Vehicle trajectory prediction method of sparse space-time fusion module

Abstract

The invention provides a vehicle trajectory prediction method of a sparse space-time fusion module, belongs to the technical field of intelligent traffic and automatic driving, and solves the technical problem that a traditional attention mechanism lacks accurate focusing on key interaction features in a dynamic scene. According to the technical scheme, the method comprises the following steps: firstly, acquiring target information from a sensor, and constructing corresponding features by taking the target information as nodes; secondly, lane line and vehicle features are extracted through a sparse space-time attention module combining diffusion convolution and tan-sigmoid; and finally, predicting a vehicle track through a full connection layer. According to the method, the extracted features are subjected to attention enhancement in space and time dimensions by applying diffusion convolution and efficient additive attention, the precision and real-time performance of vehicle trajectory prediction are improved, the driving path of the vehicle can be mastered, and the method is suitable for automatic driving and intelligent traffic systems.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation and autonomous driving technology, and in particular to a vehicle trajectory prediction method of a sparse spatiotemporal fusion module. Background Art

[0002] With the rapid development of autonomous driving technology, vehicle trajectory prediction plays a vital role in traffic safety and efficient travel. By accurately predicting future vehicle trajectories, autonomous driving systems can plan driving routes in advance, avoid collisions, reduce traffic accidents and improve traffic flow. However, existing vehicle trajectory prediction methods often have difficulty coping with complex traffic scenarios. Especially in high-density traffic environments, conventional spatiotemporal feature extraction methods cannot fully capture the complex dynamic interactions between vehicles, resulting in low prediction accuracy.

[0003] Traditional prediction methods are mostly based on single-scale convolutional neural networks (CNN) or graph convolutional networks (GCN) to extract spatiotemporal features. These methods are prone to problems such as low computational efficiency and insufficient modeling of local spatiotemporal dependencies when dealing with trajectory prediction of multiple target vehicles. Therefore, a vehicle trajectory prediction method that can capture sparse spatiotemporal features is needed. Summary of the invention

[0004] The purpose of the present invention is to provide a vehicle trajectory prediction method of a sparse spatiotemporal fusion module, which can effectively handle the complex interactive relationship between vehicles and lane lines, thereby improving the accuracy and real-time performance of target vehicle trajectory prediction.

[0005] The inventive concept of the present invention is: the present invention provides a vehicle trajectory prediction method of a sparse spatiotemporal fusion module. First, the target information is obtained from the sensor and used as a node to construct the corresponding features; secondly, the lane line and vehicle features are extracted through a sparse spatiotemporal attention module combined with diffuse convolution and tanh-sigmoid, and the TopK algorithm is used for feature screening; at the same time, the screened lane line features and vehicle node features are multimodally fused and modeled through an efficient additive attention module; finally, the vehicle trajectory is predicted through a fully connected layer. The present invention uses diffuse convolution and efficient additive attention to strengthen the attention of the extracted features in spatial and temporal dimensions, which can improve the accuracy and real-time performance of vehicle trajectory prediction, can grasp the vehicle's driving path, and is suitable for autonomous driving and intelligent transportation systems.

[0006] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is specifically: a vehicle trajectory prediction method of a sparse spatiotemporal fusion module, comprising the following steps:

[0007] S1: Get the vehicle's starting and ending positions from the sensor, and get the features of the lane the vehicle is currently in from the map, including lane lines, crosswalks, traffic lights, etc. Construct all features into corresponding feature vectors;

[0008] S2: Construct a diffuse convolution and tanh-sigmoid model, in which the lane and vehicle features are extracted through a sparse spatiotemporal attention module combining a diffuse convolution network and tanh-sigmoid, and the TopK algorithm is used for feature screening to obtain the key spatiotemporal features of the lane and vehicle.

[0009] S3: Build an efficient additive attention module to fuse and model the multimodal features between lanes and lane lines, and capture the complex interactions between different objects.

[0010] S4: Generate the future trajectory position of the vehicle using the fully connected layer.

[0011] Furthermore, the step S1 specifically includes the following steps:

[0012] S2.1: Obtain the driving data of n vehicle objects in each time frame t through sensors, which contains the initial and final positions of the target vehicle and its surrounding vehicles, vehicle identification number, area and other information. Construct the vehicle position information into the corresponding feature vector v i (i=1…n), which includes the initial position information of the vehicle node And the final location information of the vehicle node d i Contains the horizontal and vertical coordinates (x, y) of the vehicle node at each time frame t. The specific formula is expressed as:

[0013]

[0014] S2.2: Get lane information by calling the API through the map file, which includes lane lines, pedestrian crossings, traffic lights, etc. The lane line feature is represented by p i , in preparation for focusing on extracting lane line features and their interactions with vehicle nodes.

[0015] Furthermore, the step S2 specifically includes the following steps:

[0016] S3.1: Construct a diffusion convolution and tanh-sigmoid model to extract the sparse local spatiotemporal correlation features of lane lines and vehicles. Concatenate the vehicle nodes and lane line nodes in S1 as the input of the current module. First, a 1D diffusion convolutional network is used to capture spatiotemporal information, where the convolution kernel size is 3 and the expansion rate is 2. Second, a spatial and temporal attention module based on the tanh-sigmoid combination is used to better model the spatial and temporal features of lane lines and vehicle features. After the two features are diffused, the spatial attention weights are generated through the nonlinear operation of the tanh activation function and the sigmoid function. and temporal attention weights The formula is:

[0017]

[0018] in, and Denote the attention weights generated by the spatial and temporal attention mechanisms, respectively. The formula uses a combination of diffuse convolution and nonlinear activation function to perform an operation on the input feature vector v i and its transpose for weighted modeling and feature extraction, aiming to capture the feature interaction relationship in spatial and temporal dimensions. First, DiffusionConv1D is a convolution operation based on the diffusion process, which is used to diffuse and aggregate the input feature viv_ivi in ​​one-dimensional space, and can effectively capture local spatial or temporal information. In the formula, mean(v i )and Respectively represent the input vector v i and its transpose The mean operation is performed to smooth the features and reduce noise. Subsequently, the diffuse convolution result is transformed nonlinearly through the tanh and sigmoid activation functions to further enhance the network's expressiveness. The tanh function is used to maintain the smoothness of the output while introducing nonlinear features with a mapping range of [-1,1]; while the sigmoid function compresses the output value to the range of [0,1], which has the effect of normalization and weighting, thereby generating spatial attention weights. and temporal attention weights

[0019] In the entire formula, the multiplication operation represents the element-by-element weighting of the features after the diffusion convolution processing. Through the combination of two nonlinear activation functions, the representation of the features in the spatial and temporal domains is enriched, which can effectively model the spatiotemporal correlation between vehicle and lane features.

[0020] S3.2: Build a TopK screening model, select the most important K features, where K is 5, and retain the key features. The formula is:

[0021]

[0022] Among them, TopK_Selection represents the input feature vector v i Perform Top-K screening to extract the most representative and important k feature components, and the output is Among them, TopK is a selection function that sets the threshold k and selects the input vector v i The first k elements with the largest weight or importance are selected to retain key features and remove redundant information. i represents the input feature vector, which usually contains multi-dimensional feature information, and k is a parameter set by the user, indicating the number of features that need to be retained. It represents the result vector after TopK screening, which only contains the feature elements with the top k weight values. This operation plays a role in dimensionality reduction and focusing in feature selection. By screening out the most significant features, it can improve the computational efficiency of the model and its sensitivity to key information, while effectively reducing the interference of irrelevant features on subsequent processing. Therefore, this formula describes an efficient screening mechanism for the input feature vector, so that the retained features are concentrated on the most contributing parts, thereby optimizing the overall spatiotemporal feature modeling.

[0023] Furthermore, in step S3, an efficient additive attention module is constructed to input the features of vehicles and lane lines into the query matrix, key matrix and value matrix to perform multimodal feature fusion. The query matrix and key matrix are generated through linear transformation. The scaled attention score A is calculated by the scaling factor. T is 256. The weighted sum of the attention score A and the query matrix is ​​obtained to obtain the global vector G. Finally, the global vector is multiplied by the key matrix to generate the additive attention output. The formula is as follows:

[0024]

[0025] x=Proj(G·K)+Q

[0026] Among them, Q and K represent the query matrix and key matrix generated after feature transformation. is the feature vector after TopK screening, W q and W k is the weight matrix of the linear transformation of the query and key, multiplication operation and Indicates that the input feature vector is linearly mapped to generate the query vector Q and the key vector K. Normalize means normalizing the result to ensure that the feature vector is smoothly distributed within a certain range to improve the stability of the model when calculating the attention score.

[0027] The matrix A is the attention score, which is obtained by Q·W gCalculated, where W g is the weight matrix used for feature scaling. The attention score is further scaled by the scaling factor , where T represents the dimension of the vector, which is used to prevent the vector dimension from being too high, resulting in unstable numerical calculations. Subsequently, the matrix G represents the global attention vector, which is processed by weighting A i With the query vector Q i Multiply and sum to get, A i With Q i The summation formula represents the weighted aggregation of all feature points to obtain the global feature representation.

[0028] Finally, the global feature G·K is projected through the Proj operation (i.e., linear mapping function), and then added to the original query vector Q to generate the final feature x. This process not only retains the original query feature information, but also introduces the contextual interaction relationship of the key-value matrix through the global attention mechanism, realizing the fusion and enhancement of features in spatial and temporal dimensions, thereby better modeling the spatiotemporal correlation between vehicles and lanes.

[0029] Furthermore, in step S4, the spatiotemporal features of the vehicle are regressed through a fully connected layer to generate the future trajectory coordinates of the vehicle, and the formula is as follows:

[0030] y=mlp(x)

[0031] MLP stands for Multi-Layer Perceptron, a classic feedforward neural network used to perform nonlinear mapping and feature transformation on the input feature vector x. The input x is usually a feature representation generated after the previous steps, including the result of spatiotemporal feature fusion. The multi-layer perceptron extracts the deep features of the input data layer by layer through a combination of a series of linear transformations (such as matrix multiplication) and nonlinear activation functions (such as ReLU, Sigmoid, etc.), and the final output is y, which is the prediction result or target feature representation learned by the network.

[0032] The core function of the multi-layer perceptron is to introduce nonlinear transformation capabilities, break the limitations of linear models, and enable the model to fit complex data distribution and feature relationships. In addition, when processing, MLP performs weighted summation of the input feature x through the hidden layer weight W and the bias term b, and then performs nonlinear mapping through the activation function, thereby enhancing the model's ability to express the input features. The final output y represents the result after the multi-layer perceptron transformation, which can be used as a direct output for subsequent tasks (such as trajectory prediction) or input into other modules for further processing.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. The vehicle trajectory prediction method of a sparse spatiotemporal fusion module proposed in the present invention significantly improves the accuracy and real-time performance of vehicle trajectory prediction by introducing a sparse spatiotemporal fusion module, combining diffuse convolution and efficient additive attention. Specifically, the sparse spatiotemporal attention module combining diffuse convolution and tanh-sigmoid can effectively extract the key spatiotemporal features of lane lines and vehicles, reduce the impact of redundant information, and improve the efficiency of feature screening. The TopK algorithm is used to screen features, which further improves the computational efficiency of the model.

[0035] 2. The efficient additive attention module of the present invention integrates the multimodal features of vehicles and lane lines into a model, which can accurately capture the complex interactive relationships between different objects, especially in the trajectory prediction scenario of multiple target vehicles, showing higher prediction accuracy.

[0036] 3. The present invention uses a fully connected layer to predict the trajectory of the fused features, which can dynamically adapt to changes in complex traffic environments and ensure that the vehicle trajectory prediction results are more stable and reliable. Therefore, the present invention is widely applicable to autonomous driving and intelligent transportation systems, and is of great significance in improving road safety, traffic efficiency, and reducing accident risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0038] Figure 1 The present invention is a flowchart of a vehicle trajectory prediction method of a sparse spatiotemporal fusion module according to an embodiment of the present invention.

[0039] Figure 2 Schematic diagram of a sparse spatiotemporal fusion module according to an embodiment of the present invention.

[0040] Figure 3 Schematic diagram of an efficient additive attention module according to an embodiment of the present invention.

[0041] Figure 4 Schematic diagram of vehicle trajectory prediction according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0043] Example 1

[0044] See also Figures 1 to 4The technical solution provided in this embodiment is a vehicle trajectory prediction method of a sparse spatiotemporal fusion module. The overall process of the method is as shown in the attached Figure 1 The technical solution of this embodiment is further described below in conjunction with specific embodiments:

[0045] S1: Get the vehicle's starting and ending positions from the sensor, and get the features of the lane the vehicle is currently in from the map, including lane lines, crosswalks, traffic lights, etc. Construct all features into corresponding feature vectors;

[0046] S2: Construct a diffuse convolution and tanh-sigmoid model, in which the lane and vehicle features are extracted through a sparse spatiotemporal attention module combining a diffuse convolution network and tanh-sigmoid, and the TopK algorithm is used for feature screening to obtain the key spatiotemporal features of the lane and vehicle.

[0047] S3: Build an efficient additive attention module to fuse and model the multimodal features between lanes and lane lines, and capture the complex interactions between different objects.

[0048] S4: Generate the future trajectory position of the vehicle using the fully connected layer.

[0049] Specifically, the step S1 includes the following steps:

[0050] S2.1: Obtain the driving data of n vehicle objects in each time frame t through sensors, which contains the initial and final positions of the target vehicle and its surrounding vehicles, vehicle identification number, area and other information. Construct the vehicle position information into the corresponding feature vector v i (i=1…n), which includes the initial position information of the vehicle node And the final location information of the vehicle node d i Contains the horizontal and vertical coordinates (x, y) of the vehicle node at each time frame t. The specific formula is expressed as:

[0051]

[0052] S2.2: Get lane information by calling the API through the map file, which includes lane lines, crosswalks, and traffic lights. The lane line feature is represented by p i , in preparation for focusing on extracting lane line features and their interactions with vehicle nodes.

[0053] Specifically, the step S2 includes the following steps: Figure 2 As shown:

[0054] S3.1: Construct a diffusion convolution and tanh-sigmoid model to extract the sparse local spatiotemporal correlation features of lane lines and vehicles. Concatenate the vehicle nodes and lane line nodes in S1 as the input of the current module. First, a 1D diffusion convolutional network is used to capture spatiotemporal information, where the convolution kernel size is 3 and the expansion rate is 2. Second, a spatial and temporal attention module based on the tanh-sigmoid combination is used to better model the spatial and temporal features of lane lines and vehicle features. After the two features are diffused, the spatial attention weights are generated through the nonlinear operation of the tanh activation function and the sigmoid function. and temporal attention weights The formula is:

[0055]

[0056] in, and Denote the attention weights generated by the spatial and temporal attention mechanisms, respectively. The formula uses a combination of diffuse convolution and nonlinear activation function to perform an operation on the input feature vector v i and its transpose for weighted modeling and feature extraction, aiming to capture the feature interaction relationship in spatial and temporal dimensions. First, DiffusionConv1D is a convolution operation based on the diffusion process, which is used to diffuse and aggregate the input feature viv_ivi in ​​one-dimensional space, and can effectively capture local spatial or temporal information. In the formula, mean(v i )and Respectively represent the input vector v i and its transpose The mean operation is performed to smooth the features and reduce noise. Subsequently, the diffuse convolution result is transformed nonlinearly through the tanh and sigmoid activation functions to further enhance the network's expressiveness. The tanh function is used to maintain the smoothness of the output while introducing nonlinear features with a mapping range of [-1,1]; while the sigmoid function compresses the output value to the range of [0,1], which has the effect of normalization and weighting, thereby generating spatial attention weights. and temporal attention weights

[0057] In the entire formula, the multiplication operation represents the element-by-element weighting of the features after the diffusion convolution processing. Through the combination of two nonlinear activation functions, the representation of the features in the spatial and temporal domains is enriched, which can effectively model the spatiotemporal correlation between vehicle and lane features.

[0058] S3.2: Build a TopK screening model, select the most important K features, where K is 5, and retain the key features. The formula is:

[0059]

[0060] Among them, TopK_Selection represents the input feature vector v i Perform Top-K screening to extract the most representative and important k feature components, and the output is Among them, TopK is a selection function that sets the threshold k and selects the input vector v i The first k elements with the largest weight or importance are selected to retain key features and remove redundant information. i represents the input feature vector, which usually contains multi-dimensional feature information, and k is a parameter set by the user, indicating the number of features that need to be retained. It represents the result vector after TopK screening, which only contains the feature elements with the top k weight values. This operation plays a role in dimensionality reduction and focusing in feature selection. By screening out the most significant features, it can improve the computational efficiency of the model and its sensitivity to key information, while effectively reducing the interference of irrelevant features on subsequent processing. Therefore, this formula describes an efficient screening mechanism for the input feature vector, so that the retained features are concentrated on the most contributing parts, thereby optimizing the overall spatiotemporal feature modeling.

[0061] Specifically, in step S3, an efficient additive attention module is constructed to input the features of vehicles and lane lines into the query matrix, key matrix and value matrix to perform multimodal feature fusion. The query matrix and key matrix are generated through linear transformation. The scaled attention score A is calculated through the scaling factor. T is 256. The weighted sum of the attention score A and the query matrix is ​​obtained to obtain the global vector G. Finally, the global vector is multiplied by the key matrix to generate the additive attention output, as shown in the attached figure. Figure 3 As shown, the formula is as follows:

[0062]

[0063] x=Proj(G·K)+Q

[0064] Among them, Q and K represent the query matrix and key matrix generated after feature transformation. is the feature vector after TopK screening, W q and W k is the weight matrix of the linear transformation of the query and key, multiplication operation and Indicates that the input feature vector is linearly mapped to generate the query vector Q and the key vector K. Normalize means normalizing the result to ensure that the feature vector is smoothly distributed within a certain range to improve the stability of the model when calculating the attention score.

[0065] The matrix A is the attention score, which is obtained by Q·W g Calculated, where W g is the weight matrix used for feature scaling. The attention score is further scaled by the scaling factor , where T represents the dimension of the vector, which is used to prevent the vector dimension from being too high, resulting in unstable numerical calculations. Subsequently, the matrix G represents the global attention vector, which is processed by weighting A i With the query vector Q i Multiply and sum to get, A i With Q i The summation formula represents the weighted aggregation of all feature points to obtain the global feature representation.

[0066] Finally, the global feature G·K is projected through the Proj operation (i.e., linear mapping function), and then added to the original query vector Q to generate the final feature x. This process not only retains the original query feature information, but also introduces the contextual interaction relationship of the key-value matrix through the global attention mechanism, realizing the fusion and enhancement of features in spatial and temporal dimensions, thereby better modeling the spatiotemporal correlation between vehicles and lanes.

[0067] Specifically, in step S4, the spatiotemporal features of the vehicle are regressed through the fully connected layer to generate the future trajectory coordinates of the vehicle. The result is shown in the attached figure. Figure 4 As shown, the red dots represent vehicles, the blue dots represent the actual trajectories of vehicles, the green trajectories represent the predicted trajectories, and the red trajectories are the trajectories with the highest probability. The formula is as follows:

[0068] y=mlp(x)

[0069] MLP stands for Multi-Layer Perceptron, a classic feedforward neural network used to perform nonlinear mapping and feature transformation on the input feature vector x. The input x is usually a feature representation generated after the previous steps, including the result of spatiotemporal feature fusion. The multi-layer perceptron extracts the deep features of the input data layer by layer through a combination of a series of linear transformations (such as matrix multiplication) and nonlinear activation functions (such as ReLU, Sigmoid, etc.), and the final output is y, which is the prediction result or target feature representation learned by the network.

[0070] The core function of the multi-layer perceptron is to introduce nonlinear transformation capabilities, break the limitations of linear models, and enable the model to fit complex data distribution and feature relationships. In addition, when processing, MLP performs weighted summation of the input feature x through the hidden layer weight W and the bias term b, and then performs nonlinear mapping through the activation function, thereby enhancing the model's ability to express the input features. The final output y represents the result after the multi-layer perceptron transformation, which can be used as a direct output for subsequent tasks (such as trajectory prediction) or input into other modules for further processing.

[0071] Example 2

[0072] On the basis of Example 1, in order to verify the performance of the model, this example selects TPCN, SceneTransformer, HOME+GOHOME and Multipath++ models as comparative experimental models. We use input data constructed with a single target vehicle as the center of the scene, and verify the proposed model on the training set and test set of Argoverse. Compared with other models, the official indicators minADE6 and minADE6 perform better, especially in the average displacement error, showing a higher prediction accuracy. In addition, although it lags slightly behind Multipath++ in the b-minFDE6 indicator, it may be due to the detailed differences in different model architectures or optimization strategies, but overall it is close to other models in the prediction accuracy of the optimal trajectory point. Although our model lags slightly behind HOME+GOHOME in MR6, the gap is small, which may be a slight difference in prediction strategy. The results are shown in Table 1 below.

[0073] Table 1 Single target trajectory prediction performance index table of Argoverse dataset

[0074]

[0075] Example 3

[0076] Based on Example 2, in order to verify the multi-objective performance of the model, this paper selected mmTrans, LaneGCN, LTP and DenseTNT models as comparative experiments. We used input data constructed with multi-objective vehicles as the center of the scene, and verified the proposed model on the training set and test set of Argoverse. Compared with other models, our model performed better in the official indicators minADE6 and minFDE6, and it showed a certain degree of advancement.

[0077] Table 2 Multi-target trajectory prediction performance indicators of Argoverse dataset

[0078]

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A vehicle trajectory prediction method based on a sparse spatiotemporal fusion module, characterized in that: The following steps are involved: S1: Obtain the vehicle's starting and ending positions from the sensor, and obtain the features of the lane where the vehicle is currently located from the map, including lane lines, crosswalks, and traffic lights, and construct all features into corresponding feature vectors; S2: Construct a diffuse convolution and tanh-sigmoid model, in which the lane and vehicle features are extracted through a sparse spatiotemporal attention module combining a diffuse convolution network and tanh-sigmoid, and the TopK algorithm is used for feature screening to obtain the key spatiotemporal features of the lane and vehicle. S3: Build an efficient additive attention module to fuse and model the multimodal features between lanes and lane lines, and capture the complex interactions between different objects. S4: Generate the future trajectory position of the vehicle using the fully connected layer.

2. The vehicle trajectory prediction method of the sparse spatiotemporal fusion module according to claim 1 is characterized in that: The step S1 includes the following steps: S2.1: Obtain driving data of n vehicle objects in each time frame t through sensors, which includes the initial and final positions of the target vehicle and its surrounding vehicles, vehicle identification number, and area information, and construct the vehicle position information into the corresponding feature vector v i (i=1…n), including the initial location information of the vehicle node And the final location information of the vehicle node d i Contains the horizontal and vertical coordinates (x, y) of the vehicle node at each time frame t, and the formula is expressed as: S2.2: Get lane information by calling the API through the map file, which includes lane lines, crosswalks and traffic lights. The lane line features are represented by p i , in preparation for extracting lane line features and their interaction with vehicle nodes.

3. The vehicle trajectory prediction method of the sparse spatiotemporal fusion module according to claim 1 is characterized in that: The step S2 includes the following steps: S3.1: Construct a diffusion convolution and tanh-sigmoid model to extract the sparse local spatiotemporal correlation features of lane lines and vehicles, and concatenate the vehicle nodes and lane line nodes in step S1 as the input of the current module. The following steps are involved: S11, the spatiotemporal information is captured through a 1D diffusion convolutional network, where the convolution kernel size is 3 and the expansion rate is 2. Secondly, the spatial and temporal feature modeling of lane lines and vehicle features is performed using a spatial and temporal attention module based on a tanh-sigmoid combination; after the two features are diffused, the spatial attention weight is generated through the nonlinear operation of the tanh activation function and the sigmoid function. and temporal attention weights The formula is: in, and denotes the attention weights generated by the spatial and temporal attention mechanisms respectively. The formula uses a combination of diffuse convolution and nonlinear activation function to perform an operation on the input feature vector v i and its transpose for weighted modeling and feature extraction, aiming to capture the interactive relationship of features in spatial and temporal dimensions; S12, DiffusionConv1D is a convolution operation based on the diffusion process, which is used to diffuse and aggregate the input feature viv_ivi in ​​one-dimensional space to capture local spatial or temporal information. In the formula, mean(v i )and Respectively represent the input vector v i and its transpose Perform mean operation; S13, the diffuse convolution result is transformed nonlinearly through tanh and sigmoid activation functions. The tanh function is used to maintain the smoothness of the output and introduce nonlinear features. The mapping range is [-1, 1]; the sigmoid function compresses the output value to the range of [0, 1] to generate the spatial attention weight v i s and the temporal attention weight v i t ; In the whole formula, the multiplication operation represents the element-by-element weighting of the features after the diffusion convolution processing, and the spatiotemporal correlation between vehicle and lane features is modeled through the combination of two nonlinear activation functions; S3.2: Build a TopK screening model, select the most important K features, where K is 5, and retain the key features. The formula is: Among them, TopK_Selection represents the input feature vector v i Perform Top-K screening to extract the most representative and important k feature components, and the output is Among them, TopK is a selection function that sets the threshold k and selects the input vector v i Filter out the first k elements with the largest weight or importance to retain key features and remove redundant information; i represents the input feature vector, which contains multi-dimensional feature information, and k is a parameter set by the user, indicating the number of features to be retained. It means the result vector after TopK screening contains only the feature elements with the top k weight values. This operation plays the role of dimensionality reduction and focusing in feature selection, and screens out the most significant features.

4. The vehicle trajectory prediction method of the sparse spatiotemporal fusion module according to claim 1 is characterized in that: In step S3, an efficient additive attention module is constructed to input the features of vehicles and lane lines into the query matrix, the key matrix and the value matrix to perform multimodal feature fusion; wherein the filtered input The query matrix and key matrix are generated through linear transformation. The scaled attention score A is calculated by the scaling factor. T is 256. The weighted sum of the attention score A and the query matrix is ​​obtained to obtain the global vector G. Finally, the global vector is multiplied by the key matrix to generate the additive attention output. The formula is as follows: x=Proj(G·K)+Q Among them, Q and K represent the query matrix and key matrix generated after feature transformation, respectively. is the feature vector after TopK screening, W q and W k is the weight matrix of the linear transformation of the query and key, multiplication operation and Indicates that the input feature vector is linearly mapped to generate the query vector Q and the key vector K; normalize means normalizing the result; The matrix A is the attention score, which is obtained by Q·W g Calculated, where W g is the weight matrix for feature scaling, and the attention score is further scaled by the scaling factor , where T represents the dimension of the vector, which is used to prevent the vector dimension from being too high, resulting in unstable numerical calculations; then, the matrix G represents the global attention vector, which is processed by weighting A i With the query vector Q i Multiply and sum to get, A i With Q i The summation formula represents weighted aggregation of all feature points to obtain global feature representation; Finally, the global feature G·K is projected through the Proj operation and added to the original query vector Q to generate the final feature x.

5. The vehicle trajectory prediction method of the sparse spatiotemporal fusion module according to claim 1 is characterized in that: In step S4, the spatiotemporal features of the vehicle are regressed through the fully connected layer to generate the future trajectory coordinates of the vehicle. The formula is as follows: y=mlp(x) Among them, MLP stands for multi-layer perceptron, which is a classic feedforward neural network used to perform nonlinear mapping and feature transformation on the input feature vector x. The input x is the feature representation generated after processing in the previous steps, including the result of spatiotemporal feature fusion. The multi-layer perceptron extracts the deep features of the input data layer by layer through a combination of a series of linear transformations and nonlinear activation functions, and the final output is y, which is the prediction result or target feature representation learned by the network.

Images