Crossroad vehicle behavior and trajectory prediction method based on prior assistance

Abstract

The invention belongs to the technical field of automatic driving, and relates to an intersection vehicle behavior and trajectory prediction method based on prior assistance, which comprises the following steps: 1, through trajectory data segmentation and sampling, extracting spatial-temporal characteristics of a trajectory, reducing data dimensions, and generating high-fidelity low-dimensional data; 2, through a full-connection neural network behavior prediction model combining jump step prediction and LSTM assistance, predicting a left turn, right turn or straight going intention, and taking output of the full-connection neural network behavior prediction model as prior assistance of trajectory prediction; 3, key trajectory features are focused through an attention mechanism, and staged and multi-stage trajectory prediction is achieved; 4, performing integrated prediction through a prior auxiliary parallel heuristic attention prediction model; according to the method, the prediction precision is improved, the prospective time is prolonged, the generalization ability and deployment efficiency of the model are enhanced, reliable support is provided for automatic driving in an intersection scene, and the traffic accident risk and the congestion degree are effectively reduced.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology and relates to a method for predicting vehicle behavior and trajectory at intersections based on prior assistance. Background Technology

[0002] Vehicle behavior at intersections involves multi-factor interactions and trajectory planning, requiring rapid decision-making in complex and dynamic environments. To systematically describe this process, existing research typically decomposes it into the Target Selection Phase (TPS) and the Drive Toward Target Phase (DTTP). In the TPS phase, the driver or autonomous driving system selects the driving direction based on the destination and performs preparatory operations such as lane changes. In the DTTP phase, the vehicle generates a left-turn, right-turn, or straight-ahead trajectory based on the selected direction. To ensure traffic safety and efficiency, TPS identifies vehicle steering intentions through a classification task, while DTTP predicts the vehicle's future trajectory through a regression task. Compared to highways, vehicle behavior at intersections is more significantly influenced by the destination, surrounding vehicles, and road structure, especially in unlit or unstructured scenarios where trajectory interweaving and complex motion patterns significantly increase the difficulty of prediction.

[0003] The dynamic environment of intersections requires models to comprehensively consider destination, traffic signals, and vehicle interaction information. This is especially challenging in scenarios without traffic lights or in unstructured environments, where data heterogeneity and a lack of rule constraints make prediction tasks even more difficult. The stage-dependent nature of TPS and DTTP necessitates joint modeling to ensure consistency between intent and trajectory prediction. TPS needs to accurately distinguish between left-turn, right-turn, and straight-ahead intentions, while DTTP needs to precisely capture the spatiotemporal characteristics of the trajectory. In practical applications, models must balance high prediction accuracy, long lead time, cross-scenario generalization ability, and real-time performance.

[0004] Existing research on TPS and DTTP prediction has formed two major methodological systems: rule-based / mathematical derivation and data-driven approaches. In TPS prediction, rule-based models model driving relationships through mathematical expressions, such as reactive driving agent algorithms and decision tree models. However, rule-based models, due to their fixed rules, struggle to adapt to unstructured scenarios. Data-driven models improve performance by learning the mapping between motion states and intentions, such as reinforcement learning right-turn decision models and fully connected neural network models; however, data-driven models have weaker interpretability.

[0005] In DTTP prediction, mathematical derivation models use mathematical expressions to describe trajectories, such as Kalman filter models and adaptive Kalman filters. However, mathematical models are not well-suited to irregular trajectories. Data-driven models map future trajectories to historical trajectories, such as variational neural networks and LSTM trajectory prediction models. However, the computational complexity and stability of data-driven models need improvement.

[0006] Therefore, existing methods have the following limitations:

[0007] 1) Most only focus on TPS or DTTP, and the separation of models leads to high development costs and poor consistency.

[0008] 2) The prediction accuracy and forward-looking time are contradictory, and insufficient data analysis leads to limitations in accuracy or forward-looking time.

[0009] 3) The models are mostly concentrated in signalized intersections or structured scenarios, with poor generalization ability and difficulty in adapting to intersections without traffic lights or unstructured intersections.

[0010] 4) High-dimensional inputs and complex networks reduce robustness and practicality, and increase computing power requirements.

[0011] Therefore, there is an urgent need for a prediction method that combines high accuracy with good time-forward capability to provide reliable support for autonomous driving in intersection scenarios and solve the aforementioned technical problems. Summary of the Invention

[0012] The technical solution adopted by this invention to solve the technical problem is: a method for predicting vehicle behavior and trajectory at intersections based on prior assistance, comprising the following steps:

[0013] Step S1: By segmenting and sampling trajectory data, extract the spatiotemporal features of the trajectory and reduce the data dimensionality to generate high-fidelity low-dimensional data.

[0014] Step S2: By combining jump step prediction and LSTM-assisted fully connected neural network behavior prediction model, predict the intention to turn left, turn right or go straight. The output of the fully connected neural network behavior prediction model is used as a priori aid for trajectory prediction.

[0015] Step S3: Focus on key trajectory features through an attention mechanism to achieve multi-stage trajectory prediction.

[0016] Step S4: Integrate predictions using a priori-assisted parallel heuristic attention prediction model.

[0017] Preferably, step S1 includes the following sub-steps:

[0018] Step S1-1: Vehicle trajectory extraction. Based on the public dataset, select structured intersection and unstructured intersection files, filter out non-vehicle and stationary data, and obtain multiple dynamic vehicle trajectories.

[0019] Step S1-2: Vehicle trajectory segmentation. The process of a vehicle passing through an intersection is divided into three stages: approaching the intersection, crossing the intersection, and leaving the intersection. By constructing a coordinate system and considering the vehicle's directional information in the intersection, the trajectory of the intersection is segmented.

[0020] Steps S1-3: Data sampling and homogenization. Integer multiple sampling is used to sample the segmented stage trajectories.

[0021] Preferably, in step S2, the jump step prediction includes: a jump step prediction network and a jump step trajectory prediction.

[0022] The jump step prediction network includes a trajectory prediction module and a trajectory correction module. The trajectory prediction module improves the LSTM-based encoder-decoder structure by replacing the first layer of the LSTM with a lighter GRU. The trajectory correction module is based on a trajectory correction framework of a deep neural network.

[0023] The jump-step trajectory prediction is based on a jump-step prediction network. Through trajectory training and testing, using all trajectory data from both structured and unstructured intersections, preliminary trajectory prediction results are obtained. These preliminary trajectory prediction results are then used to derive the turning intention, which serves as the prior information needed for subsequent calculations.

[0024] More preferably, in step S2, an LSTM-assisted fully connected neural network behavior prediction model is constructed. This model builds an LSTM-assisted fully connected network (FCN), replacing the first layer of the traditional FCN with an LSTM. The fully connected network integrates features and outputs the probabilities of left turn, right turn, and straight-ahead intentions. A cascaded jump-step prediction network and the LSTM-assisted FCN are used to construct an intersection turning intention prediction model. The turning intention prediction network and the jump-step prediction network are two separate networks. The relationship between them is as follows: the jump-step prediction network obtains a preliminary predicted trajectory, and then the turning intention prediction network accurately predicts the turning intention based on the preliminary predicted trajectory. The turning intention serves as the prior aid needed for subsequent actions.

[0025] Preferably, in step S3, a heuristic attention-assisted encoder-decoder structure is developed, which consists of multiple attention-assisted encoders-decoders.

[0026] Better still, we design a heuristic attention-assisted encoder-decoder network based on LSTM and GRU to connect each stage; during training, the output of each stage is used as the input of the next stage, and the training label is the stage trajectory; we test the trajectory prediction of the four stages during the turning process, and after obtaining the optimal attention-assisted encoder-decoder structure for each stage, we form the optimal heuristic network; the attention-assisted encoder-decoder of the four stages has a similar structure.

[0027] Even better, the heuristic attention-assisted encoder-decoder network structure is optimized through systematic experiments for different driving behaviors and trajectory stages.

[0028] Preferably, in step S4, the prior-assisted parallel heuristic attention prediction model integrates the output intention prediction and multi-stage trajectory prediction to construct a prior-assisted parallel structure, uses the behavior prediction result as a switch signal to activate the corresponding attention-assisted encoder-decoder branch in parallel; inputs the trajectory approaching the intersection, and outputs the complete intersection prediction trajectory.

[0029] Even better, when training and testing the prior-assisted parallel heuristic attention prediction model, the model is optimized and its performance is verified using a public dataset through a cascaded behavior prediction and trajectory prediction module.

[0030] The beneficial effects of this invention are:

[0031] 1. This invention improves prediction accuracy and forward time while enhancing the model's generalization ability and deployment efficiency, providing reliable support for autonomous driving in intersection scenarios, effectively reducing traffic accident risks and congestion, and has significant practical value and application prospects.

[0032] 2. This invention features high accuracy and long lead time: By using behavior prediction results as prior aids for trajectory prediction, this invention significantly improves the accuracy and lead time of vehicle behavior TPS and trajectory DTTP prediction at intersections. Practical application has proven that the model of this invention outperforms traditional discrete models, meeting the real-time decision-making requirements of autonomous driving.

[0033] 3. This invention has strong generalization ability: This invention extracts low-dimensional spatiotemporal features through a unified trajectory segmentation and sampling method, which is suitable for various scenarios including traffic lights-free and unstructured intersections; the model maintains stable performance under different intersection shapes, overcoming the limitation of insufficient generalization ability of traditional models.

[0034] 4. This invention features low computational complexity and high practicality: Through lightweight design and low-dimensional data processing, this invention significantly reduces the computational complexity and hardware requirements of the model; compared with complex neural networks based on high-dimensional input, the computational efficiency is significantly improved, and it is easy to deploy to simulation platforms or embedded systems, providing efficient support for the rapid testing and application of autonomous driving.

[0035] 5. This invention has integrated prediction capabilities: By integrating TPS behavior prediction and DTTP trajectory prediction, this invention constructs a priori-assisted parallel prediction system, which solves the problems of high development costs and poor consistency caused by the separation of behavior and trajectory prediction in traditional methods. It supports multiple low-dimensional input data and quickly generates prediction results for complex scenarios. Attached Figure Description

[0036] Figure 1This is a filtered trajectory map of moving vehicles at an intersection, based on a priori-assisted method for predicting vehicle behavior and trajectory at intersections, according to the present invention.

[0037] Figure 2 These are a schematic diagram of trajectory segmentation and a vehicle heading coordinate system diagram of the present invention;

[0038] Figure 3 This is a diagram showing the intersection trajectory segmentation result of the present invention;

[0039] Figure 4 This invention provides a trajectory diagram of the intersection corresponding to the sampling results of the present invention.

[0040] Figure 5 This is a diagram showing the detection of changes in vehicle kinematics data at intersections according to the present invention.

[0041] Figure 6 This is a schematic diagram of the steering intention prediction and trajectory segmentation of the present invention;

[0042] Figure 7 This is the AI-1-TI trajectory prediction network diagram of the present invention;

[0043] Figure 8 This is a diagram showing the lateral and longitudinal training and testing loss of intersection 2 according to the present invention;

[0044] Figure 9 This is a diagram showing the lateral and longitudinal training and testing loss of intersection 3 in this invention;

[0045] Figure 10 This is the LSTM-assisted FCN network diagram of the present invention;

[0046] Figure 11 This invention provides loss and accuracy graphs obtained during training and testing using longitudinal and lateral input data from intersection 2.

[0047] Figure 12 This is a loss and accuracy graph obtained during training and testing using longitudinal and lateral input data from intersection 3 according to the present invention;

[0048] Figure 13 This is a detailed structural diagram of the HAaEnDe of the present invention;

[0049] Figure 14 This is the optimized HAaEnDe structure diagram of the present invention;

[0050] Figure 15 This is the behavioral network correspondence diagram of the present invention;

[0051] Figure 16 This is the encoder-decoder diagram of the cascaded prior aid and parallel heuristic attention aid of the present invention. Detailed Implementation

[0052] The related technologies of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0053] refer to Figures 1-16 This embodiment provides a prior-aid-based method for predicting vehicle behavior and trajectory at intersections. By integrating TPS behavior prediction and DTTP trajectory prediction, it solves the problems of high development costs and poor consistency caused by model separation. By using behavior prediction results as prior-aid input for trajectory prediction, it resolves the contradiction between accuracy and forward-looking time. By unifying trajectory segmentation and sampling methods and lightweight design, it solves the problems of generalization ability and computational complexity. Through parallel processing and low-dimensional input, it supports the rapid generation of prediction results from multiple data sources.

[0054] This implementation method includes the following steps:

[0055] 1. By using trajectory data segmentation and sampling methods, the spatiotemporal features of the trajectory are extracted and the data dimensionality is reduced to generate high-fidelity low-dimensional data. This implementation method is also applicable to scenarios with no traffic lights and unstructured intersections.

[0056] 2. Design a fully connected neural network behavior prediction model that combines jump step prediction and LSTM assistance to predict left turn, right turn or straight-line intention, and use its output as a prior aid for trajectory prediction.

[0057] 3. Develop a heuristic attention-assisted encoder-decoder (HAaEnDe), which consists of multiple attention-assisted encoder-decoders (AaEnDe). It focuses on key trajectory features through an attention mechanism to achieve multi-stage trajectory prediction.

[0058] 4. A prior-assisted parallel heuristic attention prediction model was developed, which integrates TPS behavior prediction and DTTP trajectory prediction. The behavior prediction results are used as prior assistance. The model achieves prediction consistency and efficiency improvement through parallel structure and joint optimization. It supports multiple low-dimensional input data and is suitable for complex intersection scenarios.

[0059] The specific steps include the following detailed steps:

[0060] Step (1): Trajectory data segmentation and sampling

[0061] 1.1 Vehicle trajectory extraction

[0062] Based on the publicly available inD dataset, five files each from intersection 2 (structured intersection) and intersection 3 (unstructured intersection) were selected. After filtering out non-vehicle and stationary data, 1924 dynamic vehicle trajectories were obtained, visualized as follows: Figure 1 As shown.

[0063] 1.2 Vehicle trajectory segmentation

[0064] The process of a vehicle passing through an intersection is divided into three stages: approaching the intersection (AI), crossing the intersection (TI), and exiting the intersection (EI). See [link to relevant documentation]. Figure 2 (a). By constructing a coordinate system, see Figure 2 (b) The intersection trajectory is segmented by considering the vehicle's directional information within the intersection. The segmentation results are shown below. Figure 3 .

[0065] 1.3 Data Sampling and Dimensionality Assimilation

[0066] The segmented AI, TI, and EI stage trajectories were sampled using an integer multiple sampling method. When the number of samplings was 9, the TI stage trajectory retained the best curve features (see [link]). Figure 4 The linear trajectory is well-suited for both the AI ​​and EI stages, and is standardized to 9 coordinate points. If the number of points exceeds 9, the smoothed points in the latter half of the TI stage are removed; if the number of points is less than 9, they are supplemented using cubic spline interpolation.

[0067] Step (2): Design of behavior prediction model

[0068] Design a fully connected neural network that combines jump step prediction and LSTM assistance to predict left turn, right turn, or straight-ahead intentions, and output the results as prior aids for steps 3 and 4.

[0069] 2.1 Detection of Change Points in Vehicle Kinematics Data (VKD)

[0070] Based on the 9-node sampling trajectory from step 1, longitudinal velocity time-series features of the vehicle are extracted, and the deceleration completion point of the AI ​​phase is detected to determine the behavioral inflection point. (See...) Figure 5 The speed variation range was analyzed using a dynamic threshold method, and the fourth data point in the AI ​​stage was selected as the key change point. The 9-node trajectory in the AI ​​stage was divided into the AI-1 stage trajectory consisting of the first 4 data points and the AI-2 stage trajectory consisting of the last 5 data points. A schematic diagram of intersection turning intention prediction is shown below. Figure 6 .

[0071] 2.2 Prediction of the jump steps from AI-1 to TI trajectory

[0072] This step designs a jump step prediction network based on input-output characteristics, and then uses the jump step prediction network to predict the trajectory from AI-1 to TI.

[0073] 2.2.1 Skip Step Prediction Network

[0074] A two-stage skip-step prediction network is constructed, comprising a trajectory prediction module and a trajectory correction module. The trajectory prediction module improves upon the traditional LSTM-based encoder-decoder structure by replacing the first layer of the LSTM with a lighter GRU; the trajectory correction module is based on a deep neural network (DNN) trajectory correction framework. The network employs a skip-step strategy, skipping redundant data points in the AI ​​stage and directly predicting trajectory features in the TI stage. For the network structure, see [link to network structure]. Figure 7 The trajectory prediction module takes the trajectory of 4 data points from the AI-1 stage as input and outputs a preliminary trajectory point sequence for the TI stage. The trajectory correction module takes the preliminary trajectory point sequence for the TI stage and 12 kinematic parameters from the AI-1 stage (including the maximum, minimum, and average values ​​of lateral and longitudinal velocities and accelerations) as input and outputs the corrected TI stage trajectory.

[0075] 2.2.2 Prediction of Jumping Step Trajectory

[0076] Based on the skip step prediction network in step 2.2.1, this involves training and testing of AI-1 to TI trajectories, using all trajectory data from intersection 2 and intersection 3. Assume the input vehicle trajectory data (AI-1 trajectory data) and label data (TI vehicle trajectories) are represented as fixed-dimensional vectors x = {x1, x2, ... x...} n} and y = {y1, y2, ... y t When sequence x is input into the encoder, the encoder first computes the hidden state h at time t. t ,as follows:

[0077] h t =f(x) t h t-1 )

[0078] When the neural network is initially initialized, the weights are randomized, and the initial hidden state h1 is automatically calculated. As backpropagation optimizes, the weights are automatically adjusted. After encoding, the fixed-dimensional semantic vector c is represented as:

[0079] c = q(h1, h2, ... h t )

[0080] Subsequently, the decoder uses the semantic vector C = {C1, C2, ... C} n The prediction result y for time t t Decode and calculate:

[0081]

[0082] Based on the time series propagation characteristics of RNNs, considering the intermediate states of the decoder's hidden layer, the probability of a certain step y is:

[0083] P(y t |y1,y2,...y t-1 c)=g(y t-1 ,s t c)

[0084] Where s t Let g represent the hidden state of the decoder at time t, and g be a non-linear function. The probability of the final output y is:

[0085]

[0086] The attention mechanism introduces input data as a reference at the decoder output. If it is added to the i-th layer, the conditional probability at time t is:

[0087] P(y t |y1, y2, ...y t-1 x)=g(y t-1 s t c t )

[0088] Where s t This represents the hidden layer output of the RNN at time t:

[0089] s t =f(s) t-1 y t-1 c t )

[0090] After obtaining the encoder-decoder output, this output is combined with the 12 VKD parameters from the AI-1 stage to form the trajectory correction model input. Assume the DNN input of this system is the sequence {x1, x2, ... x...} n} represents vehicle motion data, and the output is {y1, y2, ... y n The sequence x represents the actual trajectory during the TI phase. Before the sequence x is input into the DNN, the DNN weights initially follow a truncated normal distribution, and the weights are gradually adjusted during training. The value of each neuron is calculated forward:

[0091]

[0092] in This represents the weight of the i-th synapse of the j-th neuron in layer l during the nth iteration. The output of the previous layer is: The output of the i-th neuron at the end is:

[0093]

[0094] The DNN error is:

[0095]

[0096] After the forward propagation calculation is completed, the backward propagation optimizes the weights. The gradient calculation of the j-th neuron at the n-th iteration of the l-th layer is expressed as:

[0097]

[0098] where l = L is the input layer, l < L is the hidden layer, j is the neuron, and the weight update is:

[0099]

[0100] where α is the momentum constant and η is the learning rate, which is automatically optimized by TensorFlow.

[0101] The training and test results are shown in Figure 8 and Figure 9 , the jump step prediction network converges quickly, the convergence mean square error (C-MSE) is low, and the trajectory correction module has an increasing loss at the beginning of training but finally converges.

[0102] The present invention measures the single-data calculation time cost (SDCTC) of jump step prediction and also proposes a method to reflect the actual error (AE), and the calculation formula is as follows: [[ID=3�]]

[0103]

[0104] where represents the i-th coordinate of the true trajectory, represents the i-th coordinate of the predicted trajectory, and AE j represents the AE of the j-th trajectory. The results are shown in Table 1, and the jump step prediction network achieves good prediction results.

[0105] Table 1

[0106]

[0107] 2.3. Steering intention prediction

[0108] Use the TI trajectory predicted in step 2.2 to classify the steering behavior and deduce the vehicle steering intention.

[0109] 2.3.1. LSTM-assisted fully connected network.

[0110] The present invention constructs an LSTM-assisted fully connected network (FCN), replaces the first layer of the traditional FCN with LSTM, and the structure is shown in Figure 10Detailed information about the model is shown in Table 2. Step 2.2 of the LSTM module processes the temporal dependencies of the predicted TI trajectory and kinematic data, integrates the features through a fully connected network, and outputs the probabilities of left turn, right turn, and straight ahead intentions.

[0111] Table 2

[0112]

[0113] 2.3.2 Prediction of turning intentions.

[0114] The LSTM-assisted FCN, using data from intersections 2 and 3, was trained and tested with longitudinal and lateral trajectories as inputs. The results are shown below. Figure 11 and Figure 12 The accuracy of the longitudinal trajectory input is higher than that of the lateral trajectory, and the loss value is lower. The accuracy of the model at intersections 2 and 3 is shown in Table 3. A cascaded jump step prediction network and an LSTM-assisted FCN are used to construct an intersection turning intention prediction model.

[0115] Table 3

[0116]

[0117] Step (3): Heuristic attention trajectory prediction

[0118] We developed a heuristic attention-assisted encoder-decoder (HAaEnDe) structure, which consists of multiple attention-assisted encoder-decoders (AaEnDe) to predict trajectories in stages.

[0119] 3.1 Trajectory Prediction Model Design

[0120] 3.1.1 HAaEnDe Network

[0121] Design a heuristic attention-assisted encoder-decoder (HAaEnDe) network based on LSTM and GRU, connecting the AI-1, AI-2, TI, and EI stages. Its structure is shown below. Figure 13 During training, the output of each stage serves as the input for the next stage. The training labels are AI-2, TI, and EI stage trajectories, while testing and validation use only AI-1 input. Trajectory prediction for all four stages during turning was tested, and after obtaining the optimal attention-assisted encoder-decoder structure for each stage, an optimal heuristic network was constructed. The attention-assisted encoder-decoder structures for the four stages have similar structures; this invention only details the computational process of the attention mechanism.

[0122] After the encoder-decoder uses the attention mechanism, the semantic vector changes dynamically, calculated as follows:

[0123]

[0124] Where T is the length of the current hidden state, and h j Let a be the state of the hidden layer at step j. tj The attention weight at time t can be expressed as:

[0125]

[0126] e tj Measure s t-1 with h j The correlation is calculated as follows:

[0127] e tj =score(s t-1 h j )=Vσ(W s-1 +Uh j )

[0128] U, V, and W are weight matrices, and σ is the activation function. This invention reduces data complexity through trajectory pre-segmentation and prior assistance, eliminating the need for more advanced attention mechanisms.

[0129] 3.1.2 Optimization of HAaEnDe Network Structure

[0130] To improve trajectory prediction accuracy, this step involves systematically optimizing the attention-assisted encoder-decoder (AaEnDe) structure through experiments for different driving behaviors (left turn TL, right turn TR, straight ahead SA) and trajectory stages (AI-1→AI-2, AI→TI, TI→EI). The process is as follows:

[0131] By designing 54 sets of comparative experiments covering 3 trajectory stages (AI-1→AI-2, AI→TI, TI→EI), 3 driving behaviors (TL, TR, SA), 2 intersection types (structured intersection 2, unstructured intersection 3), and 3 candidate network structures (see Table 4), the optimal AaEnDe structure for each stage was determined using actual error (AE) as the evaluation index, thus constructing a heuristic attention-assisted encoder-decoder (HAaEnDe) network.

[0132] Table 4

[0133]

[0134] Candidate network structures include:

[0135] Ge-R-Ld-TD (GRU encoder-repeating vector-LSTM decoder)

[0136] Ge-ATT-Ld-TD (GRU encoder-attention vector-LSTM decoder)

[0137] Re-ATT-Rd-TD (RNN Encoder-Attention Vector-RNN Decoder)

[0138] This invention adds a wrapper layer at the end of each AaEnDe to unify the output data dimension and ensure consistency of output across AI-1, AI-2, TI, and EI stages. The optimized HAaEnDe structure is shown below. Figure 14 For each behavior-stage combination (such as the AI→TI stage of TL), the optimal HAaEnDe structure is determined by the following rules: If the optimal AaEnDe structure is the same at a certain stage at structured intersection 2 and unstructured intersection 3, then that structure is directly selected; if the optimal structures of the two intersections are different, then the average AE of the two intersections is calculated, and the AaEnDe structure with the minimum average AE is selected. Based on the above rules, 18 optimal AaEnDe structures (3 behaviors × 3 stages) are generated and summarized in Table 5.

[0139] Table 5

[0140]

[0141] By calculating the performance differences of 54 AaEnDe configurations, Table 5 summarizes the 18 optimal AaEnDe structures. Figure 15 This demonstrates the dedicated AaEnDe structure branches triggered by left turn (TL), right turn (TR), and straight ahead (SA) actions. The network input types and quantities are shown in Table 6, where "√4+4" indicates 4 horizontal coordinates and 4 vertical coordinates, and "×" indicates no data in the network at that stage.

[0142] Table 6

[0143]

[0144] 3.2 Parallelized Heuristic Trajectory Prediction

[0145] Using the optimized heuristic attention-assisted encoder-decoder (HAaEnDe) network from step 3.1.2, multi-stage trajectory prediction is achieved through parallel processing. The key feature is that the behavior prediction result (left turn, right turn, straight ahead) is used as a switch signal to activate the corresponding TL, TR, and SA branches of the attention-assisted encoder-decoder (AaEnDe) network in parallel. Each stage of the encoder-decoder operates independently. The first stage inputs the original trajectory AI-1, and subsequent stages input the output of the previous stage. This, combined with the 9-node sampling in step 1.3, reduces computational overhead. Figure 15 Display the logic relating behavior to branches of the AaEnDe network.

[0146] Step (4): Ensemble prediction of prior-assisted parallel heuristic attention prediction model

[0147] 4.1 Design of a Prior-Assisted Parallel Heuristic Attention Prediction Model

[0148] By fusing the TPS output intent prediction (left turn, right turn, straight-ahead intent) from step 2.3 and the DTTP multi-stage trajectory prediction from step 3.2, a priori-assisted parallel structure is constructed. The behavior prediction results are used as switching signals to activate the corresponding attention-assisted encoder-decoder (AaEnDe) branches in parallel: left turn behavior activates the TL branch, right turn behavior activates the TR branch, and straight-ahead behavior activates the SA branch. The structure is shown in [see details]. Figure 16 In model validation, the AI-1 trajectory is input, and the complete intersection prediction trajectory is output.

[0149] 4.2 Model Training and Testing

[0150] The prior-assisted parallel heuristic attention prediction model is trained and tested, characterized by the use of the inD dataset for model optimization and performance verification through cascaded behavior prediction and trajectory prediction modules. The training process uses 75% of the data for training and 25% for testing. The input is a 9-node sampled trajectory from the AI-1 stage, and the output includes behavior prediction results (left turn, right turn, straight-ahead intention) and the complete intersection prediction trajectory.

[0151] Training Process: First, the LSTM-assisted fully connected network (FCN) in step 2.3, using the AI-1 trajectory and 12 kinematic parameters as input, trains the behavior prediction module, outputting the probabilities of left turn (TL), right turn (TR), and straight ahead (SA) intentions. The behavior prediction results serve as a switch signal, triggering the attention-assisted encoder-decoder (AaEnDe) branches of the TL, TR, and SA branches in the HAaEnDe network optimized in step 3.2. Each AaEnDe branch is trained independently. The first stage inputs the original AI-1 trajectory, and subsequent stages (AI-2, TI, EI) use the output of the previous stage as input, with the training labels being the actual trajectory coordinates. The model uses TensorFlow to optimize the weights, with adaptive adjustments to the learning rate and momentum constant to ensure rapid convergence.

[0152] Test Results: In the test, the model was given a 9-node sampled trajectory from the AI-1 stage. First, the behavior prediction module generated left-turn (TL), right-turn (TR), or straight-ahead (SA) intent probabilities. Then, based on the intent probabilities, the corresponding TL, TR, and SA branch attention-assisted encoder-decoder (AaEnDe) was activated to generate predicted trajectories for the AI-2, TI, and EI stages, outputting the complete intersection trajectory. The test used 25% of the test data from the inD dataset, validating on structured intersections (intersection 2) and unstructured intersections (intersection 3). The test procedure is as follows: Figure 16 .

[0153] In summary, this invention improves prediction accuracy and forward time while enhancing the model's generalization ability and deployment efficiency, providing reliable support for autonomous driving in intersection scenarios, and effectively reducing traffic accident risks and congestion. Therefore, this invention has broad application prospects.

[0154] It should be emphasized that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for predicting vehicle behavior and trajectory at intersections based on prior assistance, characterized in that, Includes the following steps: Step S1: By segmenting and sampling trajectory data, extract the spatiotemporal features of the trajectory and reduce the data dimensionality to generate high-fidelity low-dimensional data; Step S2: By combining jump step prediction and LSTM-assisted fully connected neural network behavior prediction model, predict left turn, right turn or straight-line intention. The output of the fully connected neural network behavior prediction model is used as a priori aid for trajectory prediction. Step S3: Focus on key trajectory features through an attention mechanism to achieve multi-stage trajectory prediction. Step S4: Integrate predictions using a priori-assisted parallel heuristic attention prediction model.

2. The method for predicting vehicle behavior and trajectory at intersections based on prior assistance as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S1-1: Vehicle trajectory extraction. Based on the public dataset, select structured intersection and unstructured intersection files, filter out non-vehicle and stationary data, and obtain multiple dynamic vehicle trajectories. Step S1-2: Vehicle trajectory segmentation. The process of a vehicle passing through an intersection is divided into three stages: approaching the intersection, crossing the intersection, and leaving the intersection. By constructing a coordinate system and considering the vehicle's directional information in the intersection, the trajectory of the intersection is segmented. Steps S1-3: Data sampling and homogenization. Integer multiple sampling is used to sample the segmented stage trajectories.

3. The method for predicting vehicle behavior and trajectory at intersections based on prior assistance as described in claim 1, characterized in that, In step S2, the jump step prediction includes: a jump step prediction network and a jump step trajectory prediction; The jump step prediction network includes a trajectory prediction module and a trajectory correction module. The trajectory prediction module improves the LSTM-based encoder-decoder structure by replacing the first layer of the LSTM with a lighter GRU. The trajectory correction module is based on a deep neural network trajectory correction framework. The jump step trajectory prediction is based on a jump step prediction network. Through trajectory training and testing, using all trajectory data from structured and unstructured intersections, preliminary trajectory prediction results are obtained.

4. The method for predicting vehicle behavior and trajectory at intersections based on prior assistance according to claim 3, characterized in that, In step S2, the LSTM-assisted fully connected neural network behavior prediction model constructs an LSTM-assisted fully connected network FCN, replacing the first layer of the traditional FCN with an LSTM. The fully connected network integrates features and outputs the probability of intention to turn left, turn right, or go straight. A cascaded skip step prediction network and an LSTM-assisted FCN are used to construct a turning intention prediction model for intersections.

5. The method for predicting vehicle behavior and trajectory at intersections based on prior assistance according to claim 1, characterized in that, In step S3, a heuristic attention-assisted encoder-decoder structure is developed, which consists of multiple attention-assisted encoder-decoders.

6. The method for predicting vehicle behavior and trajectory at intersections based on prior assistance according to claim 5, characterized in that, Design a heuristic attention-assisted encoder-decoder network based on LSTM and GRU to connect the various stages; During training, the output of each stage serves as the input for the next stage, and the training label is the stage trajectory. The trajectory prediction of the four stages during the test turning process is used, and after obtaining the optimal attention-assisted encoder-decoder structure for each stage, the optimal heuristic network is formed. The attention-assisted encoder-decoder of the four stages has a similar structure.

7. The method for predicting vehicle behavior and trajectory at intersections based on prior assistance according to claim 6, characterized in that, For different driving behaviors and trajectory stages, the heuristic attention-assisted encoder-decoder network structure was optimized through systematic experiments.

8. The method for predicting vehicle behavior and trajectory at intersections based on prior assistance according to claim 1, characterized in that, In step S4, the prior-assisted parallel heuristic attention prediction model integrates the output intention prediction and multi-stage trajectory prediction to construct a prior-assisted parallel structure, and uses the behavior prediction result as a switch signal to activate the corresponding attention-assisted encoder-decoder branch in parallel. Input the trajectory of the approaching intersection, and output the complete predicted trajectory of the intersection.

9. The method for predicting vehicle behavior and trajectory at intersections based on prior assistance as described in claim 8, characterized in that, During training and testing, the prior-assisted parallel heuristic attention prediction model utilizes a cascaded behavior prediction and trajectory prediction module to optimize the model and verify its performance using a public dataset.

Images