distance from the node to the vehicle position as the position code. Combine this with the vehicle features and lane centerline environmental features, and perform MLP fully connected layer learning. Using this spatial self-attention mechanism, lane line constraints are assigned to vehicle features. S23. Filter vehicles near each vehicle that could potentially interact based on relative distance, use the vehicle-to-vehicle distance as the position code, and perform MLP fully connected layer learning on the result of concatenating the features of the vehicle and the interacting vehicle to achieve vehicle-to-vehicle interaction.

2. The vehicle trajectory prediction method considering multiple target points according to claim 1, characterized in that: The specific contents of extracting vehicle historical trajectory features in step S1 include: S11. Obtain the historical trajectories of all vehicles in the current traffic scene, select the target vehicle to establish a local coordinate system, and obtain the historical trajectory input information of the target vehicle and surrounding vehicles. The historical trajectory input information includes the relative distance, average speed, direction information at adjacent moments, and a flag indicating whether a trajectory point exists; S12. Based on the historical trajectory input information, the residual causal convolutional network under the FPN structure is used to perform multi-scale feature fusion to obtain historical trajectory features.

3. The vehicle trajectory prediction method considering multiple target points according to claim 2, characterized in that: The specific contents of extracting lane centerline environmental features in step S1 include: S13. Construct a lane centerline road topology map for the current traffic scenario, and integrate the location and shape information of lane centerline node features through MLP. S14. Using dilated convolution, we merge the features of nodes at different distances to obtain potential directional constraints. We then use learnable weights on the features of the left and right adjacent nodes to obtain lateral lane line constraints. At the same time, we add the features of the current node to obtain the lane centerline environmental features.

4. The vehicle trajectory prediction method considering multiple target points according to claim 1, characterized in that: The calculation method for feature interaction using the spatial attention mechanism is: in, It is The characteristics of the target point, It is The characteristics of the interacted object, It represents the distance feature after position encoding, which is obtained by two linear layers and ReLU. W is the weight matrix, which is implemented by the Linear linear layer. concat is the MLP fully connected layer, which includes four Linear linear layers. Normalization and nonlinear activation are performed between each layer.

5. The vehicle trajectory prediction method considering multiple target points according to claim 1, characterized in that: The specific contents of step S3 include: S31. Use the residual fully connected layer to predict the midpoint of the future trajectory and assign the lane line environment features near the midpoint to each vehicle through the spatial attention mechanism; S32. Use a composite residual fully connected layer to predict the locations of multiple future trajectory endpoints. Concatenate the relative distances between the start and end points as position codes, then decode them using a residual fully connected layer. Softmax is used to obtain the confidence score for each endpoint. S33. Assign the lane centerline environmental features whose relative distance from the endpoint with the highest confidence is within a preset threshold range to each vehicle through the spatial attention mechanism, and output the target vehicle features.

6. The vehicle trajectory prediction method considering multiple target points according to claim 5, characterized in that: The target vehicle features include vehicle historical trajectory information, environmental information, potential interaction information, and constraints on the target point location.

7. The vehicle trajectory prediction method considering multiple target points according to claim 5, characterized in that: The specific contents of step S4 include: S41. Use multiple parallel residual fully connected layers to output multiple complete trajectory points. Use the relative distance between the start and end points as the position code for concatenation and decode using a residual fully connected layer. Softmax is used to obtain the confidence score for each predicted trajectory. S42. Trajectory smoothing, based on the data smoothing method of local least squares polynomial approximation, performs n-order polynomial fitting on the data points within a certain length window, and finally outputs the smoothed predicted trajectory.

8. The vehicle trajectory prediction method considering multiple target points according to claim 1, characterized in that: The loss functions included in target point prediction and trajectory prediction are as follows: The midpoint forecast loss is shown as: in, represents the predicted midpoint position of the nth vehicle, Indicates the true midpoint position of the nth vehicle; Multi-endpoint prediction loss includes classification and regression losses: Among them, the classification loss adopts the maximum margin loss: in, Indicates the confidence value corresponding to the k-mode of multi-endpoint prediction, Represents the defined margin. Max means that if other confidence levels are less than the selected highest confidence level, they are recorded as 0. If they are larger, a loss value will be generated, and backpropagation optimization is required to ensure that the subsequent endpoint with the highest confidence level is the currently selected optimal endpoint. Regression loss: in, Indicates that the endpoint with the highest confidence is selected. Indicates the actual destination position of the nth vehicle; Trajectory prediction loss, including classification loss and regression loss: in: Finally, the total loss is: 。 9. A vehicle trajectory prediction system considering multiple target points, characterized in that: A vehicle trajectory prediction method considering multiple target points according to any one of claims 1 to 8, comprising a vehicle feature and lane centerline feature extraction network, a feature fusion network, a target point prediction network, and a feature decoding network; Vehicle feature and lane centerline feature extraction network, used to extract the vehicle historical trajectory features and lane centerline environmental features of all vehicles in the current traffic scene respectively; The feature fusion network is used to propagate and update lane centerline environmental features. The updated lane centerline environmental features are used as constraints and fused with vehicle historical trajectory features using a spatial cross-attention mechanism. The fused features of each vehicle are then used to perform constrained vehicle-to-vehicle feature interaction using a spatial self-attention mechanism. The target point prediction network is used to predict each target point separately. The target points include the midpoint and multiple end points, and the lane centerline environmental features near the midpoint and the end point with the highest confidence are assigned to the vehicle features. The feature decoding network is used to decode and output multiple predicted trajectories and confidence levels, and smooth the predicted trajectories through a filtering algorithm.

Images