A vehicle lane-changing trajectory prediction method based on multi-view and intention hierarchical fusion

Abstract

The application discloses a kind of vehicle lane-changing trajectory prediction methods based on multi-view and intention hierarchical fusion, it is related to automatic driving and intelligent transportation technical field.The method includes: vehicle trajectory data is preprocessed and driving scene is divided;From driving behavior and traffic state double view angle, feature is extracted, factor analysis dimension reduction and multi-view joint clustering are carried out, obtain multi-view driving style probability vector;Style probability vector is used as long-term prior condition, by gate mechanism, into time sequence feature matrix, input into sequence model of fusion multi-head attention mechanism, output lane-changing intention probability vector;Multi-view driving style probability vector and lane-changing intention probability vector are used as joint condition feature and input personalized trajectory prediction model, output optimal trajectory.The application solves the problem that existing method ignores long-term behavior preference and short-term decision correlation, improves long-time domain prediction accuracy and the individualization consistency of trajectory, reduces the collision risk under the environment of man-machine mixed travel.

Description

Technical Field

[0001] This invention relates to the fields of autonomous driving and intelligent transportation technology, and in particular to a method for predicting vehicle lane-changing trajectories based on multi-view and intent-level fusion. Background Technology

[0002] In real-world traffic environments where Level 3 / 4 autonomous vehicles and human-driven vehicles coexist, human drivers exhibit significant individual differences and behavioral uncertainties in their lane-changing decision-making mechanisms and trajectory evolution. Different drivers display distinct style characteristics in areas such as longitudinal acceleration selection, aggressiveness in lateral lane changes, acceptable clearance thresholds, and response strategies to interactions with surrounding vehicles. This personalized lane-changing behavior makes it difficult for autonomous driving systems to accurately predict the driving intentions and future trajectories of surrounding vehicles. Especially in human-machine mixed-traffic scenarios, misjudgments of intent and deviations in trajectory prediction can easily induce traffic conflicts or even collisions. Therefore, accurately identifying drivers' personalized style preferences and effectively embedding them into lane-changing prediction models has become a key issue in improving the environmental perception and decision-making capabilities of autonomous driving systems.

[0003] Existing methods typically process driving style modeling, lane change intention recognition, and trajectory generation in separate stages, lacking effective semantic transmission and information sharing mechanisms between modules, resulting in limited overall system performance. Specifically, the following prominent defects exist: (1) The hierarchical correlation mechanism between long-term behavioral preferences and short-term decision goals is ignored, and driving style information is simply used as a static label without dynamic fusion with real-time behavioral evidence, resulting in significant deficiencies in the personalized consistency of prediction results; (2) Driving style identification relies on only a single perspective (such as only looking at kinematic parameters or only looking at traffic flow status), lacking comprehensive consideration of different driving scenarios, with insufficient scenario-dependent ability, making it difficult to fully characterize the multi-dimensional style features of drivers; (3) The accuracy of long-term intention prediction decays significantly, and the trajectory generation process lacks high-level semantic constraints, resulting in poor multimodal coverage and large endpoint prediction errors, making it difficult to meet the safety and reliability requirements of actual autonomous driving systems.

[0004] Overall, existing technical frameworks treat the three core components of style, intent, and trajectory prediction in a fragmented manner. This fails to effectively guide short-term decisions based on long-term behavioral preferences, nor can it leverage high-level intent semantics to constrain and optimize the trajectory generation process, resulting in severe fragmentation of semantic information. Particularly in the lane-changing execution phase, existing methods fail to address the dynamic balance between style priors and real-time kinematic signals, leading to style information potentially interfering with rather than assisting predictions at critical moments. Furthermore, existing models rely heavily on human experience or traditional heuristic algorithms for hyperparameter optimization, lacking efficient adaptive optimization mechanisms, further limiting the model's prediction accuracy and generalization ability. Summary of the Invention

[0005] The purpose of this invention is to propose a vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion. This method solves the problems of missing correlation between long-term behavioral preferences and short-term decisions, fragmented semantic information, and insufficient personalization in existing lane-changing prediction methods. It achieves a significant improvement in long-term prediction accuracy and trajectory personalization consistency, and reduces the collision risk in human-machine mixed traffic environments.

[0006] To achieve the above objectives, this invention proposes a vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion, which is applied to the environmental perception and decision-making planning of autonomous vehicles. The specific steps are as follows: Step S1: Obtain historical trajectory data of the target vehicle and surrounding interactive vehicles, perform noise reduction preprocessing and driving scene segmentation, and construct a temporal feature matrix; Step S2: Extract multidimensional features from the perspectives of driving behavior and traffic status, perform factor analysis for dimensionality reduction and single-view soft clustering, concatenate the posterior probability vectors of the two perspectives and perform joint clustering to obtain the multi-view driving style probability vector. Step S3: The multi-view driving style probability vector is used as a long-term prior condition, integrated into the temporal feature matrix through a gating mechanism, and input into the sequence model that integrates the multi-head attention mechanism to output the lane change intention probability vector. Step S4: Input the multi-view driving style probability vector and lane change intention probability vector as joint conditional features into the personalized trajectory prediction model, generate future multimodal lane change trajectories through autoregression, and output the optimal trajectory.

[0007] Preferably, in step S1, the noise reduction preprocessing adopts a multi-stage filtering strategy that separates the time domain and the spatial domain: in the time domain, Savitzky-Golay filtering is used to smooth the velocity and acceleration sequences using local polynomial fitting, and in the spatial domain, Gaussian filtering is introduced to smooth the spatial position coordinates.

[0008] Preferably, the driving scenario segmentation includes: dividing non-lane-changing frames into free straight driving and following states based on the headway (THW) threshold and collision time (TTC) threshold; and dividing lane-changing behavior into three stages: lane-changing start point, lane-changing in progress point, and lane-changing end point based on lateral speed and lane centerline offset.

[0009] Preferably, in step S2, obtaining the multi-view driving style probability vector specifically includes: Step S21: Extract kinematic and interaction features for the three scenarios of free straight driving, following, and lane changing, respectively. Use factor analysis to reduce dimensionality and extract principal factors, and use Gaussian mixture model to perform soft clustering to obtain the driving behavior style probability vector. Step S22: Aggregate the macroscopic operational characteristics of local traffic flow using a fixed time window, and obtain the traffic state probability vector through factor analysis dimensionality reduction and Gaussian mixture model soft clustering. Step S23: Concatenate the driving behavior style probability vector and the traffic state probability vector into a joint feature vector, perform secondary Gaussian mixture model clustering in the joint feature space, and output the multi-view driving style probability vector.

[0010] Preferably, in step S3, the gating mechanism calculates the gating weights using learnable parameters and a sigmoid activation function, and dynamically adjusts the relative weights between prior driving style and real-time behavioral evidence. The sequence model includes a one-dimensional convolutional neural network local feature extraction module, a bidirectional long short-term memory network bidirectional temporal modeling module, and a multi-head attention enhancement module.

[0011] Preferably, an improved whale optimization algorithm is used to globally optimize the hyperparameters of the sequence model. This improved whale optimization algorithm introduces Bayesian optimization to generate a high-quality initial population and uses an exponential decay step size instead of traditional linear decay to enhance the fine-grained search capability in the later stages. Simultaneously, a differential mutation mechanism is introduced to maintain population diversity and avoid getting trapped in local optima. Specifically: Step S31: To shorten the early ineffective exploration phase, the expected improvement (EI) criterion in Bayesian optimization is used to guide the generation of the initial population, as shown in the following formula: ; in, For new candidate solutions to be evaluated, For the newly selected sampling point, The objective function value corresponding to the new sampling point. This is the current optimal solution. The objective function value corresponding to the current optimal solution. For expectation calculation; Step S32: Replace the traditional linear decay mechanism with an exponential decay mechanism, which allows for large-scale exploration in the early stage and rapid contraction in the later stage to achieve fine search. The step size factor formula is as follows: ; in, Let be the search step size in the t-th iteration. t This represents the current iteration number. The initial step size, The attenuation coefficient is... This represents the maximum number of iterations. Step S33: To mitigate the risk of premature convergence caused by a rapid decline in population diversity, differential mutation is introduced into the position update, as shown in the following formula: ; in, Scaling factor , , Three different individuals were randomly selected. Candidate individuals generated by differential mutation.

[0012] Preferably, in step S4, the personalized trajectory prediction model adopts a GRU codec network architecture, including: Encoding phase: A two-layer GRU is used to independently process the historical sequence of the target vehicle and surrounding interactive vehicles, and the hidden states of multiple vehicles are aggregated through an attention pooling mechanism to output the global hidden state; Conditional fusion stage: The joint conditional features are projected into a query vector using a cross-attention mechanism, and the global hidden state is projected into a key vector and a value vector. Attention weights are calculated and the value vectors are weighted and summed. The residual is then fused with the global hidden state to serve as the initial hidden state of the decoder. Decoding stage: A single-layer GRU is used to generate the future trajectory coordinate offset in an autoregressive manner. Independent Gaussian noise is added to the initial hidden state for sampling to generate N candidate trajectories.

[0013] Preferably, the cross-attention mechanism specifically includes: Step S41: First, generate the query vector, key vector, and value vector from the joint conditional features using linear mapping, as shown in the following formula: , , ; in, For query vector, For key vectors, For value vectors, This is a globally hidden state. , , These are the linear mapping matrices corresponding to the query vector, key vector, and value vector, respectively. Step S42: Calculate the weight of each attention head and weight the value vector; where, the first... h The formula for each attention point is as follows: ; in, For the h-th attention head, For the first h A query vector with attention heads. For the first h The key vector of each attention head. For single-head feature dimensions, For the first h A vector of values ​​for each attention head.T This is a matrix transpose operation; Step S43: Concatenate the outputs of each attention head along the feature dimension, and obtain the final attention enhancement through a linear transformation, as shown in the following formula: ; in, For the final attention-enhanced feature representation, For the first H Each attention head output, It is a linear transformation matrix. This is the feature splicing operation function.

[0014] Preferably, the loss function of the personalized trajectory prediction model is the minimum error between the candidate trajectory and the true trajectory, plus a penalty term for the average pairwise distance between each candidate trajectory, as shown in the following formula: ; in, The loss function value for the personalized trajectory prediction model. For the first n Candidate trajectories, For the actual trajectory, The weight of the average pairwise distance penalty term. The squared error of the Euclidean distance. n The number of candidate trajectories; The average pairwise distance penalty term is used to prevent the generated trajectories from becoming homogeneous. The average pairwise L2 distance between each candidate trajectory is calculated using the following formula: ; in, For the first i Candidate trajectories, N This represents the number of candidate trajectories.

[0015] Therefore, this invention proposes a vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion, which has the following advantages: (1) This invention breaks through the limitations of traditional single-perspective style recognition. It extracts features from both driving behavior and traffic conditions, performs factor analysis for dimensionality reduction and Gaussian mixture model clustering, and then concatenates the posterior probabilities of the two perspectives for joint clustering. This mechanism not only characterizes the driver's own operational preferences but also fully integrates macro-level traffic context information, realizing fine-grained and dynamic representation of driving style in various traffic contexts, providing more accurate long-term behavioral priors for subsequent intention prediction and trajectory generation.

[0016] (2) This invention uses a gating fusion mechanism to adaptively inject the prior of multi-view driving style into the temporal feature matrix, avoiding the interference of style information on kinematic signals during the lane change execution phase; at the same time, the improved whale optimization algorithm (IWOA) is used to perform global hyperparameter optimization on the sequence model with fused multi-head attention mechanism, and fast convergence and fine search are achieved by using Bayesian guided initialization, exponential decay step size and differential mutation mechanism.

[0017] (3) This invention uses driving style probability and lane change intention probability as joint conditional features and dynamically injects them into the GRU codec network through cross-attention mechanism. This realizes the effective constraint and guidance of high-level semantic conditions on the trajectory generation process, overcomes the defect of increased endpoint error in long-term prediction of traditional static splicing method, and provides reliable technical support for the safety decision of autonomous vehicles in human-machine mixed environment.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a flowchart of a vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion according to the present invention. Figure 2 This is an architecture diagram of a vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion according to the present invention. Figure 3 This is a comparative diagram of the improved IWOA algorithm in this invention with the traditional WOA, PSO and GA algorithms in the hyperparameter optimization process; where (a) is a line graph comparing F1 scores, (b) is a graph comparing optimization accuracy, (c) is a graph comparing fitness, and (d) is a graph comparing cross-entropy loss. Figure 4 This is a schematic diagram comparing the degree of deviation between cross-attention fusion and early fusion and late fusion in trajectory prediction in this invention; wherein, (a) is a comparison diagram of lateral position deviation, and (b) is a comparison diagram of longitudinal position deviation. Figure 5 This is a schematic diagram comparing the prediction results of the model proposed in this invention with those of existing technology models. Detailed Implementation

[0020] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0022] like Figures 1-2 As shown, this invention provides a vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion, with the following specific steps: Step S1: Obtain historical trajectory data of the target vehicle and surrounding interactive vehicles, perform noise reduction preprocessing and driving scene segmentation, and construct a temporal feature matrix; This embodiment uses the NGSIM real-world highway dataset as the data source. First, the original trajectory data is cleaned to remove motorcycles, trucks, and ramp vehicles, repair trajectory interruptions caused by brief occlusions, and filter out anomalies in velocity and acceleration exceeding physical limits. To suppress measurement noise and numerical differential amplification effects, a multi-stage filtering strategy separating the temporal and spatial domains is employed: in the temporal domain, Savitzky-Golay filtering is used to smooth the velocity and acceleration sequences using local polynomial fitting; in the spatial domain, Gaussian filtering is introduced to smooth the spatial coordinates and eliminate isolated jump points.

[0023] Based on the filtered trajectory data, longitudinal and lateral kinematic features (such as speed, acceleration, jerk, cumulative lateral displacement, etc.) and traffic interaction features (such as headway time-to-head distance, collision time-to-collision time, relative distance, etc.) of the target vehicle and six surrounding interactive vehicles are extracted to construct an 18-dimensional temporal feature matrix.

[0024] In this embodiment, the driving scenario is divided as follows: based on the headway (THW) and collision time (TTC) threshold, non-lane-changing frames are divided into free straight driving and following states; based on lateral speed and lane centerline offset, lane-changing behavior is further divided into three stages: lane-changing start point, lane-changing in progress point, and lane-changing end point.

[0025] Step S2: Extract multi-dimensional features from both the driving behavior and traffic state perspectives, perform factor analysis for dimensionality reduction and single-view soft clustering, concatenate the posterior probability vectors of the two perspectives, and then perform joint clustering to obtain the multi-view driving style probability vector, as detailed below: Step S21, Driving Behavior Perspective: Kinematic and interaction features are extracted for three scenarios: free straight driving, following, and lane changing. Factor analysis (FA) is used to reduce dimensionality and extract principal factors, and Gaussian mixture model (GMM) is used for soft clustering to obtain a 3-dimensional driving behavior style probability vector (including conservative probability, normal probability, and aggressive probability). Step S22, Traffic State Perspective: Aggregate the macroscopic operational characteristics of local traffic flow (such as flow rate, density, average speed, acceleration disturbance, etc.) in a fixed time window, and then perform factor analysis dimensionality reduction and GMM soft clustering to obtain a 5-dimensional traffic state probability vector (including smooth flow probability, disturbance enhancement probability, low speed restriction probability, medium density probability, congestion deceleration probability). Step S23, Multi-view Joint Clustering: The 3-dimensional behavior probability vector and the 5-dimensional state probability vector are directly concatenated into an 8-dimensional joint feature vector. Secondary GMM clustering is performed in the joint feature space, and finally an 11-dimensional multi-view driving style probability vector is output.

[0026] Step S3: The multi-view driving style probability vector is used as a long-term prior condition, and is integrated into the temporal feature matrix through a gating mechanism. It is then input into the sequence model that integrates the multi-head attention mechanism, and the output lane change intention probability vector is generated. To avoid style priors interfering with significant kinematic signals during lane change execution, this embodiment designs a gating fusion mechanism: gating weights are calculated using learnable parameters and a sigmoid activation function, and the relative weights between driving style priors and real-time behavioral evidence are dynamically adjusted to achieve adaptive conditional injection.

[0027] In this embodiment, the sequence model includes: a one-dimensional convolutional neural network (1D CNN) local feature extraction module for capturing local spatiotemporal patterns within a short time window; a bidirectional long short-term memory network (BiLSTM) bidirectional temporal modeling module for capturing long-distance dependencies between sequences; and a multi-head attention enhancement module for adaptively weighting key time steps and interaction channels to improve sensitivity to weak signals in the early stages of lane changes. Furthermore, an improved whale optimization algorithm (IWOA) is used to globally optimize the hyperparameters of the sequence model. The improved whale optimization algorithm introduces Bayesian optimization to generate a high-quality initial population and uses an exponential decay step size instead of traditional linear decay to enhance the fine-grained search capability in later stages. Simultaneously, a differential mutation mechanism is introduced to maintain population diversity and avoid getting trapped in local optima. Specifically, it consists of the following steps: Step S31: To shorten the early ineffective exploration phase, the expected improvement (EI) criterion in Bayesian optimization is used to guide the generation of the initial population, as shown in the following formula: ; in, For new candidate solutions to be evaluated, For the newly selected sampling point, The objective function value corresponding to the new sampling point. This is the current optimal solution. The objective function value corresponding to the current optimal solution. For expectation calculation; Step S32: Replace the traditional linear decay mechanism with an exponential decay mechanism, which allows for large-scale exploration in the early stage and rapid contraction in the later stage to achieve fine search. The step size factor formula is as follows: ; in, Let be the search step size in the t-th iteration. t This represents the current iteration number. The initial step size, The attenuation coefficient is... This represents the maximum number of iterations. Step S33: To mitigate the risk of premature convergence caused by a rapid decline in population diversity, differential mutation is introduced into the position update, as shown in the following formula: ; in, Scaling factor , , Three different individuals were randomly selected. Candidate individuals generated by differential mutation.

[0028] In this embodiment, the sequence model ultimately outputs a 3D probability vector of lane-changing intentions: left lane change, straight ahead, and right lane change.

[0029] Step S4: Input the multi-view driving style probability vector and lane change intention probability vector as joint conditional features into the personalized trajectory prediction model, generate future multimodal lane change trajectories through autoregression, and output the optimal trajectory.

[0030] The personalized trajectory prediction model employs a GRU encoder-decoder network architecture, specifically including: During the coding phase, a two-layer GRU is used to independently process the historical sequences of the target vehicle and surrounding interactive vehicles, and the hidden states of multiple vehicles are aggregated through an attention pooling mechanism to implicitly model the interaction relationships of multiple vehicles and output the global hidden states. In the conditional fusion stage, a cross-attention mechanism is used for dynamic injection: the 14-dimensional joint conditional features (11-dimensional style probability vector + 3-dimensional intent probability vector) are projected into a query vector, and the global hidden state output by the GRU encoder is projected into a key vector and a value vector; the cross-attention mechanism specifically includes: Step S41: First, generate the query vector, key vector, and value vector from the joint conditional features using linear mapping, as shown in the following formula: , , ; in, For query vector, For key vectors, For value vectors, This is a globally hidden state. , , These are the linear mapping matrices corresponding to the query vector, key vector, and value vector, respectively. Step S42: Calculate the weight of each attention head and weight the value vector; where, the first... h The formula for each attention point is as follows: ; in, For the h-th attention head, For the first h A query vector with attention heads. For the first h The key vector of each attention head. For single-head feature dimensions, For the first h A vector of values ​​for each attention head. T This is a matrix transpose operation; Step S43: Concatenate the outputs of each attention head along the feature dimension, and obtain the final attention enhancement through a linear transformation, as shown in the following formula: ; in, For the final attention-enhanced feature representation, For the output of the Hth attention head, It is a linear transformation matrix. This is the feature splicing operation function.

[0031] Calculate the attention weights and sum the value vectors by weights. Then, perform residual fusion of the obtained joint conditional features with the global hidden state to serve as the initial hidden state of the decoder. In the decoding and multimodal generation stages, a single-layer GRU is used to generate future trajectory coordinate offsets in an autoregressive manner. To enhance multimodal prediction capabilities, independent Gaussian noise is added to the initial hidden state of the decoder for sampling, generating... N Candidate trajectories, in this embodiment N =10.

[0032] In this embodiment, the loss function of the personalized trajectory prediction model is the minimum error between the candidate trajectory and the true trajectory, plus a penalty term for the average pairwise distance between each candidate trajectory to prevent pattern collapse. The formula is as follows: ; in, The loss function value for the personalized trajectory prediction model. For the firstn Candidate trajectories, For the actual trajectory, The weight of the average pairwise distance penalty term. The squared error of the Euclidean distance. n The number of candidate trajectories; The average pairwise distance penalty term is used to prevent the generated trajectories from becoming homogeneous. The average pairwise L2 distance between each candidate trajectory is calculated using the following formula: ; in, For the first i Candidate trajectories, N This represents the number of candidate trajectories.

[0033] To verify the technical effectiveness of the method proposed in this invention, the method proposed in this invention is compared with the prior art.

[0034] like Figure 3 As shown, the improved IWOA algorithm proposed in this invention benefits from the introduction of Bayesian optimization of the initial population, differential mutation, and exponential decay step size. It is significantly better than the comparative algorithm in terms of convergence speed and stability (it tends to converge around the 12th generation). The accuracy increases rapidly and the cross-entropy loss decreases rapidly, which proves the efficiency of the improved IWOA algorithm of this invention.

[0035] Figure 4 This visually demonstrates the degree of deviation in trajectory prediction between early fusion (simple stitching), late fusion, and the cross-attention fusion of this invention. Figure 4 It can be seen that the predicted trajectory using cross-attention fusion is closest to the real trajectory in both the horizontal and vertical directions. Cross-attention fusion reduces the average displacement error (ADE) by 20.2%~25.8% by dynamically calculating the weighted contribution of the interaction context, which significantly overcomes the defect that the traditional static stitching method is prone to increasing the prediction endpoint error in the long time domain.

[0036] Figure 5 This paper compares the prediction accuracy of the IWOA-CNN-BiLSTM model based on multi-head attention mechanism proposed in this invention with CNN, BiLSTM, CNN-BiLSTM, WOA-CNN-BiLSTM based on multi-head attention mechanism, IWOA-CNN-BiLSTM, and LSTM models. Figure 5As can be seen, the IWOA-CNN-BiLSTM model based on the multi-head attention mechanism proposed in this invention maintains the highest prediction accuracy and the smallest degradation slope across all prediction durations. Even in the longest prediction time of 3.0s, the accuracy still reaches 79.68%, which is significantly better than benchmark models such as CNN and LSTM, proving that this invention has a very strong stable modeling capability for weak signals in the early stage of lane changing.

[0037] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0038] Therefore, this invention provides a vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion. By constructing a multi-view driving style identification mechanism, a style-intent-level fusion prediction framework, and cross-attention-driven trajectory joint optimization, it effectively solves the technical problems of semantic information fragmentation, insufficient personalization consistency, and low long-term domain prediction accuracy of existing methods. It improves the accuracy of lane-changing intent and trajectory prediction and reduces the collision risk in human-machine mixed driving scenarios.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion, applied to environmental perception and decision-making planning of autonomous vehicles, characterized in that... Includes the following steps: Step S1: Obtain historical trajectory data of the target vehicle and surrounding interactive vehicles, perform noise reduction preprocessing and driving scene segmentation, and construct a temporal feature matrix; Step S2: Extract multidimensional features from the perspectives of driving behavior and traffic status, perform factor analysis for dimensionality reduction and single-view soft clustering, concatenate the posterior probability vectors of the two perspectives and perform joint clustering to obtain the multi-view driving style probability vector. Step S3: The multi-view driving style probability vector is used as a long-term prior condition, integrated into the temporal feature matrix through a gating mechanism, and input into the sequence model that integrates the multi-head attention mechanism to output the lane change intention probability vector. Step S4: Input the multi-view driving style probability vector and lane change intention probability vector as joint conditional features into the personalized trajectory prediction model, generate future multimodal lane change trajectories through autoregression, and output the optimal trajectory. In step S3, the gating mechanism calculates the gating weights using learnable parameters and the Sigmoid activation function, and dynamically adjusts the relative weights between prior driving style and real-time behavioral evidence. The sequence model includes a one-dimensional convolutional neural network local feature extraction module, a bidirectional long short-term memory network bidirectional temporal modeling module, and a multi-head attention enhancement module; An improved whale optimization algorithm is used to globally optimize the hyperparameters of the sequence model. This improved whale optimization algorithm introduces Bayesian optimization to generate a high-quality initial population and uses an exponential decay step size instead of the traditional linear decay to enhance the fine-grained search capability in the later stages. At the same time, a differential mutation mechanism is introduced to maintain population diversity and avoid getting trapped in local optima, as detailed below: Step S31: To shorten the early ineffective exploration phase, the Expected Improvement (EI) criterion in Bayesian optimization is used to guide the initial population generation, as shown in the following formula: ； in, For new candidate solutions to be evaluated, For the newly selected sampling point, The objective function value corresponding to the new sampling point. This is the current optimal solution. The objective function value corresponding to the current optimal solution. For expectation calculation; Step S32: Replace the traditional linear decay mechanism with an exponential decay mechanism, which allows for large-scale exploration in the early stage and rapid contraction in the later stage to achieve fine search. The step size factor formula is as follows: ； in, For the first t Search step size in the next iteration t This represents the current iteration number. The initial step size, The attenuation coefficient is... This represents the maximum number of iterations. Step S33: To mitigate the risk of premature convergence caused by a rapid decline in population diversity, differential mutation is introduced into the position update, as shown in the following formula: ； in, Scaling factor , , Three different individuals were randomly selected. Candidate individuals generated by differential mutation.

2. The vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion according to claim 1, characterized in that, In step S1, the noise reduction preprocessing adopts a multi-stage filtering strategy that separates the time domain and the spatial domain: in the time domain, Savitzky-Golay filtering is used to smooth the velocity and acceleration sequences using local polynomial fitting, and in the spatial domain, Gaussian filtering is introduced to smooth the spatial position coordinates.

3. The vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion according to claim 2, characterized in that, The driving scenario segmentation includes: dividing non-lane-changing frames into free straight driving and following states based on the headway (THW) threshold and collision time (TTC) threshold; and dividing lane-changing behavior into three stages: lane-changing start point, lane-changing in progress point, and lane-changing end point based on lateral speed and lane centerline offset.

4. The vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion according to claim 3, characterized in that, In step S2, obtaining the multi-view driving style probability vector specifically includes: Step S21: Extract kinematic and interaction features for the three scenarios of free straight driving, following, and lane changing, respectively. Use factor analysis to reduce dimensionality and extract principal factors, and use Gaussian mixture model to perform soft clustering to obtain the driving behavior style probability vector. Step S22: Aggregate the macroscopic operational characteristics of local traffic flow using a fixed time window, and obtain the traffic state probability vector through factor analysis dimensionality reduction and Gaussian mixture model soft clustering. Step S23: Concatenate the driving behavior style probability vector and the traffic state probability vector into a joint feature vector, perform secondary Gaussian mixture model clustering in the joint feature space, and output the multi-view driving style probability vector.

5. The vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion according to claim 4, characterized in that, In step S4, the personalized trajectory prediction model adopts a GRU codec network architecture, including: Encoding phase: A two-layer GRU is used to independently process the historical sequence of the target vehicle and surrounding interactive vehicles, and the hidden states of multiple vehicles are aggregated through an attention pooling mechanism to output the global hidden state; Conditional fusion stage: The joint conditional features are projected into a query vector using a cross-attention mechanism, and the global hidden state is projected into a key vector and a value vector. Attention weights are calculated and the value vectors are weighted and summed. The residual is then fused with the global hidden state to serve as the initial hidden state of the decoder. Decoding stage: A single-layer GRU is used to generate the future trajectory coordinate offset in an autoregressive manner. Independent Gaussian noise is added to the initial hidden state for sampling to generate N candidate trajectories.

6. The vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion according to claim 5, characterized in that, The cross-attention mechanism specifically includes: Step S41: First, generate the query vector, key vector, and value vector from the joint conditional features using linear mapping, as shown in the following formula: ， ， ； in, For query vector, For key vectors, For value vectors, This is a globally hidden state. , , These are the linear mapping matrices corresponding to the query vector, key vector, and value vector, respectively. Step S42: Calculate the weight of each attention head and weight the value vector; where, the first... h The formula for each attention point is as follows: ； in, For the h-th attention head, For the first h A query vector with attention heads. For the first h The key vector of each attention head. For single-head feature dimensions, For the first h A vector of values ​​for each attention head. T This is a matrix transpose operation; Step S43: Concatenate the outputs of each attention head along the feature dimension, and obtain the final attention enhancement through a linear transformation, as shown in the following formula: ； in, For the final attention-enhanced feature representation, For the first H Each attention head output, It is a linear transformation matrix. This is the feature splicing operation function.

7. The vehicle lane-changing trajectory prediction method based on multi-view and intent-level fusion according to claim 6, characterized in that, The loss function of the personalized trajectory prediction model is the minimum error between the candidate trajectory and the true trajectory, plus a penalty term for the average pairwise distance between each candidate trajectory, as shown in the following formula: ； in, The loss function value for the personalized trajectory prediction model. For the first n Candidate trajectories, For the actual trajectory, The weight of the average pairwise distance penalty term. The squared error of the Euclidean distance. n The number of candidate trajectories; The average pairwise distance penalty term is used to prevent the generated trajectories from becoming homogeneous. The average pairwise L2 distance between each candidate trajectory is calculated using the following formula: ； in, For the first i Candidate trajectories, N This represents the number of candidate trajectories.

Images