Interactive trajectory prediction method and system considering driving style uncertainty

Abstract

The invention discloses an interactive trajectory prediction method and system considering driving style uncertainty, and relates to the technical field of automatic driving. The method comprises the following steps: receiving a vehicle state information data set obtained based on a typical road, carrying out data preprocessing, and dividing the preprocessed data into a training set and a test set; based on the preprocessed data, the driving condition is optimized, and the vehicle driving style is identified by using a Gaussian hybrid clustering method to obtain driving style information; and a trajectory prediction model is constructed, the trajectory prediction model comprises an encoder part and a decoder part, and a lane attention mechanism and a multi-head attention mechanism are introduced into the encoder part. The method can accurately capture and predict the complex interaction behavior between the automatic driving vehicle and the human driving vehicle in the face of diversified driving styles and complex traffic states, and by combining the uncertainty of the driving styles, the model can better understand the personalized behavior mode of the driver, and the trajectory prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to an interactive trajectory prediction method and system that considers driving style uncertainty. Background Art

[0002] Autonomous driving technology has made significant progress in the past decade, and the market share of conditionally autonomous vehicles has steadily increased. However, due to problems such as imperfect road infrastructure, incomplete regulatory standards, and low driver trust and acceptance of new technologies, the large-scale market application of autonomous vehicles is still a relatively slow process. It is foreseeable that the mixed driving environment where human-driven vehicles and autonomous vehicles coexist will become the main form of road traffic for a long time in the future.

[0003] Trajectory prediction plays a vital role in improving driving safety, comfort, road and energy efficiency. Studies have shown that autonomous driving systems with integrated trajectory prediction can reduce collision rates and traffic violation rates to 1.80% and 1.24%, respectively, while controlling acceleration jitter within 0.15m / s3. Autonomous vehicles need to predict the motion trajectories of surrounding vehicles like human drivers to accurately understand long-term future traffic conditions. However, complex mixed driving environments increase the challenges of trajectory prediction: (1) uncertainty in the behavior of surrounding vehicles, such as random lane changes and excessive deceleration; (2) diversity in vehicle behavior caused by different driving styles; (3) significant differences in behavior patterns between different models (such as cars and trucks); (4) asymmetric interactive information due to different perception capabilities of autonomous vehicles and surrounding vehicles; and (5) non-humanoid autonomous driving behavior increases the sense of insecurity of human drivers.

[0004] Despite significant progress in vehicle trajectory prediction, existing research still has two major limitations. First, most existing methods only focus on the direct interaction between the autonomous vehicle and the target vehicle, ignoring the extensive interaction between the target vehicle and its surrounding vehicles. This limited perspective reduces the model's ability to capture the complex interactive information of mixed driving environments, thereby affecting the accuracy of trajectory prediction. Second, most studies ignore the impact of the uncertainty of personalized driving preferences, which is crucial to improving trajectory prediction performance and achieving human-like driving capabilities. Summary of the invention

[0005] The purpose of the present invention is to provide an interactive trajectory prediction method and system that takes into account the uncertainty of driving style, thereby solving the problem of low trajectory prediction accuracy caused by the uncertainty of the driving preference of the vehicle itself and the behavior of surrounding vehicles.

[0006] To achieve the above object, the present invention provides the following technical solution: an interactive trajectory prediction method considering driving style uncertainty, comprising the following steps:

[0007] Receiving a vehicle status information data set obtained based on a typical road for data preprocessing, and dividing the preprocessed data into a training set and a test set;

[0008] Based on the preprocessed data, the driving conditions are optimized, and the vehicle driving style is identified using the Gaussian mixture clustering method to obtain driving style information;

[0009] Constructing a trajectory prediction model, wherein the trajectory prediction model includes an encoder part and a decoder part, and the encoder part introduces a lane attention mechanism and a multi-head attention mechanism;

[0010] The training set is input into the trajectory prediction model, and the trajectory prediction model is trained based on the loss function to obtain a trained trajectory prediction model;

[0011] The test set is input into the trained trajectory prediction model to complete the testing and performance evaluation of the trajectory prediction model, and the trajectory prediction model is used to realize vehicle trajectory prediction.

[0012] Furthermore, the typical roads are selected from the highway highD dataset, and the selection rules of typical roads are as follows:

[0013] (21) The amount of road data is large and covers a variety of driving behaviors;

[0014] (22) The road features a variety of features, with six lanes in both directions;

[0015] (23) Vehicle motion characteristics are universal.

[0016] Furthermore, the vehicle status information data set includes vehicle number, frame time, global time, relative X coordinate, relative Y coordinate, vehicle length and vehicle model data in the data set.

[0017] Furthermore, the vehicle status information data set is preprocessed as follows:

[0018] (41) Exclude vehicles with an average speed lower than 15 m / s;

[0019] (42) For zero values ​​in abnormal data, the last observed non-zero value is used for replacement;

[0020] (43) Vehicles whose total trajectory duration is less than 8 seconds are excluded.

[0021] Furthermore, based on the preprocessed data, the driving conditions are optimized, and the vehicle driving style is identified using the Gaussian mixture clustering method to obtain driving style information, as follows:

[0022] (51) Optimize the driving conditions. The specific process is as follows:

[0023] Vehicles that have not changed lanes are extracted, and the driving conditions are divided according to the mixed strategy consisting of headway and collision time, which is expressed as:

[0024] d t =max{t hc v e +t cc (v e -v p ),d0} (1)

[0025] Where, d t Indicates the distance between the two vehicles that triggers the switching condition; t hc It represents the headway coefficient, and its value is 5s; t cc represents the collision time coefficient, its value is 1s; d0 is the distance constant, its value is 100m; v e is the vehicle speed, v p is the speed of the preceding vehicle;

[0026] (52) The Gaussian mixture clustering method is used to identify the vehicle driving style. The specific process is as follows:

[0027] (52.1) Select characteristic indicators, which include: vehicle speed standard deviation, maximum acceleration, minimum acceleration, average acceleration and acceleration standard deviation, maximum headway distance, minimum headway distance, average headway distance and headway distance standard deviation, maximum headway time distance, minimum headway time distance, average headway time distance, headway time distance standard deviation, maximum acceleration derivative, minimum acceleration derivative, average acceleration derivative and acceleration derivative standard deviation;

[0028] (52.2) Based on the selected feature indicators, the driving styles of the vehicles are clustered using the Gaussian mixture clustering method:

[0029]

[0030]

[0031] Where K represents the number of Gaussian components, π n represents the weight of the nth Gaussian component, Θ represents the parameter set of Gaussian mixture clustering, which can be expressed as Θ = {π m ,μ n ,Σ n}, p(χ m |Θ) represents the sample vector χ m The probability density function, N(χ m |μ n ,Σ n ) indicates that the mean vector is μ m And the covariance matrix is ​​Σ n Gaussian distribution, d represents the dimension of the sample vector.

[0032] Furthermore, a trajectory prediction model is constructed as follows:

[0033] (61)The historical trajectory can be expressed as:

[0034]

[0035] Where H represents the historical trajectory, which contains the vehicle information within a time window, t h represents the number of time points in the historical trajectory, represents the trajectory information of the heterogeneous vehicle group at the target time t, i = 1, 2, ..., 9; represents the trajectory information of the i-th vehicle at time t, which can be expressed as:

[0036]

[0037] In the formula, They represent the longitudinal position and lateral position of the i-th vehicle at time t respectively; denote the longitudinal and lateral velocities, respectively; Represents longitudinal and lateral acceleration respectively; T i represents the type of the i-th vehicle; S i represents the driving style of the i-th vehicle;

[0038] (62) Select historical trajectories as the input of the model:

[0039]

[0040] Where X is the input vector of the model, Indicates the relative motion information of surrounding vehicles relative to the target vehicle. The value of i is a natural number from 1 to 9 except 5. t=[t,t-1,…,tt h +1] is the time step, is the historical movement information of the target vehicle;

[0041] (63) The output of the model is the future trajectory coordinates of the target vehicle, expressed as:

[0042]

[0043] Where Y is the output vector, t f is the number of time points of the predicted trajectory, t j =1,2,...,t f ,

[0044]

[0045] Furthermore, the lane attention mechanism is used to enhance the vehicle's spatial perception of the target lane, and the specific implementation method is as follows:

[0046] (71) In a heterogeneous vehicle group, three vehicles in the same lane are grouped into one group, with a total of three groups, each of which is processed independently;

[0047] First, the encoder processes the vehicle information of each group through the long short-term memory network to generate a three-dimensional lane context vector, which is and Represent the left lane, middle lane, and right lane respectively; these vectors contain information at all time points in the historical trajectory;

[0048] (72) According to these three lane context vectors and Y (t-1) To determine and The weight of (t-1) is the predicted position of the target vehicle at the previous moment, which is a two-dimensional vector;

[0049] On this basis, the weighted lane context vector is obtained and They correspond to the left lane, middle lane, and right lane respectively, and are all three-dimensional vectors, where t ranges from 1 to t f ;

[0050] (73)The total context vector can be expressed as:

[0051]

[0052] In the formula, Φ represents the linear layer for dimension compression and information fusion, cat is used to connect data, and C t Represents the total context vector.

[0053] Furthermore, the multi-head attention mechanism is used to extract high-dimensional global interaction information, and the specific implementation method is as follows:

[0054] (81) The input information of the multi-head attention mechanism is divided into two types: target vehicle data, represented by X tar , and surrounding vehicle data, represented by X sur ;

[0055] (82)X tar and Xsur It is processed through three different linear layers, namely Φ qj , Φ kj and Φ vj ,These layers play a vital role in extracting key features and enhancing ,representation capabilities;

[0056] The "query" (Q), "key" (K), and "value" (V) can be expressed as:

[0057] Q j =Φ qj (X tar ) (9)

[0058] K j =Φ kj (X sur ) (10)

[0059] V j =Φ vj (X sur ) (11)

[0060] In the formula, Φ qj , Φ kj and Φ vj There are three different linear layers for dimension compression and information fusion; Q j Carrying the information of the target vehicle, and K j and V j Carrying data of surrounding vehicles;

[0061] (83) The output of the multi-head attention mechanism is obtained through the softmax operation:

[0062]

[0063] Where, d k It's Q j and K j The column dimension, Z j is the weighted information of the jth head, which indicates the attention weight of surrounding vehicles to the target vehicle in a heterogeneous vehicle group;

[0064] Finally, the data extracted from all heads are concatenated, further processed by dimension compression and information fusion, and finally the result of the multi-head attention model is output:

[0065] Z=Φ(cat(Z1,...,Z j )) (13)

[0066] Where j=1,2,...m represents the jth head, m is the number of heads, and Z represents the output of the multi-head attention model, which encapsulates the high-dimensional interaction information between the surrounding vehicles and the target vehicle; then, the output of the multi-head attention model will be fused with the output of the lane attention model.

[0067] Furthermore, the loss function is the root mean square error, which is as follows:

[0068] After obtaining the trained vehicle trajectory prediction model, the weights of the LSTM encoder and LSTM decoder in the model are fixed, and the loss function during training is as follows:

[0069]

[0070] In the formula, RMSE represents the root mean square error, x i Represents the true value of the longitudinal displacement of the trajectory, y i represents the true value of the lateral displacement of the trajectory, represents the predicted longitudinal displacement of the trajectory, Represents the predicted value of the lateral displacement of the trajectory; N represents the number of predicted points.

[0071] According to a second aspect of the present invention, the present invention provides an interactive trajectory prediction system considering driving style uncertainty, which is used to implement the above-mentioned interactive trajectory prediction method considering driving style uncertainty, comprising:

[0072] A preprocessing module, used for receiving a vehicle status information data set obtained based on a typical road, performing data preprocessing, and dividing the preprocessed data into a training set and a test set;

[0073] An identification module is used to optimize the driving conditions based on the preprocessed data and identify the vehicle driving style using a Gaussian mixture clustering method to obtain driving style information;

[0074] A construction module is used to construct a trajectory prediction model, wherein the trajectory prediction model includes an encoder part and a decoder part, and the encoder part introduces a lane attention mechanism and a multi-head attention mechanism;

[0075] A training module, used for inputting the training set into the trajectory prediction model, and training the trajectory prediction model based on the loss function to obtain a trained trajectory prediction model;

[0076] The test output module is used to input the test set into the trained trajectory prediction model, complete the testing and performance evaluation of the trajectory prediction model, and use the trajectory prediction model to realize vehicle trajectory prediction.

[0077] The present invention has at least the following beneficial effects:

[0078] 1. Accurate prediction for complex mixed driving environments: The present invention can accurately capture and predict the complex interactive behaviors between autonomous vehicles and human-driven vehicles in the face of diverse driving styles and complex traffic conditions. By combining driving style uncertainty, the model better understands the driver's personalized behavior pattern, thereby improving the accuracy of trajectory prediction.

[0079] 2. Integrated multiple attention mechanisms: The present invention enhances the performance of the model in multiple lane and vehicle interaction environments by introducing lane attention mechanism and multi-head attention mechanism. The lane attention mechanism can dynamically adjust the weights of different lanes and focus on important lane information, while the multi-head attention mechanism further enhances the learning ability of complex interactive information and can effectively respond to irregular changes in driving behavior, such as sudden lane changes or emergency braking, thereby improving the model's ability to process global and local information, thereby enhancing the robustness of trajectory prediction.

[0080] 3. Integration of driving style recognition and behavior analysis: The present invention uses a Gaussian mixture clustering method to accurately identify different driving styles, and combines vehicle motion status, vehicle model information and driving style characteristics to achieve a deep understanding of the target vehicle and the behavior of surrounding vehicles. Through the analysis of multiple driving styles, it can better respond to personalized driving behaviors and further improve the safety and accuracy of prediction results, especially in highly interactive environments such as highways, showing obvious advantages.

[0081] 4. Dynamically evaluate the uncertainty of trajectory prediction: This invention builds a trajectory prediction uncertainty evaluation model to comprehensively evaluate the reliability of the prediction results. In a complex and uncertain mixed driving environment, the safety of the prediction results is crucial. The model can quantify and evaluate the uncertainty of trajectory prediction, ensuring that the decision-making of autonomous driving vehicles is safer and more reliable, reducing the risk of collisions and traffic accidents.

[0082] 5. Improve the adaptability and safety of the autonomous driving system: By optimizing the interaction modeling between the target vehicle and surrounding vehicles, the present invention can not only predict the trajectory more accurately, but also dynamically adjust the vehicle behavior to cope with changes in the environment, especially in mixed driving environments. The trajectory prediction method of the present invention helps autonomous driving vehicles to better understand and adapt to the behavior of surrounding vehicles, thereby improving overall traffic safety and reducing potential dangers when interacting with human drivers.

[0083] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is a schematic diagram of the flow chart of the prediction method of the present invention;

[0085] Figure 2 It is a schematic diagram of the operating principle of the prediction method of the present invention;

[0086] Figure 3 The behavior characteristic diagram of the prediction method of the present invention under the following and cruising conditions, wherein (a) is the standard deviation of the longitudinal speed, (b) is the standard deviation of the longitudinal acceleration, and (c) is the average value of the head distance;

[0087] Figure 4 This is a trajectory prediction effect diagram of the prediction method described in the present invention. DETAILED DESCRIPTION

[0088] The following will be combined with the drawings in the embodiments of the present disclosure to clearly and completely describe the technical solutions in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0089] Embodiment 1:

[0090] See also Figure 1 and Figure 2 The present invention provides a technical solution: an interactive trajectory prediction method considering driving style uncertainty, comprising the following steps:

[0091] S1. Receive the vehicle status information data set obtained based on a typical road, perform data preprocessing, and divide the preprocessed data into a training set and a test set, as follows:

[0092] Taking the highly interactive scene of highway as an example, we select the highway highD dataset:

[0093] The highD dataset is a high-precision vehicle natural trajectory dataset from the German highway. There are 6 roads in total, with a sampling rate of 25Hz and an average visible duration of 13.6 seconds for each vehicle.

[0094] The data set contains the following data information: Vehicle_ID (vehicle number in the data set), Frame_ID (frame time, unit: s), Global_Time (global time, unit: s), Local_X (relative X coordinate, unit: m), Local_Y (relative Y coordinate, unit: m), v_leng (vehicle length, unit: m), v_class (vehicle type), etc. The highD data set has the advantages of bird's-eye view, large data volume, high precision, and easy processing, which ensures the integrity of the data;

[0095] Specifically, considering the consistency of the research, the scale of the data, and the richness of road characteristics, the roads in the dataset are screened. The screening rules and typical roads obtained by the screening are as follows:

[0096] 1) The amount of road data is large, basically covering a variety of driving behaviors; 2) The road features are rich, and it should be a two-way six-lane road; 3) The vehicle motion characteristics should be universal;

[0097] The road numbered 1 is the typical road in the final screening of this example. It covers 11.07 hours and involves 85,962 vehicles, of which 81.14% are cars and 18.86% are trucks. The number of samples accounts for 61.67% of the highD dataset. It has six lanes in both directions and a speed limit of 120 km / h. The above characteristics indicate that the road is highly representative.

[0098] Furthermore, the data preprocessing of the data set on a typical road is performed as follows:

[0099] Since the behavior of vehicles tends to be homogeneous in severe traffic congestion, vehicles with an average speed lower than 15 m / s are excluded. For abnormal data, such as zero values ​​in distance headway (DHW) and time headway (THW), the last observed non-zero value is used to replace them. It is assumed that THW remains unchanged when the vehicle leaves the observation area. The duration of the historical trajectory is set to 3 seconds and the predicted duration is 5 seconds, so vehicles with a total trajectory duration of less than 8 seconds are excluded.

[0100] S2. Based on the preprocessed data, the driving conditions are optimized, and the vehicle driving style is identified using the Gaussian mixture clustering method to obtain driving style information, as follows:

[0101] (S21) Optimizing the driving conditions, the specific process is as follows:

[0102] Extract vehicles that did not change lanes, accounting for about 89% of the vehicles after data preprocessing;

[0103] For the above vehicles, the driving conditions are divided according to the hybrid strategy consisting of time headway (THW) and time-to-collision (Time-to-Collision), which can be expressed as:

[0104] d t =max{t hc v e +t cc (v e -v p ),d0} (1)

[0105] Where, d t Indicates the distance between the two vehicles that triggers the switching condition; t hc It represents the headway coefficient, and its value is 5s; t cc represents the collision time coefficient, its value is 1s; d0 is the distance constant, its value is 100m; v e is the vehicle speed, v p is the speed of the preceding vehicle;

[0106] (S22) Using the Gaussian mixture clustering method to identify the vehicle driving style, the specific process is as follows:

[0107] (S22.1) First, select appropriate characteristic indicators, including: vehicle speed standard deviation (unit: m / s), maximum acceleration (unit: m / s2), minimum acceleration (unit: m / s2), average acceleration (unit: m / s2) and acceleration standard deviation (unit: m / s2), maximum headway distance (unit: m), minimum headway distance (unit: m), average headway distance (unit: m) and standard deviation of headway distance (unit: m), maximum headway time distance (unit: s), minimum headway time distance (unit: s), average headway time distance (unit: s), standard deviation of headway time distance (unit: s), maximum acceleration derivative (unit: m / s3), minimum acceleration derivative (unit: m / s3), average acceleration derivative (unit: m / s3) and standard deviation of acceleration derivative (unit: m / s3);

[0108] (S22.2) Based on the selected feature indicators, the driving styles of the vehicles are clustered using a Gaussian mixture clustering method:

[0109]

[0110]

[0111] Where K represents the number of Gaussian components, π n represents the weight of the nth Gaussian component, Θ represents the parameter set of Gaussian mixture clustering, which can be expressed as Θ = {π m ,μ n ,Σ n}, p(χ m |Θ) represents the sample vector χ m The probability density function, N(χ m |μ n ,Σ n ) indicates that the mean vector is μ m And the covariance matrix is ​​Σ n Gaussian distribution of , d represents the dimension of the sample vector;

[0112] S3. Construct a trajectory prediction model, which includes an encoder part and a decoder part, and the encoder part introduces a lane attention mechanism and a multi-head attention mechanism, as follows:

[0113] (S31) constructing a trajectory prediction model, as follows:

[0114] The essence of vehicle trajectory prediction is a time series problem, that is, predicting the future trajectory of the target vehicle based on the historical trajectory of the target vehicle and its surrounding vehicles;

[0115] The historical trajectory can be expressed as:

[0116]

[0117] Where H represents the historical trajectory, which contains the vehicle information within a time window, t h represents the number of time points in the historical trajectory, represents the trajectory information of the heterogeneous vehicle group at the target time t, i = 1, 2, ..., 9; represents the trajectory information of the i-th vehicle at time t, which can be expressed as:

[0118]

[0119] In the formula, They represent the longitudinal position and lateral position of the i-th vehicle at time t respectively; denote the longitudinal and lateral velocities, respectively; Represents longitudinal and lateral acceleration respectively; T i represents the type of the i-th vehicle (such as a car or a truck); S i represents the driving style of the ith vehicle (e.g., aggressive, mild, or conservative);

[0120] Analyzing the relative motion of surrounding vehicles helps the target vehicle identify and adapt to different traffic scenarios, thereby achieving safer, more efficient and more comfortable driving behavior. Therefore, the historical information of relative motion is selected as the input of the model to capture the interaction and dynamic changes between vehicles, thereby improving the accuracy of trajectory prediction results:

[0121]

[0122] Where X is the input vector of the model, Indicates the relative motion information of the surrounding vehicles relative to the target vehicle. The value of i is a natural number from 1 to 9 (excluding 5). t=[t,t-1,...,tt h +1] is the time step, is the historical movement information of the target vehicle;

[0123] The output of the model is the future trajectory coordinates of the target vehicle, expressed as:

[0124]

[0125] Where Y is the output vector, t f is the number of time points of the predicted trajectory, t j =1,2,...,t f ,

[0126]

[0127] (S32) The lane attention mechanism is used to enhance the vehicle's spatial perception of the target lane. The specific implementation method is as follows:

[0128] Human drivers prioritize the trajectory and intention of the vehicle in the target lane, for example paying special attention to the right lane when turning right. This lane-dependent attention, especially prominent in highway scenarios, prompted the proposal of a lane attention mechanism that dynamically adjusts lane weights to enhance attention to relevant lane information. In a heterogeneous group of vehicles, three vehicles in the same lane are grouped into three groups, each of which is processed independently;

[0129] First, the encoder processes the vehicle information of each group through a long short-term memory network (LSTM) to generate a three-dimensional lane context vector, which is and Represent the left lane, middle lane, and right lane respectively; these vectors are three-dimensional and contain information at all time points in the historical trajectory;

[0130] According to these three lane context vectors and Y (t-1) To determine and The weight of (t-1) is the predicted position of the target vehicle at the previous moment, which is a two-dimensional vector;

[0131] On this basis, the weighted lane context vector is obtained and They correspond to the left lane, middle lane, and right lane respectively, and are all three-dimensional vectors, where t ranges from 1 to t f ;

[0132] The total context vector can be expressed as:

[0133]

[0134] In the formula, Φ represents the linear layer for dimension compression and information fusion, cat is used to connect data, and C t represents the total context vector;

[0135] (S33) The multi-head attention mechanism is used to extract high-dimensional global interaction information. The specific implementation method is as follows:

[0136] The lane attention mechanism provides spatial information about lane occupancy, but does not distinguish the impact of different vehicles in the same lane on the target vehicle. Therefore, a multi-head attention mechanism is introduced to extract high-dimensional global interaction information and study the complex relationship between the target vehicle and surrounding vehicles. It complements the lane attention mechanism by enhancing the perception of autonomous driving to traffic scenes, especially in scenarios where surrounding vehicles exhibit irregular or non-standard behaviors, such as sudden lane changes or unpredictable braking;

[0137] In order to demonstrate the multi-head attention mechanism, a single-layer multi-head attention (i.e., cross attention) is used as an example. Similar to the classic Transformer structure, the cross attention mechanism uses dot product operations to capture the interaction information between vehicles. To achieve this goal, the input information of the cross attention mechanism is divided into two types: target vehicle data, represented by X tar , and surrounding vehicle data, represented by X sur ;

[0138] Then X tar and X sur It is processed through three different linear layers, namely Φ qj , Φ kj and Φ vj ,These layers play a vital role in extracting key features and enhancing ,representation capabilities;

[0139] The "query" (Q), "key" (K), and "value" (V) can be expressed as:

[0140] Q j =Φ qj (X tar ) (9)

[0141] K j =Φ kj (X sur ) (10)

[0142] V j =Φ vj (X sur ) (11)

[0143] In the formula, Φ qj , Φ kj and Φ vj There are three different linear layers for dimension compression and information fusion; Q j Carrying the information of the target vehicle, and K j and V j Carrying data of surrounding vehicles;

[0144] The output of the multi-head attention mechanism is obtained through the softmax operation:

[0145]

[0146] Where, d k It's Q j and K j The column dimension, Z j is the weighted information of the jth head, which indicates the attention weight of surrounding vehicles to the target vehicle in a heterogeneous vehicle group;

[0147] Finally, the data extracted from all heads are concatenated, further processed by dimension compression and information fusion, and finally the result of the multi-head attention model is output:

[0148] Z=Φ(cat(Z1,...,Z j )) (13)

[0149] Wherein, j=1,2,...m represents the jth head, m is the number of heads, Z represents the output of the multi-head attention model, which encapsulates the high-dimensional interaction information between the surrounding vehicles and the target vehicle; then, the output of the multi-head attention model will be fused with the output of the lane attention model; multi-head attention can learn the detailed relationship between every two vehicles, it enhances the lane attention model because it considers the weight differences of vehicles in the same lane, which should exist in the lane attention, multi-head attention is introduced to refine the weight distribution of the lane attention mechanism, because the global association it captures can be used as reinforcement after information fusion, during the training process, the lane attention and multi-head attention mechanisms run in parallel, and their outputs are fused through the contextual attention module, in addition, a unified loss function is used for optimization to ensure the balanced contribution of the two mechanisms, thereby improving the model's ability to capture local and global interactions and improving prediction accuracy;

[0150] S4. Input the training set into the trajectory prediction model, and train the trajectory prediction model based on the loss function to obtain the trained trajectory prediction model, as follows:

[0151] The loss function is the root mean square error, which is as follows:

[0152] After obtaining the trained vehicle trajectory prediction model, the weights of the LSTM encoder and LSTM decoder in the model are fixed, and the loss function during training is as follows:

[0153]

[0154] In the formula, RMSE represents the root mean square error, x i Represents the true value of the longitudinal displacement of the trajectory, yi represents the true value of the lateral displacement of the trajectory, represents the predicted longitudinal displacement of the trajectory, Represents the predicted value of the lateral displacement of the trajectory; N represents the number of predicted points, N = 25;

[0155] S5. Input the test set into the trained trajectory prediction model, complete the test and performance evaluation of the trajectory prediction model, and use the trajectory prediction model to realize vehicle trajectory prediction;

[0156] Specifically, RMSE is used to evaluate the performance of the model. The smaller the RMSE value, the better the trajectory prediction effect of the vehicle trajectory prediction model.

[0157] Next, the technical solution is further elaborated in detail:

[0158] Based on the natural driving dataset, the interactive trajectory prediction method considering driving style uncertainty in a mixed driving environment proposed in this embodiment is verified using Python language and PyTorch framework.

[0159] Figure 3 is the behavior characteristic diagram under following and cruising conditions, where Figure 3 (a) represents the standard deviation of longitudinal velocity, Figure 3 (b) represents the standard deviation of longitudinal acceleration. Figure 3 (c) represents the average distance from the front of the vehicle. It can be found that the median standard deviation of the longitudinal speed under the cruising and following conditions is 1.63km / h and 2.38km / h, respectively, which is significantly different. In addition, the data distribution of the standard deviation of the longitudinal acceleration under the cruising and following conditions is similar to Figure 3 (a) is consistent, indicating that the vehicle speed is relatively more stable in the cruising condition. In the following condition, due to the influence of the preceding vehicle, frequent acceleration and deceleration events lead to large speed changes. Figure 3 As shown in (c), the lower quartile and upper quartile of the average headway under cruising and following conditions are 157.6m and 258.5m and 29.4m and 68.5m, respectively, indicating that there are significant differences in the following distances under the two driving conditions.

[0160] The vehicle driving style characteristics obtained by the Gaussian mixture clustering method are shown in Table 1. Table 1 is as follows:

[0161] Table 1 Comparison of personalized driving styles

[0162] Characteristic parameters Radical General Conservative Speed ​​standard deviation (m / s) 0.258 0.164 0.287 <![CDATA[Maximum acceleration (m / s 2 )]]> 0.176 0.086 0.191 <![CDATA[Minimum acceleration (m / s 2 )]]> -0.232 -0.081 -0.264 <![CDATA[Average acceleration (m / s 2 )]]> -0.028 0.002 -0.050 <![CDATA[Standard deviation of acceleration (m / s 2 )]]> 0.136 0.057 0.146 Maximum distance between vehicles (m) 53.018 60.611 108.571 Minimum headway distance (m) 48.233 55.592 51.041 Average headway distance (m) 50.679 58.369 78.234 Standard deviation of headway distance (m) 1.420 1.369 21.442 Maximum headway distance(s) 1.771 2.050 3.496 Minimum headway distance (s) 1.626 1.899 1.662 Average headway time (s) 1.699 1.976 2.523 Headway standard deviation (s) 0.043 0.045 0.681 <![CDATA[Maximum value of acceleration derivative (m / s 3 )]]> 0.010 0.010 0.012 <![CDATA[Minimum value of acceleration derivative (m / s 3 )]]> -0.011 -0.010 -0.015 <![CDATA[Average value of acceleration derivative (m / s 3 )]]> -0.001 -0.001 -0.002 <![CDATA[Standard deviation of acceleration derivative (m / s 3 )]]> 0.006 0.005 0.007

[0163] From the results in Table 1, we can see that aggressive and conservative vehicles show significant speed fluctuations, while moderate vehicles have the smallest speed change, because aggressive drivers prioritize efficiency and therefore adjust their speed frequently, while conservative drivers prioritize safety and frequently change their speed due to interference from surrounding vehicles; moderate drivers focus on comfort, show a more stable speed, and are less sensitive to external interference, so their speed changes are smaller; from a spatiotemporal perspective, the average headway of aggressive vehicles fluctuates the least, because they tend to follow the vehicle in front, and the average following time distance of conservative vehicles fluctuates the most, because their cautious driving behavior will maintain a larger distance to deal with potential risks from the vehicle in front. Conservative drivers frequently adjust their speed and distance according to the behavior of the vehicle in front, resulting in large fluctuations in the following time distance. In summary, driving style has a good recognition effect and can provide reliable information support for vehicle trajectory planning.

[0164] Figure 4 This is a trajectory prediction effect diagram of the method described in this embodiment. The results show that after adopting this method, the trajectory prediction error of the vehicle within 5 seconds is effectively controlled, and the average values ​​of the total error, longitudinal error and lateral error do not exceed 0.34 meters, 0.33 meters and 0.10 meters respectively. In short, this method can significantly improve the trajectory prediction accuracy and provide more accurate information support for the vehicle decision-making planning module, thereby effectively improving the driving safety, comfort and traffic efficiency of autonomous driving vehicles.

[0165] Embodiment 2:

[0166] This embodiment provides an interactive trajectory prediction system considering driving style uncertainty, which is used to implement the interactive trajectory prediction method considering driving style uncertainty described in the first embodiment, including:

[0167] A preprocessing module, used for receiving a vehicle status information data set obtained based on a typical road, performing data preprocessing, and dividing the preprocessed data into a training set and a test set;

[0168] An identification module is used to optimize the driving conditions based on the preprocessed data and identify the vehicle driving style using a Gaussian mixture clustering method to obtain driving style information;

[0169] A construction module is used to construct a trajectory prediction model, wherein the trajectory prediction model includes an encoder part and a decoder part, and the encoder part introduces a lane attention mechanism and a multi-head attention mechanism;

[0170] A training module, used for inputting the training set into the trajectory prediction model, and training the trajectory prediction model based on the loss function to obtain a trained trajectory prediction model;

[0171] The test output module is used to input the test set into the trained trajectory prediction model, complete the testing and performance evaluation of the trajectory prediction model, and use the trajectory prediction model to realize vehicle trajectory prediction.

[0172] Specifically, the preprocessing module, identification module, construction module, training module and test output module can be embedded in a computer processing system. The computer calls the modules to complete the task of predicting the future trajectory of the vehicle according to the interactive trajectory prediction method considering the uncertainty of driving style provided above. The preprocessing module, identification module, construction module, training module and test output module can perform operations according to the specific steps given in the interactive trajectory prediction method considering the uncertainty of driving style.

[0173] It should be noted that it should be understood that the division of the various modules of the above system is only the division of logical functions. In actual implementation, they can be fully or partially integrated into one physical entity, or they can be physically separated, and these modules can all be implemented in the form of software calling through processing elements; they can also be all implemented in the form of hardware; some modules can also be implemented in the form of software calling through processing elements, and some modules can be implemented in the form of hardware. For example, the preprocessing module can be a separately established processing element, or it can be integrated in a certain chip of the above device. In addition, it can also be stored in the memory of the above device in the form of program code, and called and executed by a certain processing element of the above device. The implementation of other modules is similar. In addition, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each module above can be completed by the hardware integrated logic circuit in the processor element or the instructions in the form of software.

[0174] For example, the above modules may be one or more integrated circuits configured to implement the above methods, such as one or more application specific integrated circuits (ASIC), or one or more digital singnal processors (DSP), or one or more field programmable gate arrays (FPGA). For another example, when a module is implemented in the form of a processing element scheduling program code, the processing element may be a general-purpose processor, such as a central processing unit (CPU) or other processors that can call program code. For another example, these modules may be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0175] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0176] For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances. When an element is referred to as being "assembled on", "installed on", "fixed on" or "set on" another element, it can be directly on the other element or there can also be a centered element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be a centered element at the same time. The terms "vertical", "horizontal", "up", "down", "left", "right" and similar expressions used herein are only for illustrative purposes and are not intended to be the only implementation method.

[0177] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0178] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

Claims

1. An interactive trajectory prediction method considering driving style uncertainty, characterized in that: The following steps are involved: Receiving a vehicle status information data set obtained based on a typical road for data preprocessing, and dividing the preprocessed data into a training set and a test set; Based on the preprocessed data, the driving conditions are optimized, and the vehicle driving style is identified using the Gaussian mixture clustering method to obtain driving style information; Constructing a trajectory prediction model, wherein the trajectory prediction model includes an encoder part and a decoder part, and the encoder part introduces a lane attention mechanism and a multi-head attention mechanism; The training set is input into the trajectory prediction model, and the trajectory prediction model is trained based on the loss function to obtain a trained trajectory prediction model; The test set is input into the trained trajectory prediction model to complete the testing and performance evaluation of the trajectory prediction model, and the trajectory prediction model is used to realize vehicle trajectory prediction.

2. The interactive trajectory prediction method considering driving style uncertainty according to claim 1, characterized in that: The typical roads are selected from the highway highD dataset. The selection rules of typical roads are as follows: (21) The amount of road data is large and covers a variety of driving behaviors; (22) The road features a variety of features, with six lanes in both directions; (23) Vehicle motion characteristics are universal.

3. The interactive trajectory prediction method considering driving style uncertainty according to claim 2, characterized in that: The vehicle status information data set includes the vehicle number, frame time, global time, relative X coordinate, relative Y coordinate, vehicle length and vehicle model data in the data set.

4. The interactive trajectory prediction method considering driving style uncertainty according to claim 3, characterized in that: The vehicle status information data set is preprocessed as follows: (41) Exclude vehicles with an average speed lower than 15 m / s; (42) For zero values ​​in abnormal data, the last observed non-zero value is used for replacement; (43) Vehicles whose total trajectory duration is less than 8 seconds are excluded.

5. The interactive trajectory prediction method considering driving style uncertainty according to claim 1, characterized in that: Based on the preprocessed data, the driving conditions are optimized, and the vehicle driving style is identified using the Gaussian mixture clustering method to obtain the driving style information, as follows: (51) Optimize the driving conditions. The specific process is as follows: Vehicles that did not change lanes were extracted, and the driving conditions were divided according to the mixed strategy consisting of headway and collision time, which can be expressed as: d t =max{t hc v e +t cc (v e -v p ),d0} (1) Where, d t Indicates the distance between the two vehicles that triggers the working mode switching condition; t hc It represents the headway coefficient, and its value is 5s; t cc represents the collision time coefficient, its value is 1s; d0 is the distance constant, its value is 100m; v e is the vehicle speed, v p is the speed of the preceding vehicle; (52) The Gaussian mixture clustering method is used to identify the vehicle driving style. The specific process is as follows: (52.1) Select characteristic indicators, which include: vehicle speed standard deviation, maximum acceleration, minimum acceleration, average acceleration and acceleration standard deviation, maximum headway distance, minimum headway distance, average headway distance and headway distance standard deviation, maximum headway time distance, minimum headway time distance, average headway time distance, headway time distance standard deviation, maximum acceleration derivative, minimum acceleration derivative, average acceleration derivative and acceleration derivative standard deviation; (52.2) Based on the selected feature indicators, the driving styles of the vehicles are clustered using the Gaussian mixture clustering method: Where K represents the number of Gaussian components, π n represents the weight of the nth Gaussian component, Θ represents the parameter set of Gaussian mixture clustering, which can be expressed as Θ = {π m ,μ n ,Σ n }, p(χ m |Θ) represents the sample vector χ m The probability density function, N(χ m |μ n ,Σ n ) indicates that the mean vector is μ m And the covariance matrix is ​​Σ n Gaussian distribution, d represents the dimension of the sample vector.

6. The interactive trajectory prediction method considering driving style uncertainty according to claim 1, characterized in that: Construct a trajectory prediction model as follows: (61)The historical trajectory can be expressed as: Where H represents the historical trajectory, which contains the vehicle information within a time window, t h represents the number of time points in the historical trajectory, represents the trajectory information of the heterogeneous vehicle group at the target time t, i = 1, 2, ..., 9; represents the trajectory information of the i-th vehicle at time t, which can be expressed as: In the formula, They represent the longitudinal position and lateral position of the i-th vehicle at time t respectively; denote the longitudinal and lateral velocities, respectively; Represents longitudinal and lateral acceleration respectively; T i represents the type of the i-th vehicle; S i represents the driving style of the i-th vehicle; (62) Select historical trajectories as the input of the model: Where X is the input vector of the model, Indicates the relative motion information of surrounding vehicles relative to the target vehicle. The value of i is a natural number from 1 to 9 except 5. t=[t,t-1,...,tt h +1] is the time step, is the historical movement information of the target vehicle; (63) The output of the model is the future trajectory coordinates of the target vehicle, expressed as: Where Y is the output vector, t f is the number of time points of the predicted trajectory, t j =1,2,...,t f , 7. The interactive trajectory prediction method considering driving style uncertainty according to claim 6, characterized in that: The lane attention mechanism is used to enhance the vehicle's spatial perception of the target lane. The specific implementation method is as follows: (71) In a heterogeneous vehicle group, three vehicles in the same lane are grouped into one group, with a total of three groups, each of which is processed independently; First, the encoder processes the vehicle information of each group through the long short-term memory network to generate a three-dimensional lane context vector, which is and Represent the left lane, middle lane, and right lane respectively; these vectors contain information at all time points in the historical trajectory; (72) According to these three lane context vectors and Y (t-1) To determine and The weight of (t-1) is the predicted position of the target vehicle at the previous moment, which is a two-dimensional vector; On this basis, the weighted lane context vector is obtained and They correspond to the left lane, middle lane, and right lane respectively, and are all three-dimensional vectors, where t ranges from 1 to t f ; (73)The total context vector can be expressed as: In the formula, Φ represents the linear layer for dimension compression and information fusion, cat is used to connect data, and C t Represents the total context vector.

8. The interactive trajectory prediction method considering driving style uncertainty according to claim 7, characterized in that: The multi-head attention mechanism is used to extract high-dimensional global interaction information. The specific implementation method is as follows: (81) The input information of the multi-head attention mechanism is divided into two types: target vehicle data, represented by X tar , and surrounding vehicle data, represented by X sur ; (82)X tar and X sur It is processed through three different linear layers, namely Φ qj , Φ kj and Φ vj ,These layers play a vital role in extracting key features and enhancing ,representation capabilities; The "query" (Q), "key" (K), and "value" (V) can be expressed as: Q j =Φ qj (X tar ) (9) K j =Φ kj (X sur ) (10) V j =Φ vj (X sur ) (11) In the formula, Φ qj , Φ kj and Φ vj There are three different linear layers for dimension compression and information fusion; Q j Carrying the information of the target vehicle, and K j and V j Carrying data of surrounding vehicles; (83) The output of the multi-head attention mechanism is obtained through the softmax operation: Where, d k It's Q j and K j The column dimension, Z j is the weighted information of the jth head, which indicates the attention weight of surrounding vehicles to the target vehicle in a heterogeneous vehicle group; Finally, the data extracted from all heads are concatenated, further processed by dimension compression and information fusion, and finally the result of the multi-head attention model is output: Z=Φ(cat(Z1,…,Z j )) (13) Where j=1,2,...m represents the jth head, m is the number of heads, and Z represents the output of the multi-head attention model, which encapsulates the high-dimensional interaction information between the surrounding vehicles and the target vehicle; then, the output of the multi-head attention model will be fused with the output of the lane attention model.

9. The interactive trajectory prediction method considering driving style uncertainty according to claim 8, characterized in that: The loss function is the root mean square error, which is as follows: After obtaining the trained vehicle trajectory prediction model, the weights of the LSTM encoder and LSTM decoder in the model are fixed, and the loss function during training is as follows: In the formula, RMSE represents the root mean square error, x i Represents the true value of the longitudinal displacement of the trajectory, y i represents the true value of the lateral displacement of the trajectory, represents the predicted longitudinal displacement of the trajectory, Represents the predicted value of the lateral displacement of the trajectory; N represents the number of predicted points.

10. An interactive trajectory prediction system considering driving style uncertainty, used to implement the interactive trajectory prediction method considering driving style uncertainty according to any one of claims 1 to 9, characterized in that: include: A preprocessing module, used for receiving a vehicle status information data set obtained based on a typical road, performing data preprocessing, and dividing the preprocessed data into a training set and a test set; An identification module is used to optimize the driving conditions based on the preprocessed data and identify the vehicle driving style using a Gaussian mixture clustering method to obtain driving style information; A construction module is used to construct a trajectory prediction model, wherein the trajectory prediction model includes an encoder part and a decoder part, and the encoder part introduces a lane attention mechanism and a multi-head attention mechanism; A training module, used for inputting the training set into the trajectory prediction model, and training the trajectory prediction model based on the loss function to obtain a trained trajectory prediction model; The test output module is used to input the test set into the trained trajectory prediction model, complete the testing and performance evaluation of the trajectory prediction model, and use the trajectory prediction model to realize vehicle trajectory prediction.

Images