Method for generating vehicle track by combining natural driving data extraction features with driving conditions

Abstract

The invention belongs to the technical field of automatic driving, and relates to a method for generating a vehicle trajectory by combining natural driving data extraction features with driving conditions, which comprises the following steps of: 1, obtaining an equal number coordinate and a high-fidelity sampling trajectory through a generalized trajectory sampling method-integral multiple sampling algorithm; 2, performing correlation probability statistics on different vehicle kinematics parameters to generate a reasonable combination of the human-like vehicle kinematics parameters; 3, on the basis of a learned batch trajectory generation model, different human-like vehicle trajectories are generated based on different kinematics parameter combinations; the trajectory generated by the method has a more complex motion mode than the original trajectory, and can provide wider background traffic flow coverage for actual measurement ADS and ADAA; the track generated by the method can be quickly deployed on a simulation platform, so that the cost of real data acquisition is reduced while the test process is accelerated.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and relates to a method for generating a vehicle trajectory by combining driving conditions with features extracted from natural driving data. Background Art

[0002] During virtual simulation testing (VST), non-player character vehicles (NPC-Vs) need to provide information about different traffic flow conditions to autonomous vehicles to verify the liability of the autonomous vehicles when dealing with complex traffic flow conditions. If the NPC-Vs are also driven by the autonomous driving system (ADS) under test, then the ADS test loses its meaning. In addition, it is expected that autonomous vehicles will coexist with human-driven vehicles for a long time before reaching 100% market penetration. Considering that human driving data is variable and infinite, human driving data-driven NPC-Vs can be used to comprehensively test autonomous vehicles. Due to the above three reasons, the NPC-Vs in VST need to be driven by human driver data. However, the current natural driving data is concentrated on specific road sections and the collection cost is high, which makes it infeasible to inject a large amount of VST data.

[0003] When testing an autonomous driving system (ADS) and an autonomous driving assistance algorithm (ADAA), a comprehensive and complex variable traffic flow is required as a test background to verify the reliability of the ADS and ADAA. Therefore, the NPC-Vs used as background traffic flow generators need to have different motion patterns and conditions to generate as many traffic flow conditions as possible. In addition, AVS needs to coexist with human-driven vehicles before achieving 100% market penetration, resulting in heterogeneous traffic flows, which also requires virtual testing. However, due to the small amount of available natural driving data and the high collection cost, VST based on data injection is difficult. If different vehicle motion trajectories can be quickly generated based on partial natural driving data, the speed of ADS testing and heterogeneous traffic flow scenario testing can be improved, thereby accelerating the VST of AVS.

[0004] According to the type of generation method, the proposed models can be roughly divided into rule-based models and learning-based models. Rule-based models focus on analyzing the probabilities of vehicle actions under specific motion patterns and represent the most likely generated trajectories using mathematical formulas. However, the descriptive ability of rule-based models is limited and it is difficult to adapt to trajectory generation tasks beyond the rules. Learning-based models use a black-box learning process to extract the mapping relationship between vehicle motion states or historical trajectories and generated trajectories. Currently, quite a lot of research uses neural networks to generate future trajectories of vehicles, that is, for trajectory prediction. However, image-based models require powerful algorithm support, and powerful GPU algorithms will significantly increase the cost and energy consumption of autonomous vehicles.

[0005] There are some deficiencies in the existing technologies, which can be summarized as follows: 1. Most of the existing methods are for autonomous driving and cannot be used for NPC-V driving. 2. The existing methods are based on single-vehicle trajectories and it is difficult to generate traffic flows in batches. 3. Human-like driving can generate complex traffic flows, but lack quantitative analysis of characteristics. 4. Many methods are designed for specific applications and have low generality and reusability.

[0006] Therefore, a model that can extract features from existing natural driving data and reasonably combine different vehicle driving conditions to quickly generate trajectories is needed to solve the above technical problems. Summary of the Invention

[0007] The technical solution adopted by the present invention to solve the technical problems is: a method for generating vehicle trajectories by extracting feature combinations of driving conditions from natural driving data, including the following steps:

[0008] Step 1: Obtain equal-number coordinates and high-fidelity sampled trajectories through the generalized trajectory sampling method - integer multiple sampling algorithm;

[0009] Step 2: Conduct relevant probability statistics on different vehicle kinematic parameters to generate a reasonable combination of human-like vehicle kinematic parameters;

[0010] Step 3: Based on the learning-based batch trajectory generation model, generate different human-like vehicle trajectories based on different combinations of kinematic parameters.

[0011] Preferably, in the above Step 1, the generalized trajectory sampling method - integer multiple sampling algorithm includes the following steps:

[0012] Step 1-1: Determine a target sampling quantity, which represents the number of coordinate points to be finally retained for each trajectory;

[0013] Step 1-2: For each trajectory, calculate a sampling interval, which is used to determine how many coordinate points to select a sampling point at intervals until the sampling reaches the target sampling quantity;

[0014] Step 1-3: Select sampling points at the sampling interval. All coordinate points where the coordinate index modulo the sampling interval is zero will be selected as sampling points until the entire trajectory is traversed;

[0015] Step 1-4: Adjust the sampling results. If the number of sampling points is more than the target sampling number, remove the extra points from the beginning or end of the trajectory; if the number of sampling points is less than the target sampling number, perform interpolation in the smooth part of the trajectory to make up the difference;

[0016] Step 1-5: Determine the optimal sampling number and adjust the target sampling number to balance the data volume and accuracy.

[0017] Preferably, in Step 1-2, the calculation formula for the sampling interval is:

[0018]

[0019] In Equation (1), sampling_integer represents the sampling interval, trajectory_length represents the total number of coordinate points of the current vehicle trajectory, and sampling_times represents the number of samplings;

[0020] In Step 1-4, cubic spline interpolation method is used for interpolation to make up the difference. The expression of the cubic spline interpolation function is as follows:

[0021] S i (x) = a i +b i (x - x i ) + c i (x - x i ) 2 +d i (x - x i ) 3 (3)

[0022] In Equation (3), S i (x) represents the cubic spline curve, a i , b i , c i , d i represent the coefficients to be fitted, representing the constant term, first-order term, second-order term, and third-order term coefficients on each spline curve respectively, x ∈ [x i , x i+1 represents the interpolation interval, and x i represents the known interpolation point.

[0023] Preferably, Step 2 specifically includes the following steps:

[0024] Step 2-1: Rely on probability statistics; including:

[0025] Step 2-1-1: Divide the data provided by the data set in the form of intervals;

[0026] Step 2-1-2: Classify the original data by intervals to obtain the longitudinal speed distribution of the vehicle;

[0027] Step 2-1-3: Calculate the dependence probabilities of longitudinal speed and acceleration in different intervals to obtain the longitudinal speed - longitudinal acceleration dependence probability, longitudinal speed - lateral speed dependence probability, and longitudinal speed - lateral acceleration dependence probability;

[0028] Step 2-1-4: Set the lateral acceleration of the vehicle as the reference and obtain the dependence probability between the lateral acceleration and the heading;

[0029] Step 2-1-5: According to the dependence probabilities of all types of VKD (Vehicle Kinematic Data), form a VKD generation tree that expands upward with speed as the root node;

[0030] Step 2-2: Generation of full state parameters; including:

[0031] Step 2-2-1: Based on the VKD generation tree, obtain the longitudinal speed and longitudinal acceleration of the vehicle;

[0032] Step 2-2-2: Based on the VKD generation tree, obtain the lateral speed and lateral acceleration of the vehicle;

[0033] Step 2-2-3: Obtain the intersection behavior value;

[0034] Step 2-2-4: Obtain the maximum value, minimum value, average value, minimum speed difference, and maximum speed difference of the longitudinal speed in the data interval of the original data;

[0035] Step 2-2-5: Calculate the speed difference interval, the maximum longitudinal speed passing through the intersection, and the minimum longitudinal speed;

[0036] Step 2-2-6: Calculate the maximum and minimum values of the longitudinal acceleration, lateral speed, and lateral acceleration respectively;

[0037] Step 2-2-7: Use the data representing the vehicle steering information to represent the heading data;

[0038] Step 2-2-8: Generate a data frame for each vehicle containing the maximum value, minimum value, average value, and specific heading of the lateral speed, longitudinal speed, lateral acceleration, and longitudinal acceleration; Combine the obtained data, and each vehicle corresponds to a matrix;

[0039] Step 2-3: Generation of critical state parameters; including:

[0040] Step 2-3-1: Determine the maximum critical parameters for each of the lateral speed, longitudinal speed, lateral acceleration, and longitudinal acceleration;

[0041] Step 2-3-2: Obtain vehicle kinematic data based on the lateral critical vehicle speed and generate parameter combinations representing different vehicle behaviors; the parameter combinations include: critical longitudinal speed, critical longitudinal acceleration, and critical lateral acceleration.

[0042] Preferably, in step 3, a mapping relationship network based on an encoder-decoder structure is designed; the mapping relationship network first uses a fully connected layer to fuse input data including 15 kinds of vehicle kinematic data, and then uses an encoder-decoder structure to generate vehicle trajectory data.

[0043] More preferably, in step 3, when the neural network backpropagates, the encoder-decoder structure extracts features from the trajectory data.

[0044] Preferably, step 3 specifically includes the following steps:

[0045] Step 3-1: During the encoder-decoder training process, through a gating mechanism, the encoder dynamically selects the contributions of historical information and the current input to the hidden state, thereby capturing long-term and short-term dependencies in the time series;

[0046] Step 3-2: After the encoder processes all the input sequences, it generates a fixed-length encoded vector C representing all the features of the trajectory;

[0047] Step 3-3: Gradually generate the output sequence from the encoded vector C and input it into the decoder for calculation;

[0048] Step 3-4: Update the hidden state of the decoder;

[0049] Step 3-5: After obtaining all the hidden states, calculate the conditional probability;

[0050] Step 3-6: After obtaining the output according to the conditional probability distribution, use a time distribution layer to regularize the data to obtain the final trajectory.

[0051] More preferably, in step 3-1, the encoder is calculated as follows:

[0052]

[0053] In Equation (18), h t represents the hidden state of the encoder at the moment, h t-1 represents the hidden state at the previous moment, z represents the update gate, r represents the reset gate, h' t represents the candidate hidden state, W and U represent the training parameter matrices, σ represents the activation function, and xt represents the input vector at the current moment t;

[0054] In step 3-3, the initial hidden state of the decoder is calculated as:

[0055]

[0056] In Equation (19), h' t ' represents the hidden state of the decoder at time t, V' represents the training parameter matrix of the decoder, and c represents the final hidden state of the encoder;

[0057] In step 3-4, the hidden state update process of the decoder is:

[0058]

[0059] In Equation (20), the update gate z' and the reset gate r' control the decoder to generate a new hidden state by using the state of the previous moment and the encoding vector. W, U, c and their variants represent variable training parameter matrices;

[0060] In step 3-5, the conditional probability is:

[0061]

[0062] In Equation (21), p(y t |y t-1 ,..., y1, X) represents the conditional probability, O h , O y , O c represent the linear transformation matrices of the hidden state h”, the previous output y t , and the context vector c for the current score, which are trainable dynamic parameters. k represents the dimension of the output vocabulary, y t-1 represents the target word at the previous moment, S t-1 represents the scoring function result when the i-th word is used as the output, which is used for softmax. S i represents the output index corresponding to the maximum score among all candidates at the current moment, and g represents the normalization factor, which is used to control the smoothness of softmax. t

[0063] The beneficial effects of the present invention are:

[0064] 1. The trajectory generated by the present invention has a more complex motion pattern than the original trajectory, and can provide a wider coverage of background traffic flow for the measured ADS and ADAA.

[0065] 2. The trajectory generated by the present invention can be quickly deployed to a simulation platform, accelerating the testing process while reducing the cost of real data collection.

[0066] 3. The method of the present invention can batch generate humanoid trajectories with various motion states in a short time and is more easily transplanted into scenarios such as highways and urban expressways, providing a highly covered background traffic flow for ADS and ADAA VST in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is the structural block diagram of the sampling algorithm for the method of extracting characteristic combination driving conditions from natural driving data of the present invention to generate vehicle trajectories;

[0068] Figure 2 is the structural block diagram of the VKD generation tree of the present invention;

[0069] Figure 3 is the structural block diagram of the method for generating critical state parameters of the present invention;

[0070] Figure 4 is the structural block diagram of the learning-based batch trajectory generation model of the present invention;

[0071] Figure 5 is the network diagram for extracting the mapping relationship of the present invention;

[0072] Figure 6 is the schematic diagram of the steps of the present invention;

[0073] Figure 7 is the comparison between the original trajectory and the generated trajectory in the embodiment of the present invention;

[0074] Figure 8 is the critical trajectory generated in the embodiment of the present invention: (a) longitudinal velocity (b) longitudinal acceleration (c) lateral acceleration (d) critical trajectory at the combined orthogonality. DETAILED DESCRIPTION OF THE INVENTION

[0075] The relevant technologies in the present invention will be clearly and completely described below 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. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0076] Refer to Figures 1 to 8 As shown, the technical solution adopted in this specific embodiment includes the following steps:

[0077] 1. By using the generalized trajectory sampling method - integer multiple sampling (IM-sampling) algorithm to obtain equidistant coordinates and high-fidelity sampled trajectories.

[0078] 2. Perform relevant probability statistics on different vehicle kinematic parameters to generate a rational combination of humanoid vehicle kinematic parameters.

[0079] 3. A learning-based batch trajectory generation model is proposed to generate different humanoid vehicle trajectories based on different combinations of kinematic parameters.

[0080] Specifically as follows:

[0081] Step 1: Obtain equal-number coordinates and high-fidelity sampling trajectories through the generalized trajectory sampling method - integer multiple sampling algorithm.

[0082] 1.1 Set the sampling quantity sampling_times

[0083] First, a target sampling quantity sampling_times needs to be determined, which represents the number of coordinate points to be retained for each trajectory in the end. For example, if sampling_times = 12 is set, then each trajectory will be sampled to 12 coordinate points.

[0084] 1.2 Calculate the sampling interval sampling_integer

[0085] For each trajectory, calculate a sampling interval sampling_integer. The sampling interval is used to determine how many coordinate points to select a sampling point at intervals. The calculation formula is as follows:

[0086]

[0087] Where: trajectory_length is the total number of coordinate points of the current vehicle trajectory.

[0088] sampling_integer is the rounded sampling interval. sampling_integer means taking a point every sampling_integer coordinate points until the sampling reaches the target sampling_times quantity.

[0089] 1.3 Select sampling points according to the sampling interval

[0090] Assume that the coordinate index (i.e., the serial number of the coordinate point) of the trajectory is trajectory_index, then the selection of each sampling point follows the following conditions:

[0091] trajectory index %sampling integer =0 (2)

[0092] That is, every coordinate point where the trajectory_index modulo sampling_integer is zero will be selected as a sampling point. This condition ensures that one point is taken every sampling_integer points until the entire trajectory is traversed.

[0093] 1.4. Adjust the sampling results

[0094] After the above steps, two situations may occur:

[0095] The number of sampling points is more than sampling_times: If the number of coordinate points included in a certain trajectory after sampling is more than the target number sampling_times, then remove the extra points from the beginning or end of the trajectory. Usually, start removing from the third point of the trajectory until the number of remaining coordinate points is the same as sampling_times. The reason for this treatment is that the two ends of the trajectory are usually smoother and have less impact on the entire trajectory.

[0096] The number of sampling points is less than sampling_times: If there is a lack of points, interpolation can also be performed in the smooth part of the trajectory to make up for it. Here, the cubic spline interpolation method is used for the interpolation method. The expression of the cubic spline interpolation function is as follows:

[0097] S i (x) = a i +b i (x - x i ) + c i (x - x i ) 2 +d i (x - x i ) 3 (3)

[0098] 1.5. Determine the optimal sampling number sampling_times

[0099] To select the optimal sampling_times, different sampling_times values can be tried and the matching degree between the sampling results and the original trajectory can be observed. Generally speaking, the larger the sampling_times is set, the closer the sampled trajectory is to the original trajectory. However, when sampling_times is less than a certain value (for example, less than 12), the sampled trajectory will be distorted and cannot represent the original trajectory well. Therefore, in the research, sampling_times = 12 is set to achieve a balance between the data volume and the accuracy.

[0100] Step 2: Conduct relevant probability statistics on different vehicle kinematic parameters to generate a rational combination of humanoid vehicle kinematic parameters - a probability-limited parameter generation method. In addition to generating the VKD in the full state, this method also generates key parameters of vehicle motion to expand the scope of autonomous driving tests.

[0101] 2.1. Dependence on probability statistics

[0102] 2.1.1. The inD dataset was collected by RWTH Aachen University in Germany. UAVs were used to record the trajectories of real traffic participants at multiple signal-free intersections, covering vehicles, pedestrians, bicycles, etc. It has high precision, high frequency, and rich interaction characteristics, and is widely used in trajectory prediction and autonomous driving research.

[0103] Divide the five types of data provided by the inD dataset (lateral and longitudinal speed and acceleration, and heading types that can reflect vehicle motion states) in the form of intervals. The specific division method is as follows: Each intersection has four road junctions, denoted by [A], [B], [C], and [D] respectively. Combining the three motion states of vehicle going straight, turning left, and turning right, the heading data can be divided into 12 intervals, and the specific division is as follows:

[0104] Going straight: [A]-[C], [C]-[A], [B]-[D], [D]-[B];

[0105] Turning left: [A]-[B], [B]-[C], [C]-[D], [D]-[A];

[0106] Turning right: [A]-[D], [D]-[C], [C]-[B], [B]-[A].

[0107] 2.1.2. Classify the original data into intervals to obtain the longitudinal vehicle speed distribution.

[0108] In this step, a rule for dividing vehicle speed intervals is proposed: First, divide the speed intervals into [-2,0] m / s, [0,5] m / s, [5,10] m / s, [10,15] m / s, [15,20] m / s,

[0109] [20,25] m / s; Second, match all the data with the divided intervals. If the data only exists in one of these intervals, then this data is classified into the corresponding interval; if the data spans two of these speed intervals, then the data is attributed to the interval with the larger speed of the two intervals; if the data spans three or five speed intervals, then the data is classified into the middle interval; if the data spans four speed intervals, then it is classified into the interval with the larger middle value; if the data spans all speed intervals, then it is classified into the interval of [15,20] m / s.

[0110] Through this speed division rule, it can be obtained that the longitudinal vehicle speed at the intersection follows a normal distribution, and the speeds of most vehicles are in the range of [10, 15] m / s, that is, 36 km / h - 54 km / h, which is consistent with the daily driving situation and reflects the correctness of the interval classification carried out.

[0111] 2.1.3. Calculate the dependence probabilities of longitudinal speeds and accelerations in different intervals to obtain the dependence probabilities of longitudinal speed - longitudinal acceleration, longitudinal speed - lateral speed, and longitudinal speed - lateral acceleration.

[0112] Through this process, it can be concluded that the longitudinal acceleration is concentrated in the intervals of [0, 2] m / s 2 and [-2, 0] m / s 2 The overall acceleration shows a normal distribution. The average longitudinal acceleration of the vehicle does not exceed 1.5 m / s 2 value and shows a normal distribution throughout the interval. The lateral acceleration is mainly in the interval of [-0.2, 0] m / s 2 and the vehicle is laterally stable in this interval, which conforms to normal driving behavior.

[0113] 2.1.4. To avoid generating data that violates vehicle driving rules, such as turning the vehicle when the lateral acceleration is zero, set the lateral acceleration of the vehicle as a reference and obtain its dependence probability with the heading.

[0114] 2.1.5. According to the dependence probabilities of all types of VKD calculated previously, form a VKD generation tree that expands upward with speed as the root node.

[0115] 2.2. Full state parameter generation

[0116] 2.2.1. Based on the VKD generation tree, obtain the longitudinal speed and longitudinal acceleration of the vehicle.

[0117] In this step, we assume that the 6 intervals of the longitudinal speed are {A - B}, and the specific interval is [A - B]. Then the calculation formula for the longitudinal speed is:

[0118] long vel =P [a-b] *{A - B}→P [random] *[α - b] (4)

[0119] In the formula, P [a-b] represents the probability of selecting the interval [a - b], and * represents a value taken from a certain range. Therefore, P [a-b] *{A - B} means that the interval [A - B] is selected from all intervals {A - B} with a probability of P [a-b] Similarly, P [random]*[a - b] represents the longitudinal speed obtained with a random probability P [random] from the interval [a - b].

[0120] After obtaining the longitudinal speed, the longitudinal acceleration is as follows:

[0121] long acc = P [a-b] *{A - B} → P [c-d] *{C - D} → P′ [random] *[α - b] (5)

[0122] where [c - d] is a specific interval among all four intervals {c - d} of the longitudinal acceleration, and P' [random] is the new random probability.

[0123] 2.2.2. Using the same method as in 2.2.1, the lateral speed and lateral acceleration of the vehicle can be calculated and given by the following formulas:

[0124] lat vel = P [a-b] *{A - B} → P [e-f] *{E - F} → P″ [random] *[e - f] (6)

[0125] lat acc = P [a-b] *{A - B} → P [g-h] *{G - H} → P[P″′ [random] *[g - h] (7)

[0126] where [e - f], [e - f], [g - h], and {g - h} represent the specific intervals and all intervals of the lateral speed and acceleration respectively; P″ [random] and P″' [random] are the probabilities of randomly selecting values within specific intervals.

[0127] 2.2.3. Since the heading depends on the lateral acceleration, the values of typical intersection behavior can be given by the following formula:

[0128] [X] - [X] = P [a-b] *{A - B} → P [g-h] *{G - H} → P [X]-[X] *{[X] - [X]} (8)

[0129] where [X] - [X] represents the vehicle information entering and leaving the intersection, and {[X] - [X]} contains 12 behaviors of the vehicles at the intersection; P [X]-[X] in the equation is the probability of entering and leaving the intersection.

[0130] 2.2.4. The maximum, minimum, and average values of the longitudinal speed in the data range [a - b] in the original data are respectively:

[0131]

[0132] Among them, n represents the nth data value [a - b] in the longitudinal speed range. The minimum and maximum speed differences in the original data interval [a - b] are respectively:

[0133]

[0134]

[0135] Among them represents the maximum value from the first to the nth data value.

[0136] 2.2.5. Calculate the maximum and minimum longitudinal speeds in the speed difference range and passing through the intersection.

[0137] The speed difference range is defined by the following formula:

[0138]

[0139] Based on this speed difference range formula and the longitudinal speed given in step 2.2.1, the maximum and minimum longitudinal speeds of the vehicles passing through the intersection are calculated respectively by the following formulas:

[0140]

[0141] Among them, represents the maximum value corresponding to the nth longitudinal speed value long in the range [a - b] vel corresponding, represents the range in the random value. [[ID=Heading]]

[0142] 2.2.6. Calculate the maximum and minimum values of longitudinal acceleration, lateral speed, and acceleration in the same way as in steps 2.2.4 and 2.2.5.

[0143] 2.2.7. Represent the heading data with the vehicle steering information data. Since the maximum and minimum values of the heading are stable, the present invention only fine-tunes the heading values within the range of 0 - 2 degrees.

[0144] Generate three types of turning information data as follows:

[0145] <Turning_indicator, Entrance_heading, Exit_heading> (17)

[0146] The steering indicator values come from the interval [1 - 12], where the three groups of values [1, 2, 3, 4], [5, 6, 7, 8], and [9, 10, 11, 12] represent the straight, right - turn, and left - turn behaviors of four different routes respectively.

[0147] 2.2.8. Each vehicle generates a data frame containing the maximum, minimum, average, and specific heading of the lateral / longitudinal speed / acceleration. Combine the obtained data to get a [10000×15] matrix.

[0148] 2.3. Generation of critical state parameters

[0149] 2.3.1. In the present invention, four types of kinematic parameters (lateral speed / acceleration, longitudinal speed / acceleration) are selected. First, four maximum critical parameters for each of the four kinematic parameters are determined.

[0150] 2.3.2. According to Figure 3 the critical state parameters, vehicle kinematic data is obtained based on the critical lateral vehicle speed. The following three parameter combinations representing different vehicle behaviors are generated:

[0151] Critical longitudinal speed: The vehicle passes through the intersection at a high speed of [20 - 25] m / s and with a sharp acceleration, concentrating on straight and left - turn behaviors;

[0152] Critical longitudinal acceleration: The vehicle passes through the intersection with a high longitudinal acceleration of [2 - 5] m / s² and a sharp acceleration;

[0153] Critical lateral acceleration: The vehicle passes through an intersection with a lateral acceleration of [-6 - 2] m / s² or [2 - 5] m / s², showing a sharp - turn phenomenon, concentrating on left - turn and right - turn behaviors.

[0154] In this step, a total of 1500 (500×3) key data combinations are generated, forming a [1500×15] matrix, which is input into the trajectory generation model together with the [10,000×15] matrix generated in the full - state mode.

[0155] Step 3: Based on the learning - based batch trajectory generation model (the network structure is as Figure 4 shown), different human - like vehicle trajectories are generated based on different combinations of kinematic parameters. The present invention proposes a mapping relationship network based on an encoder - decoder structure. The specific mapping relationship extraction network is as Figure 5 shown. The mapping relationship extraction network first uses a fully - connected layer to fuse the input data including 15 kinds of vehicle kinematic data, and then uses an encoder - decoder structure to generate vehicle trajectory data. In particular, when the neural network back - propagates, the encoder - decoder structure can well extract features from the trajectory data.

[0156] 3.1. During the encoder-decoder training process, the auxiliary label or context information y corresponding to the input x participates in the loss calculation and node operations, unlike a general Recurrent Neural Network (RNN) which is only used for loss supervision. Assume the encoder input is represented by {x n , y n}, then the encoder calculates according to the following process:

[0157]

[0158] h t is the hidden state of the encoder at time t, h t-1 is the hidden state of the previous time step, z is the update gate, which controls how much of the current time step's hidden state comes from the previous time step. h' t is the candidate hidden state, W, U are training parameter matrices, and σ is the activation function.

[0159] The update gate z controls how the current hidden state switches between the previous state and the candidate state. The reset gate r controls how much historical information is retained in the candidate hidden state. Through these gating mechanisms, the encoder can dynamically select the contributions of historical information and the current input to the hidden state, thus capturing long-term and short-term dependencies in the time series.

[0160] 3.2. After the encoder processes the entire input sequence through the calculation in 3.1, it generates a fixed-length encoding vector C, representing all the features of the trajectory.

[0161] 3.3. After the calculations in 3.1 and 3.2, the encoding vector C generated by the encoder is gradually used to generate the output sequence and input into the decoder for calculation. The initial hidden state of the decoder is calculated as follows:

[0162] h″ t = tanh(V′·c) (19)

[0163] Here V' is the training parameter matrix of the decoder, and c is the final hidden state of the encoder.

[0164] 3.4. The hidden state update process of the decoder is as follows:

[0165]

[0166] h' t ' is the hidden state of the decoder at time t. The update gate z' and the reset gate r' control how the decoder uses the previous state and the encoding vector to generate a new hidden state. Among them, W, U, c and their variants represent variable training parameters.

[0167] 3.5. After obtaining all hidden states through the previous four steps, calculate the conditional probability using the following formula:

[0168]

[0169] where O h 、O y 、O c are dynamic training parameters, and k is the output dimension.

[0170] 3.6. After obtaining the output according to the probability distribution of the decoder in 3.5, use the time distribution layer to regularize the data to obtain the final trajectory.

[0171] Embodiment

[0172] The IM-Sampling sampling algorithm proposed in this embodiment aims to unify the original natural driving trajectories into a fixed length, ensuring the consistency of data input and the integrity of the trajectory structure. Experimental results show that this method exhibits good adaptability and fidelity on different types of trajectories, providing a reliable data basis for subsequent model training.

[0173] In terms of input construction, this embodiment introduces a kinematic parameter combination method based on probability constraints, constructing a generation logic by mining the statistical dependence relationships between various parameters, effectively avoiding the emergence of unreasonable states. Experiments verify that this method not only ensures physical rationality but also improves the diversity and representativeness of samples.

[0174] In terms of model design, this embodiment constructs a trajectory generation model based on the Encoder-Decoder framework, realizing batch trajectory synthesis under multiple motion states. Experimental results show that this model can generate results with natural trajectory shapes and diverse behaviors, and still maintains high stability and expressiveness under key driving states, demonstrating the generalization ability and practicality of the method.

[0175] In addition, in this embodiment, a humanized analysis of the generation results was carried out using a variety of trajectory similarity metrics. The results show that the generated trajectories are highly consistent with the real trajectories in terms of spatial structure, temporal evolution, and behavioral boundaries, demonstrating that the proposed method has good ability to retain human driving styles, the generated quality is reliable, and it can be used to support highly credible autonomous driving simulation tests. Table 1 summarizes the scores of the generated trajectories and the real trajectories under each metric, showing that the generated results are highly consistent with the real samples in terms of spatial structure, temporal synchronization, and behavioral boundary characteristics. In addition, the similarity characteristics of the critical state are slightly greater than those of the full state, indicating that the trajectories of the critical state are more complex than those of the full state. The following 10 metrics were used in this embodiment: PartialCurve Mapping (PCM), Dynamic Time Warping (DTW), Curve Length (CL), Discrete Fréchet Distance (DF), Area Between Two Curves (ABTC), Longitudinal VelocitySimilarity, Lateral Velocity Similarity, Longitudinal Acceleration Similarity, Lateral Acceleration Similarity, Yaw Rate Similarity. The first five are commonly used trajectory shape similarity metrics, which evaluate the overall matching degree of the trajectories from dimensions such as spatial distance, path length, and temporal variation; the last five are kinematic feature metrics, which are used to reflect the similarity of the generated trajectories in terms of behavioral attributes such as speed and acceleration, and comprehensively reflect the "human style" and naturalness of the trajectories. The similarity metrics of different behaviors of the generated trajectories in this embodiment are shown in Table 1:

[0176] Table 1

[0177]

[0178] In summary, the trajectories generated by the present invention have more complex motion patterns than the original trajectories, and can provide a wider coverage of background traffic flow for the measured ADS and ADAA; the trajectories generated by the present invention can be quickly deployed to the simulation platform, accelerating the test process while reducing the cost of real data collection; the method of the present invention can batch generate human-like trajectories with multiple motion states in a short time, and is more easily transplanted into scenarios such as highways and urban expressways, providing a highly covered background traffic flow for ADS and ADAA VST under different scenarios.

[0179] 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 method for generating a vehicle trajectory by combining driving conditions with features extracted from natural driving data, characterized in that, Including the following steps: Step 1: Obtain equal-number coordinates and high-fidelity sampling trajectories through the generalized trajectory sampling method - integer multiple sampling algorithm; Step 2: Conduct relevant probability statistics on different vehicle kinematic parameters to generate a reasonable combination of humanoid vehicle kinematic parameters; Step 3: Based on the learning-based batch trajectory generation model, generate different humanoid vehicle trajectories based on different combinations of kinematic parameters.

2. The method for generating a vehicle trajectory by extracting a feature combination driving condition from natural driving data according to claim 1, wherein In the said Step 1, the generalized trajectory sampling method - integer multiple sampling algorithm includes the following steps: Step 1-1: Determine a target sampling quantity, which represents the number of coordinate points to be finally retained for each trajectory; Step 1-2: For each trajectory, calculate a sampling interval, which is used to determine how many coordinate points to select a sampling point at intervals until the sampling reaches the target sampling quantity; Step 1-3: Select sampling points according to the sampling interval. Coordinate points whose coordinate indices modulo the sampling interval are zero will be selected as sampling points until the entire trajectory is traversed; Step 1-4: Adjust the sampling result. If the number of sampling points is more than the target sampling quantity, remove the redundant points from the beginning or end of the trajectory; if the number of sampling points is less than the target sampling quantity, perform interpolation in the smooth part of the trajectory to make up the difference; Step 1-5: Determine the optimal sampling quantity and adjust the target sampling quantity to achieve a balance between the data volume 3. The method for generating a vehicle trajectory by extracting characteristic combined driving conditions from natural driving data according to claim 2, wherein, and accuracy. In the said Step 1-2, the calculation formula of the sampling interval is: In formula (1), sampling_integer represents the sampling interval, trajectory_length represents the total number of coordinate points of the current vehicle trajectory, and sampling_times represents the sampling quantity; S i f(x) = a i + b i (x - x i ) + c i (x - x i ) 2 + d i (x - x i ) 3 (3) In formula (3), S i (x) represents a cubic spline curve, and a i , b i , c i , d i represent the coefficients of the constant term, the first-order term, the second-order term, and the third-order term on each segment of the spline curve respectively. x ∈ [x i , x i+1 represents the interpolation interval, and x i represents the known interpolation point.

4. The method for generating a vehicle trajectory by extracting a feature combination driving condition from natural driving data according to claim 1, wherein In the said Step 1-4, the interpolation to make up the difference uses the cubic spline interpolation method; the expression of the cubic spline interpolation function is as follows: The said Step 2 specifically includes the following steps: Step 2-1: Dependence probability statistics; including: Step 2-1-1: Divide the data provided by the data set in the form of intervals; Step 2-1-2: Classify the original data by intervals to obtain the longitudinal speed distribution of the vehicle; Step 2-1-3: Calculate the dependence probabilities of longitudinal speed and acceleration in different intervals to obtain the longitudinal speed - longitudinal acceleration dependence probability, longitudinal speed - lateral speed dependence probability, and longitudinal speed - lateral acceleration dependence probability; Step 2-1-4: Set the lateral acceleration of the vehicle as the reference and obtain the dependence probability between the lateral acceleration and the heading; Step 2-1-5: According to the dependence probabilities of all types of VKD, form a VKD generation tree with speed as the root node and expand upward; Step 2-2: Full state parameter generation; including: Step 2-2-1: Based on the VKD generation tree, obtain the longitudinal speed and longitudinal acceleration of the vehicle; Step 2-2-2: Based on the VKD generation tree, obtain the lateral speed and lateral acceleration of the vehicle; Step 2-2-3: Obtain the intersection behavior value; Step 2-2-4: Obtain the maximum value, minimum value, average value, minimum speed difference, and maximum speed difference of the longitudinal speed in the data interval of the original data; Step 2-2-5: Calculate the speed difference interval, the maximum longitudinal speed, and the minimum longitudinal speed for passing through the intersection; Step 2-2-6: Calculate the maximum and minimum values of the longitudinal acceleration, lateral speed, and lateral acceleration respectively; Step 2-2-7: Represent the heading data using the data of the vehicle steering information; Step 2-2-8: Each vehicle generates a data frame containing the maximum, minimum, average values of the lateral speed, longitudinal speed, lateral acceleration, and longitudinal acceleration, and the specific heading; Combine the obtained data, and each vehicle corresponds to a matrix; Step 2-3: Generate critical state parameters; including: Step 2-3-1: Determine the maximum critical parameter of each of the lateral speed, longitudinal speed, lateral acceleration, and longitudinal acceleration; Step 2-3-2: Obtain vehicle kinematic data based on the lateral critical vehicle speed, and generate a parameter combination representing different vehicle behaviors; The parameter combination includes: critical longitudinal speed, critical longitudinal acceleration, and critical lateral acceleration.

5. The method for generating a vehicle trajectory by extracting a feature combination driving condition from natural driving data according to claim 1, characterized in that In the said Step 3, design a mapping relationship network based on the encoder-decoder structure; The mapping relationship network first uses a fully connected layer to fuse the input data including 15 kinds of vehicle kinematic data, and then uses the encoder-decoder structure to generate vehicle trajectory data.

6. The method for generating a vehicle trajectory by extracting feature combination driving conditions from natural driving data according to claim 5, wherein In the said Step 3, when the neural network backpropagates, the encoder-decoder structure extracts features from the trajectory data.

7. The method for generating a vehicle trajectory by extracting feature combination driving conditions from natural driving data according to claim 5, wherein The said Step 3 specifically includes the following steps: Step 3-1: During the encoder-decoder training process, through the gating mechanism, the encoder dynamically selects the contributions of the historical information and the current input to the hidden state, so as to capture the long-term and short-term dependencies in the time series; Step 3-2: After the encoder processes all the input sequences, generate a fixed-length encoded vector C, representing all the features of the trajectory; Step 3-3: Gradually generate the output sequence from the encoded vector C and input it into the decoder for calculation; Step 3-4: Update the hidden state of the decoder; Step 3-5: After obtaining all the hidden states, calculate the conditional probability; Step 3-6: After obtaining the output according to the conditional probability distribution, use the time distribution layer to regularize the data to obtain the final trajectory.

8. The method for generating a vehicle trajectory by extracting a feature combination driving condition from natural driving data according to claim 7, wherein In the said Step 3-1, the encoder calculates according to the following process: In Equation (18), h t represents the hidden state of the encoder at the moment, h t-1 represents the hidden state of the previous moment, z represents the update gate, r represents the reset gate, h' t represents the candidate hidden state, W and U represent the training parameter matrices, σ represents the activation function, x t represents the input vector at the current moment t; In the said Step 3-3, the initial hidden state of the decoder is calculated as: h″ t = tanh(V′·c) (19) In Equation (19), h″ t represents the hidden state of the decoder at time t, V' represents the training parameter matrix of the decoder, and c represents the final hidden state of the encoder; In the said Step 3-4, the process of updating the hidden state of the decoder is: In Equation (20), W, U, c and their variants represent variable training parameter matrices; In the said Step 3-5, the conditional probability is: In Equation (21), p(y t |y t-1 ,..., y1, X) represents the conditional probability, O h , O y , O c represent the linear transformation matrices of the hidden state h″ t , the previous output y t-1 , and the context vector c for the current score, which are trainable dynamic parameters, k represents the dimension of the output vocabulary, y t-1 represents the target word at the previous moment, s i represents the result of the scoring function when the i-th word is used as the output, s t represents the output index corresponding to the maximum score among all candidates at the current moment, and g represents the normalization factor.

Images