Vehicle following behavior prediction method and electronic equipment

Abstract

The invention provides a vehicle following behavior prediction method and electronic equipment, and belongs to the technical field of intelligent traffic and vehicle automatic driving, and the method comprises the steps: firstly extracting the local spatial-temporal characteristics of a vehicle track through a local region self-attention mechanism; then capturing a long-term dependency relationship in the car-following behavior by using an XLSTM module, especially a matrix memory of an mLSTM unit of the XLSTM module; and finally, the multi-scale features are adaptively weighted and fused through a dynamic feature fusion module, and optimal feature representation is generated for final prediction. The method effectively solves the problems that a traditional model cannot describe heterogeneity of driving behaviors sufficiently and is weak in long-term dependence capture capability, and the accuracy of vehicle following behavior prediction is remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and autonomous driving technology, and in particular to a method and electronic device for predicting vehicle following behavior based on the fusion of local area self-attention and dynamic features. Background Technology

[0002] With the continuous growth of road traffic flow and the significant increase in the complexity of traffic scenarios, the driving behavior exhibited by individual drivers is highly heterogeneous. Traditional rule-based car-following models struggle to adapt to this complexity, while existing deep learning models (such as LSTM) do not adequately explore the local interaction relationships between vehicles and cannot dynamically adjust feature weights, resulting in limited prediction accuracy. Chinese patent document CN116611327A provides a car-following model calibration method, but it suffers from rigid driving behavior classification, difficulty in handling long-term dependencies, and the inability to dynamically adjust feature weights. Therefore, there is an urgent need in this field for a vehicle car-following behavior prediction scheme that can more accurately capture the heterogeneity of driving behavior, effectively process long-sequence data, and adaptively fuse key features. Summary of the Invention

[0003] In view of the technical defects and drawbacks existing in the prior art, the present invention provides a vehicle following behavior prediction method and electronic device to overcome the above problems or at least partially solve the above problems, the specific solution of which is as follows;

[0004] As a first aspect of the present invention, a method for predicting vehicle following behavior is provided, comprising the following steps:

[0005] Local feature extraction stage: The input vehicle trajectory time series data is divided into overlapping local regions. In each local region, a multi-head self-attention mechanism is applied to extract local feature representations that reflect the heterogeneity of driving behavior.

[0006] Temporal dependency modeling stage: The obtained local feature sequence is input into the XLSTM module, and the long-term dependencies in the trajectory data are captured through matrix memory units;

[0007] Multi-scale feature fusion stage: Dynamic feature fusion is performed on the multi-scale features output by the XLSTM module to generate an optimized fused feature representation;

[0008] Behavior prediction output stage: Based on the obtained fused feature representation, output the prediction results of vehicle following behavior.

[0009] Furthermore, the process of dividing the input vehicle trajectory time series data into overlapping local regions and applying a multi-head self-attention mechanism within each local region to extract local feature representations reflecting the heterogeneity of driving behavior specifically includes:

[0010] Input vehicle trajectory time series data According to the preset local region size p s The patch sequence, which overlaps with the sliding step size, is obtained by dividing the data into M local regions. Each local region represents the original input. ;

[0011] For each local region Perform layer normalization preprocessing to obtain the transformation preparation state. This preparation state is used for subsequent linear transformations and attention calculations;

[0012] Based on the transformation preparation state By querying the weight matrix respectively Key weight matrix Value weight matrix Perform a linear transformation to generate the query matrix Q= Bond matrix K= Sum matrix V= ;

[0013] A multi-head self-attention mechanism is used to calculate the Q, K, and V matrices to obtain the attention-weighted feature representation:

[0014] ;

[0015] Multi-head attention output Local feature enhancement is performed using a ConvFFN convolutional feedforward network to obtain optimized local feature representations.

[0016] Among them, This represents the total length of the time series. p represents the dimension of the feature vector. s This represents the time step size contained in each local region, where step represents the sliding step size, M is the number of local regions obtained after partitioning, and d q d k d v These represent the dimensions of the query, key, and value vector, respectively.

[0017] Furthermore, the XLSTM module uses mLSTM as its basic processing unit, and its memory unit is a matrix. Information storage and updating are achieved through the following formula:

[0018] Memory unit update formula: ;

[0019] in, This represents the memory cell matrix at the current time step, with dimensions d×d. This represents the memory matrix of the previous time step. Let k be the value vector at the current time step. t The key vector at the current time step;

[0020] Hidden state calculation formula:

[0021] in, Let n be the query vector at the current time step. t is the normalization factor vector, and l is the preset minimum threshold parameter.

[0022] Furthermore, the computation process of the mLSTM unit includes the following steps:

[0023] After performing layer normalization on the input data, the feature dimensions are expanded through an up projection operation, and the expanded features are divided into left and right paths for parallel processing.

[0024] The left-side features are first processed by causal convolution, then nonlinear transformation is performed using the Swish activation function, and then the processed features are input into the projection matrix and then processed by the mLSTM unit.

[0025] The output of the mLSTM unit is processed by group normalization, and the output after group normalization is residually connected with the result of the left-path feature after Swish activation.

[0026] The output after residual connection is multiplied by the features after right-side up-projection processing to achieve interaction and fusion between features;

[0027] The output after the dot product operation is down-projected to map the high-dimensional features back to the target dimension;

[0028] The output after downprojection is residually concatenated with the original input to obtain the final output of the mLSTM unit.

[0029] Furthermore, dynamic feature fusion specifically includes:

[0030] Feature sequences at different levels The concatenation operation is performed along the feature dimension to obtain the fused feature X, where t represents the number of feature levels, calculated using the following formula:

[0031]

[0032] Global average pooling is performed on the concatenated feature sequence X to extract global channel statistics. Then, channel weight vectors are generated through a fully connected layer and a sigmoid activation function. The calculation formula is:

[0033]

[0034] Generate channel weights The original concatenated feature X is multiplied element-wise to achieve channel-level feature calibration, and then the dimension is restored through a fully connected layer to obtain the channel-calibrated feature. The calculation formula is:

[0035] .

[0036] Furthermore, the dynamic feature fusion also includes the generation and fusion of time step weights:

[0037] For each level of features One-dimensional convolutional layers are applied to extract features at each time step to capture the local dependencies of each feature sequence in the time dimension.

[0038] The outputs of each convolutional layer are added element by element to fuse the time step feature information from different layers.

[0039] Apply the Sigmoid activation function to the fused temporal feature representation to generate global time step weights w. ts The calculation formula is:

[0040]

[0041] The channel-calibrated feature X ch With global time step weight w ts Element-wise multiplication is performed to achieve dynamic fusion of spatiotemporal features, resulting in the final optimized feature representation. The calculation formula is:

[0042] .

[0043] Furthermore, the local region self-attention mechanism achieves refined modeling of heterogeneous driving behavior in the following ways:

[0044] By dividing the vehicle trajectory time series data into overlapping local regions, and independently applying a multi-head self-attention mechanism in each local region, the system captures the short time distance and high acceleration characteristics of aggressive driving behavior during car following, the long time distance and smooth acceleration and deceleration characteristics of conservative driving behavior, and the moderate car following characteristics of normal driving behavior.

[0045] By utilizing the ability to capture long-distance dependencies within local regions, we can distinguish the feature differences of different driving behavior types at local time scales. Among them, aggressive driving behavior shows higher attention weight fluctuations in local regions, conservative driving behavior shows a more uniform attention distribution, and ordinary driving behavior is between aggressive and conservative driving behavior.

[0046] The ConvFFN convolutional feedforward network is used to enhance local features, thereby improving the ability to extract local features for different driving behavior types.

[0047] Furthermore, the XLSTM module captures long-term dependencies using the following method:

[0048] Matrix memory storage structure using mLSTM cells This enables parallel storage and processing of multi-dimensional feature relationships in vehicle trajectory data, where feature information such as speed, acceleration, and relative distance are stored collaboratively in matrix form.

[0049] Through matrix key-value pairs (k t ,v t The parallel update mechanism synchronously processes long-term dependencies between multiple feature dimensions, where the key vector k t Responsible for feature selection, value vector v t Responsible for feature storage and query vectors Implement feature-based retrieval;

[0050] By employing extended memory capacity and parallel computing architecture, it effectively captures multi-scale temporal dependencies from short-term to long-term in car-following behavior, solving the gradient vanishing or exploding problem of traditional LSTM models when processing long-sequence vehicle trajectory data.

[0051] By using a hybrid stacked architecture of sLSTM and mLSTM, combining the advantages of local feature interaction and global memory storage, accurate modeling of long-term car-following patterns in heterogeneous driving behaviors can be achieved.

[0052] Furthermore, the prediction result of vehicle following behavior based on the obtained fused feature representation specifically includes:

[0053] The obtained fusion feature representation The input is fed into a fully connected layer for feature dimension transformation, mapping it to a feature space that matches the dimension of the prediction target, where the prediction target includes the speed v of the following vehicle. t+1 acceleration a t+1 and the relative distance Δd to the vehicle in front. t+1 At least one parameter;

[0054] A time-distributed fully connected layer is used to perform temporal decoding on the transformed features, generating a vehicle trajectory prediction sequence within the future time interval [t+1, t+τ] step by step, where τ is the prediction time span, which is dynamically adjusted according to the complexity of the car-following scenario;

[0055] The decoded features are normalized by the output layer, and continuous numerical prediction results are generated by using a linear activation function or discrete driving behavior classification probabilities are generated by using a Softmax activation function. The driving behavior classification includes at least three types: aggressive, conservative, and normal.

[0056] Supervised training is performed on the predicted results and actual trajectory data using the mean squared error loss function or cross-entropy loss function. The model parameters are optimized using the gradient descent algorithm to make the predicted results approximate the actual car-following behavior data as closely as possible. The calculation formula is as follows:

[0057]

[0058] in For the true value, is the predicted value, and N is the number of training samples.

[0059] As a second aspect of the present invention, an electronic device is provided, comprising:

[0060] One or more processors;

[0061] Memory, used to store one or more programs;

[0062] When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described above.

[0063] The present invention has the following beneficial effects:

[0064] This invention patent utilizes a vehicle car-following model based on local region self-attention and dynamic feature fusion to predict vehicle trajectory data. The local region self-attention module refines the local features of different driving behavior types, effectively distinguishing the differences in local features between aggressive, conservative, and normal driving behaviors, thus improving the accuracy of car-following predictions for each type of driving behavior. An XLSTM model is used to handle long-term dependencies in long sequences of trajectory data. Dynamic feature fusion adaptively selects and fuses multi-scale features, dynamically adjusting the importance weight of each feature based on global information of the car-following task. Attached Figure Description

[0065] Figure 1 A flowchart illustrating a vehicle following behavior prediction method provided in an embodiment of the present invention;

[0066] Figure 2 This invention provides a local region self-attention structure diagram.

[0067] Figure 3 A diagram of an mLSTM structure provided in an embodiment of the present invention;

[0068] Figure 4 This is a dynamic feature fusion structure diagram provided in an embodiment of the present invention;

[0069] Figure 5 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0070] To enable those skilled in the art to better understand the technical solutions of the present invention, exemplary embodiments of the present invention are described below in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0071] Where there is no conflict, the various embodiments of the present invention and the features thereof may be combined with each other.

[0072] As used herein, the term “and / or” includes any and all combinations of one or more related enumerated entries.

[0073] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used herein, the singular forms “a” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that when the terms “comprising” and / or “made of” are used in this specification, the presence of the stated feature, integral, step, operation, element, and / or component is specified, but the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof is not excluded. Terms such as “connected” or “linked” are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect.

[0074] Unless otherwise specified, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. It will also be understood that terms such as those defined in commonly used dictionaries should be interpreted as having the meaning consistent with their meaning in the context of the relevant art and the invention, and will not be interpreted as having an idealized or overly formal meaning unless expressly so defined herein.

[0075] In the technical solution of this invention, the collection, storage, use, processing, transmission, provision, and disclosure of user personal information all comply with relevant laws and regulations and do not violate public order and good morals. The use of user data in this technical solution follows relevant national laws and regulations (e.g., the "Information Security Technology - Personal Information Security Specification"). For example: appropriate measures are taken for personal information access control; restrictions are imposed on the display of personal information; the purpose of using personal information does not exceed the scope of direct or reasonable association; and explicit identity targeting is eliminated when using personal information to avoid precisely locating a specific individual.

[0076] To address at least one of the technical problems existing in the aforementioned related technologies, the present invention provides a method for predicting vehicle following behavior. Figure 1 A flowchart illustrating a vehicle following behavior prediction method provided in an embodiment of the present invention includes the following steps:

[0077] S1. The input vehicle trajectory time series data is divided into overlapping local regions. A multi-head self-attention mechanism is applied in each local region to extract local feature representations that reflect the heterogeneity of driving behavior.

[0078] S2. Input the local feature sequence obtained in S1 into the XLSTM module, and capture the long-term dependencies in the trajectory data through matrix memory units;

[0079] S3. Perform dynamic feature fusion on the multi-scale features output by the XLSTM module to generate an optimized fused feature representation;

[0080] S4. Based on the fused feature representation obtained in step S3, output the prediction result of vehicle following behavior.

[0081] This invention addresses the systemic deficiency of traditional car-following models in failing to adequately characterize the heterogeneity of driving behavior by employing a process of local feature extraction, temporal dependency modeling, multi-scale feature fusion, and behavior prediction. By integrating heterogeneous driving behavior analysis, long-term dependency capture, and dynamic feature selection into a unified framework, it significantly improves the accuracy and robustness of vehicle trajectory prediction. Furthermore, through a dynamic feature fusion mechanism, the model can adapt to complex and ever-changing traffic scenarios, providing reliable technical support for intelligent transportation systems and autonomous driving decision-making.

[0082] In some embodiments, S1 specifically includes:

[0083] Input vehicle trajectory time series data According to the preset local region size p s The patch sequence, which overlaps with the sliding step size, is obtained by dividing the data into M local regions. Each local region represents the original input. ;

[0084] For each local region Perform layer normalization preprocessing to obtain the transformation preparation state. This preparation state is used for subsequent linear transformations and attention calculations;

[0085] Based on the transformation preparation state By querying the weight matrix respectively Key weight matrix Value weight matrix Perform a linear transformation to generate the query matrix Q= Bond matrix K= Sum matrix V= ;

[0086] A multi-head self-attention mechanism is used to calculate the Q, K, and V matrices to obtain the attention-weighted feature representation:

[0087] ;

[0088] Multi-head attention output Local feature enhancement is performed using a ConvFFN convolutional feedforward network to obtain optimized local feature representations.

[0089] Among them, This represents the total length of the time series. The dimension of the feature vector includes car-following behavior features such as velocity, acceleration, and relative distance, p s The time step represents the size of each local region, which determines the size of the local receptive field. `step` represents the sliding step size, which controls the degree of overlap between local regions. `M` is the number of local regions obtained after segmentation, and `d` is the time step size. q d k d v These represent the dimensions of the query, key, and value vectors, respectively, and are typically set to d. q =d k =dv=d.

[0090] See Figure 2 The diagram shows the self-attention structure of a local region.

[0091] In the above embodiments, by overlapping local region partitioning, the subtle feature differences between different driving styles such as aggressive and conservative driving are effectively captured during the car-following process; by applying a multi-head self-attention mechanism in each local region, the limitations of the local receptive field of traditional CNNs are broken through, and long-distance dependencies on the local temporal scale are established; by decomposing the global attention calculation into multiple parallel subtasks through local region partitioning, the computational complexity is significantly reduced while maintaining model performance; the 3×3 convolutional kernel in ConvFFN further enhances the extraction of local spatiotemporal features and improves the model's sensitivity to capturing dynamic changes in car-following behavior.

[0092] In some embodiments, the XLSTM module uses mLSTM as the basic processing unit, and its memory unit is a matrix. Information storage and updating are achieved through the following formula:

[0093] Memory unit update formula: ;

[0094] in, This represents the memory cell matrix at the current time step, with dimensions d×d. This represents the memory matrix of the previous time step. The value vector at the current time step is obtained from the input features through a linear transformation, k t The key vector at the current time step is obtained by linear transformation of the input features;

[0095] Hidden state calculation formula: ;

[0096] in, Let n be the query vector at the current time step. t is a normalization factor vector used to calculate the stable hidden state. Its value is obtained by linear transformation of the input features at the current time step. l is a preset minimum threshold parameter to ensure that the denominator is not zero.

[0097] In the above embodiments, the mLSTM unit adopts a matrix. As a memory storage structure, compared to the vector-based memory of traditional LSTM, it significantly improves memory capacity and feature relationship modeling capabilities. Here, 'd' represents the dimension of the feature vector. This design supports the collaborative storage and processing of multi-dimensional features such as speed, acceleration, and relative distance in vehicle trajectory data. Memory unit updates employ... The covariance update rule is based on the value vector. and key vector k t The outer product operation enables incremental updates of historical memory, effectively solving the gradient vanishing problem in traditional RNN models for long sequence processing; the hidden state calculation introduces a normalization factor vector n. t and stability constant l, through The formula ensures numerical stability in the calculation process, making it particularly suitable for scenarios where numerical fluctuations may occur in vehicle trajectory data.

[0098] In the specific implementation process, the following preferred scheme is adopted for the calculation of each gating vector:

[0099] In the specific implementation process, the following preferred scheme is adopted for the calculation of each gating vector:

[0100] Query vector generation:

[0101] ;

[0102] in ∈R d×d To query the weight matrix, For bias terms, The input features are for the current time step.

[0103] Key vector calculation:

[0104] ;

[0105] Using scale factor Normalize the key vectors to enhance training stability.

[0106] Value vector transformation:

[0107] ;

[0108] The input features are mapped to a value vector through a linear transformation.

[0109] Specific implementation of gating mechanism

[0110] Input gate control:

[0111] ;

[0112] in An exponential function is used to ensure that the input gate is always positive.

[0113] Forgot Door Configuration:

[0114] Two options are provided:

[0115] Sigmoid function scheme: ;

[0116] Exponential function scheme: ;

[0117] in ;

[0118] Output gate design:

[0119] ;

[0120] in The Sigmoid function is used to control the information output ratio.

[0121] The structure of the mLSTM block is as follows: Figure 3 As shown.

[0122] In some embodiments, the computation process of the mLSTM unit includes the following steps:

[0123] After performing layer normalization on the input data, the feature dimension is expanded through an up projection operation, where the parameter expansion factor is set to 2, and the expanded features are processed in parallel in two paths.

[0124] The left-side features are first processed by a causal convolution with a convolution window size of 4, then the Swish activation function is applied for non-linear transformation, and then the processed features are input into a block object projection matrix with a block size of 4, and then processed by the mLSTM unit.

[0125] The output of the mLSTM unit is processed by group normalization, and the output after group normalization is residually connected with the result of the left-path feature after Swish activation.

[0126] The output after residual connection is multiplied by the features after right-side up-projection processing to achieve interaction and fusion between features;

[0127] The output after the dot product operation is subjected to downprojection, and the expansion factor of the downprojection parameter is set to 2 to map the high-dimensional features back to the target dimension.

[0128] The output after downprojection is residually concatenated with the original input to obtain the final output of the mLSTM unit.

[0129] In the above embodiments, the dual normalization strategy of layer normalization and group normalization effectively suppresses internal covariate shift, accelerates model convergence, and improves training stability. The upprojection operation (parameter expansion factor 2) maps features to a high-dimensional space. Combined with the smooth nonlinear characteristics of the Swish activation function, it significantly improves the model's ability to represent complex car-following patterns. The residual connection mechanism ensures the effective propagation of gradients in deep networks, and the dot product operation realizes deep interaction of left and right path features, forming a multi-level feature fusion architecture. Causal convolution ensures that the model only uses historical information for prediction, which conforms to the temporal causal relationship of car-following behavior and improves reliability in practical applications.

[0130] In some embodiments, dynamic feature fusion is specifically implemented as follows:

[0131] Existing feature fusion methods, such as direct concatenation or simple convolution, lack utilization of global information, resulting in less robust fused features. Dynamic mechanisms, however, can adaptively select important features, thereby improving fusion performance. This patent proposes a dynamic feature fusion module that adaptively fuses local features from multi-scale sequences based on global time-series information. The dynamic feature fusion structure is as follows: Figure 4 As shown, where, Corresponding to Figure 4 F1, F2, ..., Fn in the code, and the final output. correspond Figure 4 The final output F is calculated as follows:

[0132] Feature sequences at different levels The concatenation operation is performed along the feature dimension to obtain the fused feature X, calculated using the following formula:

[0133]

[0134] Global average pooling is performed on the concatenated feature sequence X to extract global channel statistics. Then, channel weight vectors are generated through a fully connected layer and a sigmoid activation function. The calculation formula is:

[0135]

[0136] Generate channel weights The original concatenated feature X is multiplied element-wise to achieve channel-level feature calibration, and then the dimension is restored through a fully connected layer to obtain the channel-calibrated feature. The calculation formula is:

[0137] .

[0138] In the above embodiments, the channel weights are generated through global average pooling and the Sigmoid activation function. This enables intelligent importance allocation across different feature dimensions and channel calibration operations. Effectively integrate multi-scale feature information from local details to global statistics.

[0139] In some embodiments, the dynamic feature fusion further includes the generation and fusion of time step weights:

[0140] For each level of features One-dimensional convolutional layers are applied to extract features at each time step to capture the local dependencies of each feature sequence in the time dimension.

[0141] The outputs of each convolutional layer are added element by element to fuse the time step feature information from different layers.

[0142] Apply the Sigmoid activation function to the fused temporal feature representation to generate global time step weights w. ts The calculation formula is:

[0143]

[0144] The channel-calibrated feature X ch With global time step weight w ts Element-wise multiplication is performed to achieve dynamic fusion of spatiotemporal features, resulting in the final optimized feature representation. The calculation formula is:

[0145] .

[0146] In the above embodiments, parallel one-dimensional convolution operations effectively extract local dependencies at different time scales, enhancing the model's ability to collaboratively perceive short-term car-following behavior fluctuations and long-term trends. The time-step weights generated by Sigmoid activation automatically identify key time nodes (such as emergency braking, acceleration, and overtaking), improving the sensitivity to sudden driving behaviors. This complements the channel weight mechanism, and through dual calibration in both the time and feature dimensions, significantly improves the spatiotemporal consistency of trajectory prediction in complex traffic scenarios.

[0147] In some embodiments, the local region self-attention mechanism described in S1 achieves refined modeling of heterogeneous driving behavior in the following ways:

[0148] By dividing the vehicle trajectory time series data into overlapping local regions, and independently applying a multi-head self-attention mechanism in each local region, the system captures the short time distance and high acceleration characteristics of aggressive driving behavior during car following, the long time distance and smooth acceleration and deceleration characteristics of conservative driving behavior, and the moderate car following characteristics of normal driving behavior.

[0149] By utilizing the ability to capture long-distance dependencies within local regions, we can distinguish the feature differences of different driving behavior types at local time scales. Among them, aggressive driving behavior shows higher attention weight fluctuations in local regions, conservative driving behavior shows a more uniform attention distribution, and ordinary driving behavior is between aggressive and conservative driving behavior.

[0150] In the above embodiments, the ConvFFN convolutional feedforward network is used to enhance local features, thereby strengthening the ability to extract local features for different types of driving behavior. The 3×3 convolutional kernel can effectively capture the differences in local spatiotemporal patterns in car-following behavior. Fine-grained behavioral feature extraction is achieved for the feature differences of three typical driving behaviors at local time scales (such as high acceleration fluctuations in aggressive driving and smooth acceleration and deceleration in conservative driving). By visualizing the attention distribution patterns of different driving behaviors, intuitive decision-making basis is provided for driving behavior analysis, enhancing the transparency of the model.

[0151] In some embodiments, the XLSTM module in S2 captures long-term dependencies using the following method:

[0152] Matrix memory storage structure using mLSTM cells This enables parallel storage and processing of multi-dimensional feature relationships in vehicle trajectory data, where feature information such as speed, acceleration, and relative distance are stored collaboratively in matrix form.

[0153] Through matrix key-value pairs (k t ,v t The parallel update mechanism synchronously processes long-term dependencies between multiple feature dimensions, where the key vector k t Responsible for feature selection, value vector v t Responsible for feature storage and query vectors Implement feature-based retrieval;

[0154] Employing extended memory capacity and parallel computing architecture, it effectively captures multi-scale temporal dependencies in car-following behavior, ranging from short-term (several seconds) to long-term (tens of seconds), solving the gradient vanishing or exploding problem of traditional LSTM models when processing long-sequence vehicle trajectory data.

[0155] By using a hybrid stacked architecture of sLSTM and mLSTM, combining the advantages of local feature interaction and global memory storage, accurate modeling of long-term car-following patterns in heterogeneous driving behaviors can be achieved.

[0156] In the above embodiments, historical information is stored in matrix form, which increases the memory capacity by d times compared with the traditional LSTM (d is the feature dimension), effectively solving the gradient vanishing problem in long sequence prediction; multi-dimensional features such as speed, acceleration, and relative distance are processed synchronously in matrix operations, which enhances the modeling efficiency of collaborative relationships between features. It can maintain stable long-term dependency modeling capabilities for differentiated scenarios such as frequent starts and stops on urban roads and continuous following on highways.

[0157] In some embodiments, the step of outputting a prediction result of vehicle following behavior based on the fused feature representation obtained in step S3 specifically includes:

[0158] The fusion feature representation obtained in S3 The input is fed into a fully connected layer for feature dimension transformation, mapping it to a feature space that matches the dimension of the prediction target, where the prediction target includes the speed v of the following vehicle. t+1 acceleration a t+1 and the relative distance Δd to the vehicle in front. t+1 At least one parameter;

[0159] A time-distributed fully connected layer is used to perform temporal decoding on the transformed features, generating a vehicle trajectory prediction sequence within the future time interval [t+1, t+τ] step by step, where τ is the prediction time span, which is dynamically adjusted according to the complexity of the car-following scenario;

[0160] The decoded features are normalized by the output layer, and continuous numerical prediction results are generated by using a linear activation function or discrete driving behavior classification probabilities are generated by using a Softmax activation function. The driving behavior classification includes at least three types: aggressive, conservative, and normal.

[0161] Supervised training is performed on the predicted results and actual trajectory data using the mean squared error loss function or cross-entropy loss function. The model parameters are optimized using the gradient descent algorithm to make the predicted results approximate the actual car-following behavior data as closely as possible. The calculation formula is as follows:

[0162]

[0163] in For the true value, is the predicted value, and N is the sample size.

[0164] In the above embodiments, numerical trajectory predictions (speed, acceleration, etc.) and categorical behavior judgments are generated simultaneously to meet the multi-dimensional needs of autonomous driving system decision-making and planning; the accuracy of numerical predictions is ensured by the mean square error loss function, while the reliability of behavior classification is optimized by cross-entropy loss, forming a dual quality assurance; and the standardized output format can be directly connected to the vehicle control module, providing a plug-and-play solution for practical applications such as adaptive cruise control and collision warning.

[0165] Based on the same inventive concept, embodiments of the present invention also provide an electronic device. Figure 5 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Figure 5As shown, an embodiment of the present invention provides an electronic device including: one or more processors 101, a memory 102, and one or more I / O interfaces 103. The memory 102 stores one or more programs, which, when executed by the one or more processors, enable the one or more processors to implement an asynchronous tracking method for a vehicle remote control link as described in any of the above embodiments; the one or more I / O interfaces 103 are connected between the processor and the memory, configured to enable information interaction between the processor and the memory.

[0166] The processor 101 is a device with data processing capabilities, including but not limited to a central processing unit (CPU); the memory 102 is a device with data storage capabilities, including but not limited to random access memory (RAM, more specifically SDRAM, DDR, etc.), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), and flash memory (FLASH); the I / O interface (read / write interface) 103 is connected between the processor 101 and the memory 102, and can realize information interaction between the processor 101 and the memory 102, including but not limited to a data bus (Bus).

[0167] In some embodiments, the processor 101, memory 102, and I / O interface 103 are interconnected via bus 104, and thus connected to other components of the computing device.

[0168] In some embodiments, the one or more processors 101 include a field-programmable gate array.

[0169] This invention also provides a computer-readable medium. The computer-readable medium stores a computer program, which, when executed by a processor, implements the steps in any of the asynchronous tracking methods for a vehicle remote control link described in the above embodiments. The computer-readable storage medium may be volatile or non-volatile.

[0170] This invention also provides a computer program product, including computer-readable code, or a non-volatile computer-readable storage medium carrying computer-readable code. When the computer-readable code is run in the processor of an electronic device, the processor in the electronic device executes the above-described asynchronous tracking method for a vehicle remote control link.

[0171] Those skilled in the art will understand that all or some of the steps, systems, and apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software can be distributed on a computer-readable storage medium, which may include computer storage media (or non-transitory media) and communication media (or transient media).

[0172] As is known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable program instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), static random access memory (SRAM), flash memory or other memory technologies, portable compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, it is known to those skilled in the art that communication media typically contain computer-readable program instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0173] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0174] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0175] The computer program product described herein can be implemented specifically through hardware, software, or a combination thereof. In one alternative embodiment, the computer program product is specifically embodied in a computer storage medium; in another alternative embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0176] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0177] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0178] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0179] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction, which contains one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0180] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A method for predicting vehicle following behavior, characterized in that, Includes the following steps: The input vehicle trajectory time series data is divided into overlapping local regions. A multi-head self-attention mechanism is applied to each local region to extract local feature representations that reflect the heterogeneity of driving behavior. The obtained local feature sequence is input into the XLSTM module, and the long-term dependencies in the trajectory data are captured through matrix memory units. Dynamic feature fusion is performed on the multi-scale features output by the XLSTM module to generate an optimized fused feature representation. Based on the obtained fusion feature representation, the prediction results of vehicle following behavior are output.

2. The method according to claim 1, characterized in that, The process involves dividing the input vehicle trajectory time series data into overlapping local regions, and applying a multi-head self-attention mechanism within each local region to extract local feature representations reflecting the heterogeneity of driving behavior. Specifically, this includes: Input vehicle trajectory time series data According to the preset local region size p s The patch sequence, which overlaps with the sliding step size, is obtained by dividing the data into M local regions. Each local region represents the original input. ; For each local region Perform layer normalization preprocessing to obtain the transformation preparation state. This preparation state is used for subsequent linear transformations and attention calculations; Based on the transformation preparation state By querying the weight matrix respectively Key weight matrix Value weight matrix Perform a linear transformation to generate the query matrix Q= The bond matrix K = Sum matrix V= ; A multi-head self-attention mechanism is used to calculate the Q, K, and V matrices to obtain the attention-weighted feature representation: ； Multi-head attention output Local feature enhancement is performed using a ConvFFN convolutional feedforward network to obtain optimized local feature representations. Among them, This represents the total length of the time series. p represents the dimension of the feature vector. s This represents the time step size contained in each local region, where step represents the sliding step size, M is the number of local regions obtained after partitioning, and d q d k d v These represent the dimensions of the query, key, and value vector, respectively.

3. The method according to claim 2, characterized in that, The XLSTM module uses mLSTM as its basic processing unit, and its memory unit is a matrix. Information storage and updating are achieved through the following formula: Memory unit update formula: ; in, This represents the memory cell matrix at the current time step, with dimensions d×d. This represents the memory matrix of the previous time step. Let k be the value vector at the current time step. t The key vector at the current time step; Hidden state calculation formula: ; in, Let n be the query vector at the current time step. t is the normalization factor vector, and l is the preset minimum threshold parameter.

4. The method according to claim 3, characterized in that, The calculation process of the mLSTM unit includes the following steps: After performing layer normalization on the input data, the feature dimensions are expanded through an up projection operation, and the expanded features are divided into left and right paths for parallel processing. The left-side features are first processed by causal convolution, then nonlinear transformation is performed using the Swish activation function, and then the processed features are input into the projection matrix and then processed by the mLSTM unit. The output of the mLSTM unit is processed by group normalization, and the output after group normalization is residually connected with the result of the left-path feature after Swish activation. The output after residual connection is multiplied by the features after right-side up-projection processing to achieve interaction and fusion between features; The output after the dot product operation is down-projected to map the high-dimensional features back to the target dimension; The output after downprojection is residually concatenated with the original input to obtain the final output of the mLSTM unit.

5. The method according to claim 1, characterized in that, The dynamic feature fusion specifically includes: Feature sequences at different levels The concatenation operation is performed along the feature dimension to obtain the fused feature X, where t represents the number of feature levels, calculated using the following formula: ； Global average pooling is performed on the concatenated feature sequence X to extract global channel statistics. Then, channel weight vectors are generated through a fully connected layer and a sigmoid activation function. The calculation formula is: ； Generate channel weights The original concatenated feature X is multiplied element-wise to achieve channel-level feature calibration, and then the dimension is restored through a fully connected layer to obtain the channel-calibrated feature. The calculation formula is: 。 6. The method according to claim 5, characterized in that, The dynamic feature fusion also includes the generation and fusion of time step weights: For each level of features One-dimensional convolutional layers are applied to extract features at each time step to capture the local dependencies of each feature sequence in the time dimension. The outputs of each convolutional layer are added element by element to fuse the time step feature information from different layers. Apply the Sigmoid activation function to the fused temporal feature representation to generate global time step weights w. ts The calculation formula is: ； The channel-calibrated feature X ch With global time step weight w ts Element-wise multiplication is performed to achieve dynamic fusion of spatiotemporal features, resulting in the final optimized feature representation. The calculation formula is: 。 7. The method according to claim 1, characterized in that, The local self-attention mechanism achieves refined modeling of heterogeneous driving behavior in the following ways: By dividing the vehicle trajectory time series data into overlapping local regions, and independently applying a multi-head self-attention mechanism in each local region, the system captures the short time distance and high acceleration characteristics of aggressive driving behavior during car following, the long time distance and smooth acceleration and deceleration characteristics of conservative driving behavior, and the moderate car following characteristics of normal driving behavior. By utilizing the ability to capture long-distance dependencies within local regions, we can distinguish the feature differences of different driving behavior types at local time scales. Among them, aggressive driving behavior shows higher attention weight fluctuations in local regions, conservative driving behavior shows a more uniform attention distribution, and ordinary driving behavior is between aggressive and conservative driving behavior. The ConvFFN convolutional feedforward network is used to enhance local features, thereby improving the ability to extract local features for different driving behavior types.

8. The method according to claim 1, characterized in that, The XLSTM module captures long-term dependencies using the following method: Matrix memory storage structure using mLSTM cells This enables parallel storage and processing of multi-dimensional feature relationships in vehicle trajectory data, where feature information such as speed, acceleration, and relative distance are stored collaboratively in matrix form. Through matrix key-value pairs (k t ,v t The parallel update mechanism synchronously processes long-term dependencies between multiple feature dimensions, where the key vector k t Responsible for feature selection, value vector v t Responsible for feature storage and query vectors Implement feature-based retrieval; By employing extended memory capacity and parallel computing architecture, it effectively captures multi-scale temporal dependencies from short-term to long-term in car-following behavior, solving the gradient vanishing or exploding problem of traditional LSTM models when processing long-sequence vehicle trajectory data. By using a hybrid stacked architecture of sLSTM and mLSTM, combining the advantages of local feature interaction and global memory storage, accurate modeling of long-term car-following patterns in heterogeneous driving behaviors can be achieved.

9. The method according to claim 1, characterized in that, The prediction result of vehicle following behavior based on the obtained fusion feature representation is output, specifically including: The obtained fusion feature representation The input is fed into a fully connected layer for feature dimension transformation, mapping it to a feature space that matches the dimension of the prediction target, where the prediction target includes the speed v of the following vehicle at a future time t+l. t+l acceleration a t+l and the relative distance Δd to the vehicle in front. t+l At least one parameter, where l∈[1,τ]; A time-distributed fully connected layer is used to perform temporal decoding on the transformed features, generating a vehicle trajectory prediction sequence for the future time interval [t+1, t+τ] step by step, where τ is the prediction time span; The decoded features are normalized by the output layer, and continuous numerical prediction results are generated by using a linear activation function or discrete driving behavior classification probabilities are generated by using a Softmax activation function. The driving behavior classification includes at least three types: aggressive, conservative, and normal. Supervised training is performed on the predicted results and actual trajectory data using the mean squared error loss function or cross-entropy loss function. The model parameters are optimized using the gradient descent algorithm to make the predicted results approximate the actual car-following behavior data as closely as possible. The calculation formula is as follows: ； in For the true value, is the predicted value, and N is the number of training samples.

10. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-9.

Images