Fine-grained expressway automatic driving vehicle lane changing track prediction method

Abstract

The invention belongs to the technical field of vehicle trajectory prediction, and relates to a fine-grained highway automatic driving vehicle lane changing trajectory prediction method, which comprises the following steps: step 1, a data processing stage: data acquisition and screening, lane changing process segmentation and data same-dimensional processing are included; 2, a lane changing track prediction scheme: designing a heuristic attention-assisted encoder, determining an optimal structure of each stage of lane changing, finally constructing a network and verifying the network, and performing error analysis; 3, vehicle kinematics data prediction: analyzing the vehicle kinematics data, and selecting a prediction method to predict the kinematics data; according to the lane changing track segmentation and sampling algorithm, the lane changing process is divided into four stages, a data and behavior analysis basis is provided for a track prediction model and other related researches of lane changing, and the lane changing process can be deeply understood.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle trajectory prediction, and relates to a fine-grained highway autonomous driving vehicle lane change trajectory prediction method. Background Technique

[0002] Lane change is one of the most basic driving behaviors in the field of autonomous driving. Improper lane change may cause collisions, resulting in property losses and even casualties. As an important traffic scenario, highways have characteristics such as fast vehicles, diverse vehicle types, and complex lane change patterns, which make lane change behaviors on highways play a decisive role in traffic safety.

[0003] With the increasing maturity of the development of the vehicle networking, advanced driving assistance systems with vehicle-to-vehicle and vehicle-to-infrastructure functions have been applied to intelligent connected vehicles and will be applied to autonomous driving vehicles in the future. If the future driving trajectories of surrounding vehicles are provided to the drivers in intelligent connected vehicles or the autonomous driving systems in autonomous driving vehicles through advanced driving assistance systems, some key tasks (such as forward collision avoidance, trajectory planning, etc.) can be executed in advance, thereby improving traffic safety and efficiency.

[0004] Lane change is divided into two stages, namely the lane change decision (LCD) stage and the lane change implementation (LCI) stage. Lane change decision refers to operations such as judgment and prediction before a vehicle performs a lane change action. For example, similar studies such as lane change feasibility analysis, lane change behavior prediction, and lane change intention recognition can be classified as studies in the lane change decision stage; while operations during the period from the start to the end of the lane change can be classified as studies in the lane change implementation stage. For example, similar studies such as lane change trajectory prediction, lane change trajectory planning, and lane change recognition can be classified as studies in the lane change implementation stage. Since lane change decision belongs to one-to-many classification, its mode is simpler than the lane change implementation stage which belongs to many-to-many regression. Many studies have achieved quite high recognition and prediction accuracies in the research of lane change decision. Compared with lane change decision, there are also studies on trajectory prediction in the lane change implementation stage.

[0005] However, existing studies on trajectory prediction in the lane change implementation stage focus on the macroscopic prediction of the entire road trajectory. In addition to the short prediction lead time and short prediction distance, the lane change process is not predicted and finely described, resulting in the inability to track the position of the vehicle during the lane change process.

[0006] Therefore, a method or device capable of improving the accuracy and reliability of vehicle lane change trajectory prediction is needed to solve the deficiencies of existing models in terms of prediction accuracy, long-term prediction ability, and fine description of the lane change process. Summary of the Invention

[0007] The object of the present invention is to provide a method for predicting the lane-changing trajectory of highway autonomous driving vehicles based on a heuristic attention-assisted codec model, so as to solve the deficiencies of existing models in terms of prediction accuracy, long-term prediction ability, and fine description of the lane-changing process, improve the accuracy and reliability of predicting the lane-changing trajectory of highway autonomous driving vehicles, and provide more effective technical support for intelligent transportation systems.

[0008] The technical solution adopted by the present invention to solve the technical problems is: a fine-grained method for predicting the lane-changing trajectory of highway autonomous driving vehicles, including the following steps:

[0009] A fine-grained method for predicting the lane-changing trajectory of highway autonomous driving vehicles, including the following steps:

[0010] Step 1: Data processing stage: including data acquisition and screening, lane-changing process segmentation, and data dimensionality processing;

[0011] Step 2: Lane-changing trajectory prediction scheme: design a heuristic attention-assisted encoder, then determine the optimal structure for each stage of lane-changing, finally construct a network and verify it, and conduct error analysis;

[0012] Step 3: Prediction of vehicle kinematic data: analyze vehicle kinematic data and select a prediction method to predict the kinematic data.

[0013] Preferably, the specific steps of Step 1 include:

[0014] Step 1-1: Data acquisition and screening, select the HighD natural driving dataset containing highway vehicle driving information for model development, and the HighD natural driving dataset includes vehicle position coordinates and motion state information;

[0015] Step 1-2: Lane-changing trajectory extraction, select sections with different shapes from the highway sections in the HighD natural driving dataset to obtain lane-changing data samples and obtain lane-changing trajectories;

[0016] Step 1-3: Data acquisition and screening, take time as the abscissa, and analyze the lateral speed, longitudinal speed, lateral acceleration, and longitudinal acceleration data in the HighD natural driving dataset as the ordinate. According to the lane intersection points, peaks and valleys of longitudinal acceleration, the lane-changing process is divided into four stages: the LC-B stage before lane-changing, the LC-1 stage not entering the target lane, the LC-2 stage entering the target lane, and the LC-A stage after lane-changing, and preprocess the data for each stage;

[0017] Step 1-4: Data dimensionality processing, perform data sampling by the head-to-tail sampling method. Assume that the original trajectory consists of n coordinates, and the calculation formula for the sampling points is given:

[0018]

[0019] In formula (1), sp1…sp t represent the sampling points in the vehicle trajectory, and x1…x n represent the respective coordinate points from the starting point to the ending point in the vehicle trajectory, where x1 represents the starting point of the trajectory and x n represents the ending point of the trajectory.

[0020] More preferably, step 2 specifically includes:

[0021] Step 2-1: Heuristic attention-assisted codec design. Use LSTM and GRU to design a heuristic attention-assisted codec, and link the LC-B stage, LC-1 stage, LC-2 stage, and LC-A stage in sequence; during the training process, the output of the previous stage is used as the input of the next stage, and there are training labels in the LC1 stage and LC2 stage; only the input of the LC-B stage and the output of the LC-A stage are required in the testing and verification processes;

[0022] Step 2-2: Optimal structure selection; conduct verification experiments to test the lateral trajectory and the longitudinal trajectory respectively; during training, input the sampling coordinates of the previous stage and output the coordinates of the next stage. After structure testing and index evaluation, determine the optimal lateral structure and the optimal longitudinal structure, and finally construct a heuristic network accordingly.

[0023] More preferably, in step 2-1, the hidden state h of the encoder at time t t is:

[0024] h t = f(x t , h t-1 ) (2)

[0025] In formula (2), x t represents the vehicle motion data or state information input to the model at time t, and h t-1 represents the hidden state of the encoder at time t-1;

[0026] The semantic vector c with a fixed dimension is:

[0027] c = q(h1, h2…h t ) (3)

[0028] The probability calculation formula is:

[0029]

[0030] In formula (5), y t-1 represents the prediction result at time t-1, s tDenotes the state at time t, which usually represents the hidden state of the model or the current state of the decoder;

[0031] The semantic vector after adding the attention mechanism is:

[0032]

[0033] In Equation (6), T represents the current length of the hidden state, a tj Denotes the attention weight at time t, h j Denotes the encoder hidden state at time j, which usually represents the output of the model at that time or the memory of the input sequence.

[0034] More preferably, the structure test and index evaluation include: MSEAC, CE, SDCToT, AE.

[0035] More preferably, step 3 specifically includes:

[0036] Step 3-1: Vehicle kinematic data analysis. After separating the HighD natural driving dataset, we obtain: lateral velocity, lateral acceleration, longitudinal velocity, and longitudinal acceleration; extract and predict the steering angle of the vehicle in the LC-1 stage and LC-2 stage, and combine the predicted trajectory of the heuristic network to achieve real-time tracking of the vehicle position, providing a more accurate decision-making basis for the vehicle safety algorithm;

[0037] Step 3-2: Vehicle kinematic data prediction method. For sequence-to-sequence prediction, a heuristic network is used, and for many-to-one prediction, a deep neural network is used; when predicting the steering angle, the lateral velocity output by the heuristic network is utilized;

[0038] Step 3-3: Output of the vehicle kinematic data prediction result.

[0039] The beneficial effects of the present invention are:

[0040] 1. The lane-changing trajectory segmentation and sampling algorithm proposed by the present invention divides the lane-changing process into four stages, providing a data and behavior analysis basis for the trajectory prediction model and other lane-changing related research, and helping to more deeply understand the lane-changing process.

[0041] 2. The heuristic attention-assisted encoder-decoder network of the present invention can effectively predict vehicle trajectories and kinematic data, and has a reasonable computational load, improving the practicality of the prediction model and enabling more accurate predictions without adding too much computational burden.

[0042] 3. The cascade prediction model of the present invention combines trajectory prediction and vehicle kinematic data prediction, featuring high integration, long-term prediction ability, strong fault tolerance, and high modifiability. It can more comprehensively and accurately describe the lane-changing process, provide a more reliable basis for vehicle control decisions, and contribute to improving the safety and efficiency of autonomous vehicles on highways. Brief Description of the Drawings

[0043] Figure 1 is a segmented result diagram of four randomly selected trajectories of a fine-grained highway autonomous vehicle lane-changing trajectory prediction method of the present invention;

[0044] Figure 2 is a schematic diagram of the steering angle of the present invention;

[0045] Figure 3 is a schematic diagram of the variation of the first lateral prediction error of the present invention at different time and space positions;

[0046] Figure 4 is a schematic diagram of the variation of the second lateral prediction error of the present invention at different time and space positions;

[0047] Figure 5 is a schematic diagram of the steps of the present invention. Detailed Embodiment

[0048] Next, the related technologies in the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0049] The technical solutions adopted by the present invention include the following steps:

[0050] Refer to Figures 1 - 5 shown, the detailed embodiment includes the following steps:

[0051] Step 1, data processing stage: including data acquisition and screening, lane-changing process segmentation, and data dimensionality processing;

[0052] Step 2, lane-changing trajectory prediction scheme: design a heuristic attention-assisted encoder, then determine the optimal structure for each stage of lane-changing, finally construct a network and verify it, and perform error analysis;

[0053] Step 3, vehicle kinematic data prediction: analyze vehicle kinematic data and select a prediction method to predict the kinematic data.

[0054] Specifically include:

[0055] Data Acquisition and Data Preprocessing

[0056] Step 1.1: Data Acquisition and Screening

[0057] Since the proposed model requires high-precision vehicle position coordinates and motion state information, the HighD natural driving dataset containing highway vehicle driving information developed by the IKA research team is selected for model development.

[0058] Step 1.2: Lane Change Trajectory Extraction

[0059] Six sections with different shapes are selected from 60 highway sections in the HighD natural driving dataset to obtain 1051 lane change data samples and obtain the lane change trajectory.

[0060] Step 1.3: Lane Change Process Segmentation

[0061] Taking time as the abscissa and the lateral speed, longitudinal speed, lateral acceleration, and longitudinal acceleration data in the HighD natural driving dataset as the ordinate for analysis, it is concluded that there is no obvious change in the lateral speed and acceleration when the vehicle changes lanes, while the longitudinal speed and acceleration change significantly and regularly, especially the longitudinal acceleration has obvious peaks and valleys. By randomly selecting a trajectory and comparing it with the longitudinal acceleration, it is further concluded that the longitudinal trajectory changes regularly with the longitudinal acceleration and the lane change trajectory is asymmetric. According to the lane intersection points, peaks and valleys of the longitudinal acceleration, the lane change process is divided into four stages: LC-B (before lane change), LC-1 (not entering the target lane), LC-2 (entering the target lane), and LC-A (after lane change). Figure 1 The segmentation results of four randomly selected trajectories are shown, and the segmentation effect is obvious and meets the expectations. After completing the lane change process segmentation, the data of each stage is preprocessed

[0062] Step 1.4: Data Homogeneous Dimension Processing

[0063] A head-to-tail sampling (H2T-sampling) method is proposed for data sampling. Assuming that the original trajectory consists of n coordinates (n is different for different vehicle trajectories), the calculation formula for the sampling points is given:

[0064]

[0065] Lane Change Trajectory Prediction Scheme

[0066] Step 2.1: Heuristic Attention-Assisted Encoder-Decoder Design

[0067] Use LSTM and GRU to design a heuristic attention-assisted encoder-decoder, divide the lane change process into four stages (LC-B, LC-1, LC-2, LC-A) and link them together;

[0068] During the training process, the output of each stage is used as the input of the next stage, and there are training labels in the LC1 and LC2 stages; only the LC-B input and LC-A output are required during the testing and validation processes.

[0069] After this decoder uses the attention mechanism, the semantic vector changes dynamically. For the input sequence x = {x1, x2... x n}, with the label y = {y1, y2... y n}, the hidden state h t of the encoder at time t is calculated as follows:

[0070] h t = f(x t , h t-1 ) (2)

[0071] The initial hidden state h1 is randomly initialized and automatically calculated when the neural network framework is first calculated, and the weights are optimized through backpropagation subsequently.

[0072] After encoding is completed, the semantic vector c with a fixed dimension is expressed as:

[0073] c = q(h1, h2... h t ) (3)

[0074] The decoder uses the semantic vector C = {C1, C2,... C n} to calculate the prediction result y t at time t, and the probability calculation formula is:

[0075]

[0076] When considering the intermediate state of the decoder hidden layer, P(y t | y1, y2,... y t-1 , c) = g(y t-1 , s t , c), and the probability of the final output y is:

[0077]

[0078] After the attention mechanism is added, the semantic vector becomes:

[0079]

[0080] (T is the current length of the hidden state, a tj is the attention weight at time t e tj = score(s t-1 , h j ) = Vσ(Ws t-1 + Uh j), where V, W, and U are weight matrices to be trained by the neural network, and σ is the activation function)

[0081] Step 2.2: Optimal structure selection

[0082] To obtain the optimal attention-aided codec structures for the four lane-changing stages, 48 validation experiments were carried out to test the lateral and longitudinal trajectories respectively. Taking the LC-B to LC-1 stage as an example, during training, the sampled coordinates of the LC-B stage were input, and the coordinates of the LC-1 stage were output. After testing with various structures and evaluating indicators (including MSEAC, CE, SDCToT, AE, etc.), the optimal lateral structure was determined to be LeALdT, and the longitudinal one was GeRLdT. A similar method was used to determine the optimal structure for the LC-1 to LC-2 stage (the loss change is not shown). Through analysis, it was determined that the optimal lateral structure for the LC-2 to LC-A stage was LeAGdT and the longitudinal one was GeRLdT. Finally, a heuristic network was constructed accordingly.

[0083] Vehicle kinematic data prediction

[0084] Step 3.1: Analysis of vehicle kinematic data

[0085] The HighD natural driving dataset contains two types of vehicle kinematic data, speed and acceleration. After separating them into lateral and longitudinal directions, four types of data are obtained. It is found that the lateral speed and acceleration are generally stable with small fluctuations, while the longitudinal speed and acceleration fluctuate regularly and have different distributions in the four lane-changing stages. Since the vehicle steering angle changes in the LC-1 and LC-2 stages, it is difficult to finely describe the vehicle state with the macroscopic trajectory, and the lane changes in these two stages are asymmetric. Therefore, it is necessary to extract and predict the steering angles in these two stages. A straight line is formed using the critical points from LC-B to LC-1 and LC-1 to LC-2, and its slope is used as the vehicle steering angle (as Figure 2 shown). Combining with the trajectory predicted by the heuristic network, real-time tracking of the vehicle position can be achieved, providing a more accurate decision-making basis for vehicle safety algorithms

[0086] Step 3.2: Vehicle kinematic data prediction method

[0087] For sequence-to-sequence prediction, a heuristic network is used, and for many-to-one prediction, a deep neural network is used. There are corresponding historical data for speed and acceleration, and their prediction methods are similar to the trajectory prediction method, with the LC-B stage as the input to predict subsequent relevant data. When predicting the steering angle, the lateral speed output by the heuristic network is used. Specifically, the speed sequence of the LC-B stage is input into the heuristic network and the DNN simultaneously to predict the speed and the first steering angle of the LC-1 stage, and then the speed of the LC-1 stage is used to predict the second steering angle.

[0088] The input of the experimental DNN is the kinematic data sequence x = {x1, x2…x n}, the output is the trajectory correction matrix y = {y1, y2... y n} Before the input sequence, the DNN weights are initialized as random numbers obeying a truncated normal distribution and optimized during the training process. When inputting the sequence, through the formula (where represents the weight of the i-th synapse of the j-th neuron in the l-th layer at the n-th iteration, is the output of the i-th neuron; when l = 1, y i = x i ), calculate the output of each neuron. After completing the forward propagation calculation, the DNN optimizes the weights through backpropagation. In the backpropagation calculation method

[0089] represents the gradient of the l-th neuron at the n-th iteration, and its calculation method is:

[0090]

[0091] Step 3.3: Prediction results of vehicle kinematic data.

[0092] Table 1

[0093] Test index Average value of mean square error Computational efficiency Time required for each data point calculation (microseconds) LAT 0.002711 LAT: 290 LAT: 6.125 LONG 0.003125 LONG: 350 LONG: 5.973 SA1 0.125 SA1: 7600 SA1: 0.5236 SA2 1.325 SA2: 22,000 SA2: 0.6215

[0094] Table 2

[0095] Actual prediction error Lateral data Longitudinal data Trajectory error 1.225 meters 0.4571 meters Velocity error 0.7367 m / s 0.3426 m / s Acceleration error <![CDATA[0.2411 m / s 2 > <![CDATA[0.0577 m / s 2 >

[0096] Table 1 shows the lateral and longitudinal test metrics of the cascaded prediction model, demonstrating the test results of the cascaded prediction model on lateral and longitudinal data, including errors and computational efficiency at different stages. Table 2 shows the actual prediction errors of different data types, used to evaluate the performance of the model in lateral and longitudinal predictions. The magnitude of the prediction error reflects the accuracy of the model in each dimension. The errors of the lateral trajectory and speed are relatively large because these variables are more affected by complex road and driving behaviors; while the errors of the longitudinal acceleration and speed are relatively small, indicating that the model's prediction of longitudinal motion is relatively accurate. Figure 3 、 Figure 4 Two figures show the variation of the lateral prediction error at different time and space positions, helping to analyze the error distribution of the model in different situations. Figure (a) mainly focuses on the time dimension, while Figure (b) focuses on the space dimension.

[0097] In summary, the lane-changing trajectory segmentation and sampling algorithm of the present invention divides the lane-changing process into four stages, providing a data and behavior analysis basis for the trajectory prediction model and other lane-changing related research, and helping to understand the lane-changing process more deeply.

[0098] It should be emphasized that the above are only the preferred embodiments of the present invention, and there is no restriction on the present invention in any form. Any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A fine-grained lane-changing trajectory prediction method for highway autonomous vehicles, characterized in that It includes the following steps: Step 1: Data processing stage: including data acquisition and screening, lane change process segmentation, and data dimensionality processing; Step 2: Lane change trajectory prediction scheme: design a heuristic attention-aided encoder, then determine the optimal structure for each stage of the lane change, finally construct and verify the network, and conduct error analysis; Step 3: Prediction of vehicle kinematic data: analyze vehicle kinematic data and select a prediction method to predict the kinematic data.

2. The fine-grained lane-changing trajectory prediction method for highway autonomous vehicles according to claim 1, wherein, The specific content of Step 1 includes: Step 1-1: Data acquisition and screening, select the HighD natural driving dataset containing highway vehicle driving information for model development. The HighD natural driving dataset includes vehicle position coordinates and motion state information; Step 1-2: Lane change trajectory extraction, select sections with different shapes from the highway sections in the HighD natural driving dataset to obtain lane change data samples and obtain lane change trajectories; Step 1-3: Data acquisition and screening, use time as the abscissa, and analyze the lateral speed, longitudinal speed, lateral acceleration, and longitudinal acceleration data in the HighD natural driving dataset as the ordinate. According to the lane intersection points, peaks and valleys of the longitudinal acceleration, divide the lane change process into four stages: the LC-B stage before lane change, the LC-1 stage not entering the target lane, the LC-2 stage entering the target lane, and the LC-A stage after lane change, and preprocess the data for each stage; Step 1-4: Data dimensionality processing, use the head-to-tail sampling method for data sampling. Assume that the original trajectory consists of n coordinates, and the calculation formula for the sampling points is given: In formula (1), sp1...sp t represent the sampling points in the vehicle trajectory, and x1...x n represent the respective coordinate points from the starting point to the ending point in the vehicle trajectory, where x1 represents the starting point of the trajectory and x n represents the ending point of the trajectory.

3. A fine-grained lane-changing trajectory prediction method for highway autonomous vehicles according to claim 2, characterized in that The specific content of Step 2 includes: Step 2-1: Design of heuristic attention-aided codec, use LSTM and GRU to design a heuristic attention-aided codec, and link the LC-B stage, LC-1 stage, LC-2 stage, and LC-A stage in sequence; during training, the output of the previous stage is used as the input of the next stage, and there are training labels for the LC1 stage and LC2 stage; only the input of the LC-B stage and the output of the LC-A stage are required during the test and verification process; Step 2-2: Selection of the optimal structure; conduct verification experiments, and test the lateral trajectory and longitudinal trajectory respectively; during training, input the sampling coordinates of the previous stage and output the coordinates of the next stage. After structure testing and index evaluation, determine the optimal lateral structure and optimal longitudinal structure, and finally construct a heuristic network accordingly.

4. A fine-grained lane-changing trajectory prediction method for highway autonomous vehicles according to claim 3, characterized in that, In the above step 2-1, the hidden state h of the encoder at time t t is as follows: h t = f(x t , h t-1 ) (2) In formula (2), x t represents the vehicle motion data or state information input into the model at time t, and h t-1 represents the hidden state of the encoder at time t - 1; The semantic vector c with a fixed dimension is: c = q(h1, h2…h t ) (3) The probability calculation formula is: In formula (5), y t-1 represents the prediction result at time t - 1, and s t represents the hidden state or system state at time t; The semantic vector after adding the attention mechanism is: In Equation (6), T represents the current length of the hidden state, a tj represents the attention weight at time t, and h j represents the encoder hidden state at time j.

5. A fine-grained lane-changing trajectory prediction method for highway autonomous vehicles according to claim 3, characterized in that, The structure testing and index evaluation include: MSEAC, CE, SDCToT, AE.

6. A fine-grained lane-changing trajectory prediction method for highway autonomous vehicles according to claim 2, characterized in that, The specific content of Step 3 includes: Step 3-1: Analysis of vehicle kinematic data, after separating the HighD natural driving dataset, obtain: lateral speed, lateral acceleration, longitudinal speed, and longitudinal acceleration; extract and predict the steering angles of the vehicle in the LC-1 stage and LC-2 stage, and combine with the trajectory predicted by the heuristic network to achieve real-time tracking of the vehicle position, providing a more accurate decision-making basis for the vehicle safety algorithm; Step 3-2: For the vehicle kinematic data prediction method, a heuristic network is used for sequence-to-sequence prediction, and a deep neural network is used for many-to-one prediction; when predicting the steering angle, the lateral velocity output by the heuristic network is utilized. Step 3-3: Output the prediction result of the vehicle kinematic data.

Images