A vehicle trajectory prediction method based on wave semantic network

Abstract

The present invention discloses a vehicle trajectory prediction method based on a wave semantic network, comprising: obtaining a high-precision map centered on a target vehicle; dividing the high-precision map into a number of data blocks equal to the number of semantic channels; inputting the data blocks into a prediction model to predict the trajectory of the target vehicle in a future time period. The execution process of the prediction model includes: inputting the data blocks into a first first module and inputting the output result into a next first module; performing global average pooling and fully connected layer processing on the output data of the last first module to obtain a total feature; and inputting the total feature into a Laplacian decoder to predict the trajectory of the target vehicle in the future time period. The present invention introduces "waves" to characterize complex traffic scenarios and uses the superposition of waves to implicitly model multi-flow social interactions, enabling the model to more deeply understand the structure and dynamic changes of complex traffic scenarios and improve the ability to predict vehicle trajectories in traffic scenarios.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and more particularly to a vehicle trajectory prediction method based on a wave semantic network. Background Art

[0002] The core of the trajectory prediction task lies in accurately inferring the future motion trends of the intelligent agent by understanding the semantics of the contextual scene and the historical trajectories of traffic participants. As an important component of the autonomous driving system, vehicle trajectory prediction closely connects the upstream environmental perception module with the downstream planning and decision-making module. However, in actual deployment, accurately predicting the future trajectory of autonomous vehicles is a daunting challenge. The root of this problem lies in how to effectively model the intricate interaction network, which requires full consideration of various necessary constraints, evolving time dependencies, and continuously changing social interaction patterns in dynamic traffic environments. For these reasons, the trajectory prediction modules proposed in the existing technology have low trajectory accuracy. Therefore, there is an urgent need for a trajectory prediction method to improve the ability to understand complex traffic scenarios and thus improve the accuracy of trajectory prediction. Summary of the Invention

[0003] The present invention is proposed based on the above-mentioned requirements of the prior art. The technical problem to be solved by the present invention is to propose a vehicle trajectory prediction method based on wave semantic network to deeply understand the structure and dynamic changes of traffic scenes through wave superposition, thereby improving the accuracy of trajectory prediction.

[0004] In order to solve the above problems, the present invention is implemented by adopting the following technical solutions:

[0005] A vehicle trajectory prediction method based on a wave semantic network, the method comprising: obtaining a high-precision map centered on a target vehicle, the high-precision map having multiple semantic channels, the semantic channels comprising drivable areas, lane boundaries, and traffic lights; the semantic channels storing historical trajectory data of the target vehicle and other vehicles; dividing the high-precision map into multiple data blocks equal in number to the semantic channels; inputting the data blocks into a prediction model to predict the trajectory of the target vehicle in a future time period, the prediction model comprising multiple first modules connected in sequence, a global average pooling layer and a fully connected layer, and a Laplace decoder, the first module comprising a fully connected layer, a wave semantic interaction block, and a channel learning block connected in sequence, the execution process of the prediction model comprising: inputting the data block into a first first module, and inputting the output result into a next first module; performing global average pooling and fully connected layer processing on the output data of the last first module to obtain a total feature; and inputting the total feature into a Laplace decoder to predict the target vehicle in a future time period. trajectory; the execution process of the first module includes: inputting input data into the fully connected layer to obtain embedded features; inputting the embedded features into the wave semantic interaction block to obtain time domain features, including: using the first multi-layer perceptron and the second multi-layer perceptron to process the embedded features respectively to obtain the first amplitude and the first phase; based on the first amplitude and the first phase, fusing the waves corresponding to any two embedded features to obtain the second amplitude and the second phase; obtaining the corresponding synthetic wave data based on the second amplitude and the second phase; using the third multi-layer perceptron and the fourth multi-layer perceptron to process the synthetic wave data respectively to obtain the first data and the second data; performing global wave fusion processing on the first data to obtain the first feature; performing local wave fusion processing on the second data to obtain the second feature; obtaining time domain features based on the first feature and the second feature; inputting the time domain features into the channel learning block to obtain channel learning features, including: converting the time domain features to obtain frequency domain features; using the multi-layer perceptron to process the frequency domain features, and performing inverse transformation on the processed features to obtain channel learning features.

[0006] Optionally, global wave fusion processing is performed on the first data to obtain a first feature, including: dividing the first data along the channel dimension and the width dimension respectively to obtain multiple first data segments and multiple second data segments; permuting each first data segment in the height dimension and the channel dimension to obtain first permuted data; permuting each second data segment in the width dimension and the channel dimension to obtain second permuted data; splicing the first permuted data and the second permuted data along the channel dimension respectively, and fusing the spliced ​​data to obtain first fused data and second fused data; processing the first data using a fully connected layer to obtain third fused data; determining the learning attention parameters of the first fused data, the second fused data and the third fused data, and fusing the first fused data, the second fused data and the third fused data based on the learning attention parameters to obtain the first feature.

[0007] Optionally, the first multilayer perceptron and the second multilayer perceptron are used to process the embedded features of the block input respectively to obtain the first amplitude and the first phase, which are expressed as follows: i =MLP1(h i ,W i 0 ),i∈[1,2,...,m],θ i =MLP2(h i ,W i 1 ),i∈[1,2,...,m], where z i represents the first amplitude corresponding to the i-th embedded feature, i represents the identification number of the embedded feature, MLP1 represents the first multi-layer perceptron, and h i represents the i-th embedding feature, W i 0 represents the amplitude weight corresponding to the i-th embedded feature, m represents the number of embedded features, θ i represents the first phase corresponding to the i-th embedded feature, MLP2 represents the second multi-layer perceptron, and W i 1 represents the phase weight corresponding to the i-th embedded feature.

[0008] Optionally, based on the first amplitude and the first phase, the waves corresponding to any two embedded features are fused to obtain the second amplitude and the second phase, which are expressed as follows:

[0009] θ r =θ i +atan2(|z a |⊙sin(θ a -θ i ),|z i |+|z a |⊙cos(θa -θ i )), where z r represents the second amplitude with identification number r obtained by fusion of the i-th embedded feature and the a-th embedded feature, z a represents the first amplitude corresponding to the a-th embedded feature, θ a represents the first phase corresponding to the ath embedded feature, θ r It represents the second phase with identification number r obtained by fusing the i-th embedded feature with the a-th embedded feature, and atan2() represents the two-parameter inverse tangent function.

[0010] Optionally, obtaining corresponding synthetic wave data based on the second amplitude and the second phase includes: in, represents the rth synthetic wave data, b represents an imaginary unit, and n represents the total number of synthetic wave data.

[0011] Optionally, determining the learning attention parameters of the first fusion data, the second fusion data and the third fusion data includes: adding the first fusion data, the second fusion data and the third fusion data to obtain total fusion data; performing adaptive average pooling on the total fusion data to obtain pooled data; processing the pooled data using a Softmax function to obtain a first learning attention parameter corresponding to the first fusion data, a second learning attention parameter corresponding to the second fusion data and a third learning attention parameter corresponding to the third fusion data.

[0012] Optionally, the first fused data, the second fused data, and the third fused data are fused based on the learned attention parameter to obtain a first feature, which is expressed as: in, represents the first feature, α[0] represents the first learning attention parameter, represents the first fusion data, α[1] represents the second learning attention parameter, represents the second fusion data, α[2] represents the third learning attention parameter, Represents the third fused data.

[0013] Optionally, performing local wave fusion processing on the second data to obtain a second feature includes: processing all the second data based on the Euler formula and the quantum measurement method to obtain the second feature, which is expressed as: in, represents the second feature, j represents the identification number of the second data, represents the first weight of the j-th second data, Z j represents the amplitude of the jth second data, θ j represents the phase of the j-th second data, represents the second weight of the j-th second data, and n represents the total number of second data.

[0014] Optionally, obtaining a time domain feature based on the first feature and the second feature includes: adding the first feature and the second feature to obtain an addition result; convolving the addition result, and performing a residual connection between the convolution result and the embedded feature to obtain the time domain feature.

[0015] Optionally, the Laplacian decoder includes three fully connected layers, and every two fully connected layers are connected by an activation function and a normalization layer.

[0016] Compared with the existing technology, the present invention provides a vehicle trajectory prediction method based on wave semantic networks, which introduces "waves" to characterize complex traffic scenes and uses the superposition of waves to implicitly model multi-flow social interactions, so that the model can have a deeper understanding of the structure and dynamic changes of complex traffic scenes, and improve the ability to predict vehicle trajectories in complex traffic scenes. By deeply fusing wave semantic information in spatial and channel dimensions through wave semantic interaction blocks, the relationship between global and local features can be effectively captured. By introducing a channel fusion module, the model can efficiently aggregate complementary information between different channels, improve the richness of feature expression, and significantly enhance the abstraction and understanding of scene semantics, thereby improving the overall performance of vehicle trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this specification. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0018] Figure 1 is a flow chart of a vehicle trajectory prediction method based on wave semantic network provided in this embodiment;

[0019] Figure 2 is a schematic diagram of the architecture of the prediction model provided in this embodiment;

[0020] Figure 3 This is a data transmission diagram of the wave semantic interaction block provided in this embodiment. DETAILED DESCRIPTION

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. 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 shall fall within the scope of protection of the present invention.

[0022] To facilitate understanding of the embodiments of the present invention, the following will be further explained with reference to specific embodiments in conjunction with the accompanying drawings. The embodiments do not constitute a limitation on the scope of protection of the present invention.

[0023] This embodiment provides a vehicle trajectory prediction method based on wave semantic network, the process of which is as follows: Figure 1 As shown, including:

[0024] S1 obtains a high-precision map centered on the target vehicle.

[0025] The vehicle's historical trajectory and surrounding environment are encoded to create a high-definition map centered on the target vehicle, providing semantic information for scene understanding. This high-precision map is a high-precision map used for autonomous driving. It serves as prior data for autonomous driving and has multiple semantic channels, including drivable areas, lane boundaries, and traffic lights. The historical trajectory of the target vehicle and other vehicles is represented in the semantic channels of the HD map in a dynamically changing matrix format.

[0026] S2 divides the high-precision map into multiple data blocks with the same number as the semantic channels.

[0027] Assume that S1 obtains high-precision maps Among them, H represents the height of the HD map, W represents the width of the HD map, and C represents the number of channels of the HD map.

[0028] This step will be high-precision map Divide into a series of discrete data blocks, each of which has a size of V×V×C, and a total of m data blocks are obtained, where V represents the height and width of the data block. Each data block represents a different aspect of the traffic scene.

[0029] S3 inputs the data block into the prediction model to predict the trajectory of the target vehicle in the future time period.

[0030] like Figure 2As shown, the prediction model includes multiple first modules connected in sequence, a global average pooling layer and a fully connected layer, and a Laplacian decoder, and the first module includes a fully connected layer, a wave semantic interaction block and a channel learning block connected in sequence.

[0031] In the S3 step, include:

[0032] S30 inputs the data block into the first first module, and inputs the output result into the next first module.

[0033] The execution process of the first module includes:

[0034] S300 inputs the input data into the fully connected layer to obtain embedded features.

[0035] The data block is input into the fully connected layer in the first first module for linear transformation to obtain the corresponding embedding features. For other first modules, the output result of the previous first module is input into the fully connected layer to obtain the corresponding embedding features to merge the waves.

[0036] Assume that m embedding features Η = [h1,h2,...,h i ...,h m ], where i represents the identification number of the embedded feature, h i represents the i-th embedded feature, and m represents the number of embedded features.

[0037] S310 inputs the embedded features into the wave semantic interaction block to obtain time domain features.

[0038] The wave semantic interaction block is used to capture the complex dependencies generated by semantic interactions in traffic scenarios. The complex interactions occurring in traffic scenarios can be regarded as interactions between various data blocks. Complex interactions are achieved by superimposing waves at both the global and local levels to fuse amplitude and phase information.

[0039] like Figure 3 As shown, in this step, it includes:

[0040] Step 1: Use a first multi-layer perceptron and a second multi-layer perceptron to process the embedded features respectively to obtain a first amplitude and a first phase.

[0041] This embodiment inputs the embedded features into the wave semantic representation layer to describe the embedded features as a wave function with amplitude and phase. Specifically:

[0042] Input the embedded features into the first multi-layer perceptron to obtain the first amplitude, which is expressed as: i =MLP1(h i ,W i 0),i∈[1,2,...,m], where z i represents the first amplitude corresponding to the i-th embedded feature, MLP1 represents the first multi-layer perceptron, and W i 0 Represents the amplitude weight corresponding to the i-th embedded feature, which is a learnable parameter.

[0043] The first phase can be derived by estimating the function θ, i.e., θ i =Θ(h i ,W i 1 ),i∈[1,2,...,m], in order to ensure computational efficiency, this embodiment inputs the embedded features into the second multi-layer perceptron to obtain the first phase, which is expressed as: θ i =MLP2(h i ,W i 1 ),i∈[1,2,...,m], where θi represents the first phase corresponding to the i-th embedded feature, MLP2 represents the second multi-layer perceptron, and W i 1 Represents the phase weight corresponding to the i-th embedded feature, which is a learnable parameter.

[0044] Step 2: Based on the first amplitude and the first phase, the waves corresponding to any two embedded features are fused to obtain the second amplitude and the second phase; based on the second amplitude and the second phase, the corresponding synthetic wave data is obtained.

[0045] Specifically:

[0046] Enter the first amplitude and first phase of any two blocks of data into the expression θ r =θ i +atan2(|z a |⊙sin(θ a -θ i ),|z i |+|z a |⊙cos(θ a -θ i )), the corresponding second amplitude and second phase are obtained, where z r represents the second amplitude with identification number r obtained by fusion of the i-th embedded feature and the a-th embedded feature, z a represents the first amplitude corresponding to the a-th embedded feature, θ a represents the first phase corresponding to the ath embedded feature, θ r It represents the second phase with identification number r obtained by fusing the i-th embedded feature with the a-th embedded feature, and atan2() represents the two-parameter inverse tangent function.

[0047] Enter the second amplitude and the second phase into the formula Get all synthetic wave data, among which, represents the rth synthetic wave data, b represents an imaginary unit, and n represents the total number of synthetic wave data.

[0048] During the wave superposition process, the increase or decrease in the amplitude of the composite wave is related to the phase difference, so the phase can fine-tune the intensity of complex social interactions within the traffic scene.

[0049] Step 3: Use the third multi-layer perceptron and the fourth multi-layer perceptron to process the synthetic wave data to obtain the first data and the second data respectively.

[0050] The third multi-layer perceptron is used to process the input data to obtain the first data Z~ g , use the fourth multi-layer perceptron to process the input data to obtain the second data Z~ l .

[0051] Step 4: Perform global wave fusion processing on the first data to obtain the first feature.

[0052] The global wavelet fusion block includes three branches, namely, coded wavelet information along the height dimension, coded wavelet information along the width dimension, and coded wavelet information along the channel dimension. The first data is input into the three branches respectively to focus on different wavelet semantic information in space, specifically:

[0053] The first data is divided along the channel dimension and the width dimension respectively to obtain a plurality of first data segments and a plurality of second data segments.

[0054] Specifically, the first data is divided along the channel dimension to obtain multiple first data segments. in, represents the Uth first data segment, and U represents the total number of first data segments.

[0055] Divide the first data along the width dimension to obtain multiple second data segments in, represents the U-th second data segment, U represents the total number of second data segments, and in this embodiment, the number of the first data segments and the second data segments obtained by division are the same.

[0056] Each first data segment is permuted in terms of height dimension and channel dimension to obtain first permuted data.

[0057] Each second data segment is permuted in terms of width and channel dimensions to obtain second permuted data.

[0058] The first permuted data and the second permuted data are spliced ​​along the channel dimension, and the spliced ​​data are fused to obtain first fused data and second fused data.

[0059] This embodiment uses a first fully connected layer with a weight of size C×C to fuse the data after the first permuted data is spliced ​​along the channel dimension, and outputs the first fused data Use the second fully connected layer with a weight of size C×C to fuse the data after the second permutation data is spliced ​​along the channel dimension, and output the second fused data

[0060] The first data is processed using a fully connected layer to obtain third fused data.

[0061] This embodiment uses a third fully connected layer with a weight of size C×C to perform linear projection on the first data to fuse the height information to obtain the third fused data

[0062] Learning attention parameters of the first fused data, the second fused data, and the third fused data are determined.

[0063] In order to calibrate the importance of different branches, the prediction model of this embodiment adaptively learns attention parameters. Specifically, the first fusion data, the second fusion data, and the third fusion data are added to obtain total fusion data; the total fusion data is adaptively averaged pooled to obtain pooled data; the pooled data is processed using the Softmax function to obtain a first learning attention parameter corresponding to the first fusion data, a second learning attention parameter corresponding to the second fusion data, and a third learning attention parameter corresponding to the third fusion data, which are expressed as follows: Among them, α represents the learning attention parameter, including the first learning attention parameter, the second learning attention parameter and the third learning attention parameter, and AAP represents adaptive average pooling.

[0064] The first feature is obtained by fusing the first fused data, the second fused data and the third fused data based on the learned attention parameter.

[0065] The expression for determining the first characteristic is:

[0066]

[0067] in, represents the first feature, α[0] represents the first learning attention parameter, α[1] represents the second learning attention parameter, and α[2] represents the third learning attention parameter.

[0068] Step 5: Perform local wave fusion processing on the second data to obtain the second feature.

[0069] All the second data are processed based on the Euler formula and quantum measurement method to obtain the second feature, which is expressed as:

[0070]

[0071] in, represents the second feature, j represents the identification number of the second data, Represents the first weight of the j-th second data, which is a learnable parameter, Z j represents the amplitude of the jth second data, θ j represents the phase of the j-th second data, represents the second weight of the j-th second data, which is a learnable parameter, and n represents the total number of second data.

[0072] In this step, different vehicles interact with each other in terms of amplitude and phase information at the local level. The interaction results are encapsulated into the second feature in the form of wave superposition, reflecting the interactive dynamics in the traffic scene.

[0073] Step 6: Obtain time domain features based on the first feature and the second feature.

[0074] In this step, the first feature and the second feature are input to the branch fusion layer, and the following operations are performed:

[0075] The first feature and the second feature are added to obtain an addition result; the addition result is convolved, and the convolution result is residually connected with the embedded feature to obtain a time domain feature.

[0076] S320 inputs the time domain features into the channel learning block to obtain channel learning features.

[0077] To enhance the channel representation capability of the prediction model in trajectory prediction, this embodiment designs a channel learning block. Within this module, the model is allowed to capture the interactions and correlations between different variables, thereby significantly improving the accuracy and reliability of the prediction results.

[0078] This step includes:

[0079] The time domain features are converted into frequency domain features, the frequency domain features are processed using a multi-layer perceptron, and the processed features are inversely transformed to obtain channel learning features.

[0080] The multilayer perceptron in the channel learning block has learnable weights and biases. The frequency domain is converted to the time domain through inverse transformation to obtain the learned features.

[0081] Furthermore, a layer normalization layer is connected before the wave semantic interaction block, and layer normalization is implemented between the wave semantic interaction block and the channel learning block to stably accelerate the model training process and improve the generalization ability of the model.

[0082] Furthermore, the output of the wave semantic interaction block is residually connected with the output of the channel learning block. This solves the gradient vanishing and gradient exploding problems in deep neural networks and improves the performance and depth of the model.

[0083] Preferably, the prediction model has four first modules connected in sequence.

[0084] S31 performs global average pooling and full connection layer processing on the output data of the last first module to obtain the total features.

[0085] Through global average pooling and multi-stream interaction aggregation of fully connected layers, more accurate and comprehensive future trajectories are output.

[0086] S32 inputs the total features into the Laplace decoder to predict the trajectory of the target vehicle in the future time period.

[0087] Considering the multimodality of vehicle motion, this step uses Laplace distribution to model vehicle trajectories. For each prediction instance, the Laplace decoder outputs a set of trajectories Among them, k represents the kth mode, K represents the number of predicted modes, π i,k represents the mixing coefficient corresponding to the k-th Laplace density of the i-th target vehicle, i represents the identification number of the target vehicle, t represents the time, t f Represents the forecast time period, Laplace() represents the Laplace function, represents the predicted position of the i-th target vehicle at the t-th moment in the prediction time period, It represents the deviation of the i-th target vehicle at the t-th moment in the prediction time period.

[0088] Compared with the existing technology, this embodiment provides a vehicle trajectory prediction method based on a wave semantic network, which introduces "waves" to characterize complex traffic scenes and uses the superposition of waves to implicitly model multi-flow social interactions, so that the model can have a deeper understanding of the structure and dynamic changes of complex traffic scenes, and improve the ability to predict vehicle trajectories in complex traffic scenes. By deeply fusing wave semantic information in the spatial and channel dimensions through the wave semantic interaction block, the relationship between global and local features can be effectively captured. By introducing the channel fusion module, the model can efficiently aggregate complementary information between different channels, improve the richness of feature expression, and significantly enhance the abstraction and understanding of scene semantics, thereby improving the overall performance of vehicle trajectory prediction.

[0089] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A vehicle trajectory prediction method based on wave semantic network, characterized in that: include: Obtaining a high-precision map centered on the target vehicle, the high-precision map having multiple semantic channels, the semantic channels including a drivable area, lane boundaries, and traffic lights; The semantic channel stores historical trajectory data of the target vehicle and other vehicles; Dividing the high-precision map into a plurality of data blocks equal in number to the semantic channels; Inputting a data block into a prediction model to predict the trajectory of a target vehicle in a future time period, the prediction model comprising a plurality of first modules, a global average pooling layer and a fully connected layer, and a Laplace decoder connected in sequence, wherein the first module comprises a fully connected layer, a wave semantic interaction block, and a channel learning block connected in sequence, and the execution process of the prediction model comprises: inputting the data block into a first first module and inputting the output result into a next first module; performing global average pooling and fully connected layer processing on the output data of the last first module to obtain a total feature; and inputting the total feature into a Laplace decoder to predict the trajectory of the target vehicle in the future time period; The execution process of the first module includes: inputting input data into the fully connected layer to obtain embedded features; inputting the embedded features into the wave semantic interaction block to obtain time domain features, including: using the first multi-layer perceptron and the second multi-layer perceptron to process the embedded features respectively to obtain the first amplitude and the first phase; based on the first amplitude and the first phase, fusing the waves corresponding to any two embedded features to obtain the second amplitude and the second phase; obtaining corresponding synthetic wave data based on the second amplitude and the second phase; using the third multi-layer perceptron and the fourth multi-layer perceptron to process the synthetic wave data respectively to obtain the first data and the second data; performing global wave fusion processing on the first data to obtain the first feature; performing local wave fusion processing on the second data to obtain the second feature; obtaining time domain features based on the first feature and the second feature; inputting the time domain features into the channel learning block to obtain channel learning features, including: converting the time domain features to obtain frequency domain features; using the multi-layer perceptron to process the frequency domain features, and performing inverse transformation on the processed features to obtain channel learning features.

2. The vehicle trajectory prediction method based on wave semantic network according to claim 1 is characterized in that: Performing global wave fusion processing on the first data to obtain a first feature includes: The first data is divided along the channel dimension and the width dimension respectively to obtain multiple first data segments and multiple second data segments; each first data segment is permuted in the height dimension and the channel dimension to obtain first permuted data; each second data segment is permuted in the width dimension and the channel dimension to obtain second permuted data; the first permuted data and the second permuted data are spliced ​​along the channel dimension respectively, and the spliced ​​data are fused to obtain first fused data and second fused data; the first data is processed by using a fully connected layer to obtain third fused data; learning attention parameters of the first fused data, the second fused data and the third fused data are determined, and the first fused data, the second fused data and the third fused data are fused based on the learning attention parameters to obtain a first feature.

3. The vehicle trajectory prediction method based on wave semantic network according to claim 1 is characterized in that: The first multi-layer perceptron and the second multi-layer perceptron are used to process the embedded features of the block input respectively to obtain the first amplitude and the first phase, which are expressed as follows: z i =MLP1(h i ,W i 0 ),i∈[1,2,...,m] θ i =MLP2(h i ,W i 1 ),i∈[1,2,...,m] Among them, z i represents the first amplitude corresponding to the i-th embedded feature, i represents the identification number of the embedded feature, MLP1 represents the first multi-layer perceptron, and h i represents the i-th embedding feature, W i 0 represents the amplitude weight corresponding to the i-th embedded feature, m represents the number of embedded features, θ i represents the first phase corresponding to the i-th embedded feature, MLP2 represents the second multi-layer perceptron, and W i 1 represents the phase weight corresponding to the i-th embedded feature.

4. The vehicle trajectory prediction method based on wave semantic network according to claim 3 is characterized in that: Based on the first amplitude and the first phase, the waves corresponding to any two embedded features are fused to obtain the second amplitude and the second phase, which are expressed as follows: i r =θ i +atan2(|z a |⊙sin(θ a -θ i ),|z i |+|z a |⊙cos(θ a -θ i )) Among them, z r represents the second amplitude with identification number r obtained by fusion of the i-th embedded feature and the a-th embedded feature, z a represents the first amplitude corresponding to the a-th embedded feature, θ a represents the first phase corresponding to the ath embedded feature, θ r It represents the second phase with identification number r obtained by fusing the i-th embedded feature with the a-th embedded feature, and atan2() represents the two-parameter inverse tangent function.

5. The vehicle trajectory prediction method based on wave semantic network according to claim 4 is characterized in that: The corresponding synthetic wave data is obtained based on the second amplitude and the second phase, including: in, represents the rth synthetic wave data, b represents an imaginary unit, and n represents the total number of synthetic wave data.

6. The vehicle trajectory prediction method based on wave semantic network according to claim 2 is characterized in that: Determining learning attention parameters of the first fused data, the second fused data, and the third fused data, including: Adding the first fused data, the second fused data, and the third fused data to obtain total fused data; Performing adaptive average pooling processing on the total fusion data to obtain pooled data; The pooled data is processed using a Softmax function to obtain a first learning attention parameter corresponding to the first fusion data, a second learning attention parameter corresponding to the second fusion data, and a third learning attention parameter corresponding to the third fusion data.

7. The vehicle trajectory prediction method based on wave semantic network according to claim 6 is characterized in that: And based on the learned attention parameter, the first fusion data, the second fusion data and the third fusion data are fused to obtain the first feature, which is expressed as: in, represents the first feature, α[0] represents the first learning attention parameter, represents the first fusion data, α[1] represents the second learning attention parameter, represents the second fusion data, α[2] represents the third learning attention parameter, Represents the third fused data.

8. The vehicle trajectory prediction method based on wave semantic network according to claim 1 is characterized in that: Performing local wave fusion processing on the second data to obtain a second feature includes: All the second data are processed based on the Euler formula and quantum measurement method to obtain the second feature, which is expressed as: in, represents the second feature, j represents the identification number of the second data, represents the first weight of the j-th second data, Z j represents the amplitude of the jth second data, θ j represents the phase of the j-th second data, represents the second weight of the j-th second data, and n represents the total number of second data.

9. The vehicle trajectory prediction method based on wave semantic network according to claim 1 is characterized in that: The time domain features are obtained based on the first feature and the second feature, including: Adding the first feature and the second feature to obtain an addition result; The addition result is convolved, and the convolution result is residually connected with the embedded feature to obtain a time domain feature.

10. The vehicle trajectory prediction method based on wave semantic network according to claim 1, characterized in that: The Laplacian decoder includes three fully connected layers, and every two fully connected layers are connected by an activation function and a normalization layer.

Images