A method for vehicle driving intent recognition and trajectory prediction for autonomous driving

Abstract

This invention describes a method for vehicle driving intention recognition and trajectory prediction for autonomous driving, comprising the following steps: constructing a dynamic threshold sliding sampling window and dynamically adjusting the window length according to road condition complexity; extracting the lateral displacement change of the interactive vehicle within the dynamic threshold sliding sampling window and outputting the driving intention of the interactive vehicle; based on the driving intention, constructing a bimodal adaptive graph attention network, fusing the physical interaction graph and semantic interaction graph of the bimodal adaptive graph attention network, outputting the motion feature vectors of the final node and the interactive vehicle, calculating the probability distribution of the driving intention, and outputting the probability distribution of the driving intention features of the interactive vehicle; based on the motion feature vectors of the interactive vehicle and the probability distribution of the driving intention of the interactive vehicle, outputting the spatiotemporal interaction features; and constructing an intention feature fusion mechanism to integrate the motion feature vectors of the interactive vehicle, the driving intention features of the interactive vehicle, and the spatiotemporal interaction features to predict the future trajectory of the interactive vehicle.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, and specifically relates to a method for vehicle driving intention recognition and trajectory prediction for autonomous driving. Background Technology

[0002] In autonomous driving, accurately identifying the driving intentions of interacting vehicles and predicting their trajectories is a crucial problem that needs to be solved to ensure the safe operation of autonomous vehicles. Current trajectory prediction methods primarily rely on the target vehicle's own state and vehicle interactions to make single trajectory predictions, rarely considering influencing factors such as the vehicle's driving intentions, the motion states of surrounding vehicles, and environmental information. This results in the following shortcomings or limitations:

[0003] (1) In terms of driving intention recognition, existing methods are poorly adapted to complex road conditions. For example, the traditional sliding window algorithm is difficult to extract lane change data under complex road conditions, resulting in inaccurate data and affecting the accuracy of intention recognition. At the same time, existing methods have shortcomings in dealing with the diverse impact of vehicle intentions on target vehicles at different times and spaces in complex traffic scenarios. They cannot fully integrate the spatiotemporal interaction information between vehicles and make it difficult to accurately identify driving intentions.

[0004] (2) In terms of trajectory prediction, existing methods are prone to feature confusion when distinguishing between the target vehicle and surrounding vehicles, making it impossible to accurately grasp the vehicle's motion state. Moreover, existing models generally ignore the uncertainty and spatiotemporal interaction characteristics of driving behavior, resulting in limited long-term prediction accuracy and failing to meet the stringent safety and reliability requirements of autonomous driving systems. Summary of the Invention

[0005] To more accurately determine the driving intention of interactive vehicles and reduce the prediction bias of autonomous vehicles regarding the future trajectories of surrounding vehicles, thereby at least partially addressing the problems existing in the prior art, this invention provides a method for vehicle driving intention recognition and trajectory prediction for autonomous driving. The specific technical solution is as follows:

[0006] A method for vehicle driving intent recognition and trajectory prediction for autonomous driving includes the following steps:

[0007] S1. Construct a dynamic threshold sliding sampling window and dynamically adjust its length according to the road condition complexity; extract the lateral displacement change of the interactive vehicle within the dynamic threshold sliding sampling window and output the driving intention of the interactive vehicle.

[0008] S2, based on the driving intention of the interactive vehicle, construct a bimodal adaptive graph attention network, set the bottom physical interaction graph and the upper semantic interaction graph of the bimodal adaptive graph attention network, fuse the physical interaction graph and the semantic interaction graph, and output the motion feature vectors of the final node and the interactive vehicle.

[0009] S3, calculate the probability distribution of driving intention for the final node and output the probability distribution of the driving intention features of the interactive vehicle;

[0010] S4, based on the motion feature vector of the interactive vehicle and the probability distribution of the driving intention of the interactive vehicle, constructs a long-period attention encoder to capture the global trend features of the target vehicle; constructs a gated local feature extraction module to identify the key frame abrupt behavior of surrounding vehicles and capture the motion changes of surrounding vehicles; outputs spatiotemporal interaction features.

[0011] S5 constructs an intent feature fusion mechanism to integrate the motion feature vector of the interactive vehicle, the driving intent feature of the interactive vehicle, and the spatiotemporal interaction feature to predict the future trajectory of the interactive vehicle.

[0012] Furthermore,

[0013] I. The road condition complexity is calculated using the following formula:

[0014] ,

[0015] In the formula, T is the road condition complexity index; Let be the standard deviation of the lateral acceleration of the interacting vehicles; k is the lane curvature. For the density of interactive vehicles; They are respectively The corresponding empirical weighting coefficients;

[0016] II. Set the window length L of the dynamic threshold sliding sampling window according to the following rules:

[0017] set up Due to complex road conditions, ;set up The road conditions are of moderate complexity. ;set up For simple road conditions, ;in, Base window length;

[0018] III. Through formula Calculate the lateral displacement change of the interactive vehicle, and output the driving intention of the interactive vehicle under the following conditions:

[0019] when The longitudinal acceleration of the target vehicle The output of the interactive vehicle's driving intention is to maintain a straight course; when Meters and lane change direction angle The output of the interactive vehicle's driving intention is to change lanes to the left; when Meters and lane change direction angle The output vehicle's driving intention is to change lanes to the right;

[0020] In the formula, This represents the lateral displacement change of the interactive vehicles; The coordinates of the horizontal position of the Lth frame within the dynamic threshold sliding sampling window; The coordinates of the horizontal position of the first frame within the dynamic threshold sliding sampling window.

[0021] Furthermore,

[0022] I. The spatiotemporal graph structure of the physical interaction graph is constructed using the following expression:

[0023] ,

[0024] In the formula, This is a holistic representation of the physical interaction diagram; A set of nodes in a physical interaction graph; It is the set of edges in the physical interaction graph;

[0025] The combined edge weights of the physical interaction graph are calculated using the following expression:

[0026] ,

[0027] In the formula, The combined edge weights between node i and node j in the physical interaction graph; For spatiotemporal distance attenuation; The influence of relative motion; For the Sigmoid function; Let i be the node feature vector of node i; Let be the node feature vector of node j; This is a learnable weight matrix used to calculate the relationships between nodes; This is a bias term used to calculate the correlation between nodes;

[0028] The spatiotemporal distance decay is calculated using the following expression:

[0029] ,

[0030] In the formula, For spatiotemporal distance attenuation; Let i be the position coordinates of the interactive vehicle; Let J be the position coordinates of the interactive vehicle j; For the observation timestamp of interactive vehicle i; For the observation timestamp of interactive vehicle j; This is a hyperparameter of the spatiotemporal distance attenuation coefficient; This is a hyperparameter representing the spatial weighting in spatiotemporal distance calculation;

[0031] The effect of relative motion is calculated using the following expression:

[0032] ,

[0033] In the formula, The influence of relative motion; Let i be the three-dimensional velocity vector of the interactive vehicle i; Let be the three-dimensional velocity vector of the interactive vehicle j; The attenuation coefficient hyperparameter for the effects of relative motion;

[0034] II. Construct the semantic interaction graph structure using the following expression:

[0035] ,

[0036] In the formula: This is a holistic representation of the semantic interaction graph; A set of nodes for a semantic interaction graph; It is the set of edges of the semantic interaction graph;

[0037] The combined edge weights of the semantic interaction graph are calculated using the following expression:

[0038] ,

[0039] In the formula, To integrate edge weights; For intent similarity; Intended influence degree; Let i be the position coordinates of the interactive vehicle; Let J be the position coordinates of the interactive vehicle j; For the observation timestamp of interactive vehicle i; For the observation timestamp of interactive vehicle j; This is a hyperparameter representing the attenuation coefficient of spatial distance on intentional interaction.

[0040] Intent similarity Calculated using the following expression:

[0041] ,

[0042] Intentional Influence Calculated using the following expression:

[0043] ,

[0044] In the formula: For the Sigmoid function; Let i be the feature vector of the driving intention of the interactive vehicle i. Let be the feature vector of the driving intention of the interactive vehicle j; To calculate the learnable weight matrix for the influence of intent; This is a bias term used to calculate the degree of influence of the intention.

[0045] Furthermore,

[0046] The following steps are used to output the motion feature vector of the interactive vehicle:

[0047] I. The physical interaction graph and the semantic interaction graph are fused using the following expression:

[0048] ,

[0049] In the formula, The fusion gating coefficients for the physical interaction graph and the semantic interaction graph; The learnable weight matrix for the gating mechanism; Use the Sigmoid activation function; The feature representation of node i in the physical interaction graph; The feature representation of node i in the semantic interaction graph; For learnable bias terms of the gating mechanism; For gating mechanisms;

[0050] II. Input the historical trajectory sequence into the bidirectional GRU, and output the interactive vehicle driving intention feature vector. Calculate the semantic association between nodes using the following expression:

[0051] ,

[0052] In the formula, For semantic relationships between nodes; This represents the learnable weight matrix for the query in the attention mechanism; This represents the learnable weight matrix of the keys in the attention mechanism; Let i be the feature vector of the driving intention of the interactive vehicle i. Let be the feature vector of the driving intention of the interactive vehicle j; Let be the driving intention feature vector of the interactive vehicle k; n is the index of the neighboring node of node i. Let i be the set of neighbors of node i;

[0053] III. Output the motion feature vector of the interactive vehicle using the following expression:

[0054] ,

[0055] In the formula, The motion feature vector of the interactive vehicle; The fusion gating coefficients for the physical interaction graph and the semantic interaction graph; The feature representation of node i in the physical interaction graph; The feature representation of node i in the semantic interaction graph; For gating mechanisms;

[0056] The expression for the final node in the bimodal computation is as follows:

[0057] ,

[0058] In the formula, For graph attention network Feature representation of node i in layer; GAT is a graph attention network structure; This represents the feature representation of node i in the l-th layer of the graph attention network; This is the feature representation of node j in the l-th layer of the graph attention network; This refers to the semantic relationships between nodes.

[0059] Furthermore,

[0060] The final node is subjected to a multi-layer nonlinear transformation using the following expression:

[0061] ,

[0062] In the formula, This is the output feature vector of the first layer of the multilayer perceptron; This is the output feature vector of the second layer of the multilayer perceptron; , N Number of intent categories; This is the comprehensive feature vector after dimensional transformation; This refers to the bias term of the second layer of the multilayer perceptron; This refers to the bias term of the third layer of the multilayer perceptron.

[0063] The driving intent characteristics are output using the following expression:

[0064] ,

[0065] In the formula, P represents the driving intention feature; This is the output feature vector of the second layer of the multilayer perceptron; is the c-th element in the output feature vector of the second layer of the multilayer perceptron; c is the intent category index; N is the total number of intent categories.

[0066] Furthermore,

[0067] The long-cycle attention encoder for the target vehicle is constructed using the following steps:

[0068] I. Input the historical trajectory sequence of the target vehicle , Let D be the time step and D be the feature dimension. The attention weights are calculated using the following expression:

[0069] ,

[0070] In the formula, , These are the query, key, and value matrices for the attention mechanism. Learnable weights for generating the query matrix, Let be the feature vector of the target vehicle in frame t; The dimension of the key vector; For time step;

[0071] II. Output the global trend characteristics of the target vehicle using the following expression:

[0072] ,

[0073] In the formula, This represents the overall trend characteristics; For layer normalization; For regularization; For the attention of the bulls.

[0074] Furthermore,

[0075] The gated local feature extraction module is constructed using the following steps:

[0076] I. Enter the trajectory of surrounding vehicles , Let D be the time step and D be the feature dimension. Local temporal features are extracted using the following expression:

[0077] ,

[0078] In the formula, The output features of a one-dimensional causal convolution. The trajectory feature sequence of surrounding vehicles;

[0079] II. Calculate the characteristic response at the abrupt change time using the following expression:

[0080] ,

[0081] In the formula, Let be the gating weight for frame t; For the Sigmoid function; The learnable weight matrix for the gating mechanism; The output features of a one-dimensional causal convolution; The difference features between adjacent frames; For learnable bias terms of the gating mechanism;

[0082] III. Output the local mutation features of surrounding vehicles using the following expression:

[0083] ,

[0084] In the formula, This represents the local mutation feature vector of the surrounding vehicles; For the first t Frame gating weights; The output features of a one-dimensional causal convolution;

[0085] The global trend features of the target vehicle are adaptively fused with the local abrupt change features of surrounding vehicles to output spatiotemporal interaction features, expressed as follows:

[0086] ,

[0087] In the formula, MLP is used to coordinate the feature scales of different modalities to ensure spatiotemporal consistency; Features of spatiotemporal interaction; This represents the local mutation feature vector of the surrounding vehicles; This represents the overall trend characteristics.

[0088] Furthermore,

[0089] A gated fusion module is constructed to dynamically adjust the contribution weights of multimodal features, wherein the feature importance weights are calculated using the following expression:

[0090] ,

[0091] In the formula, These represent the weights of motion, interaction, and intent features, respectively. This is a learnable matrix used to calculate the feature importance weights; Features of spatiotemporal interaction; is the motion feature vector of the interactive vehicle; p is the driving intention feature.

[0092] Weighted feature fusion is calculated using the following expression:

[0093] ,

[0094] In the formula, For weighted feature fusion; The weights for the motion features; The weights of the interaction features; The weights of the intent features; Features of spatiotemporal interaction; is the motion feature vector of the interactive vehicle; p is the driving intention feature.

[0095] Furthermore,

[0096] The distribution of interactive vehicle multimodal trajectory prediction is performed using the following expression:

[0097] ,

[0098] In the formula, Given input features X, let X be the probability distribution of the future trajectory Y. The mixing coefficient for intention category k; Let be the mean and covariance of the Gaussian distribution;

[0099] The final output is obtained by sampling multiple possible predicted trajectories from the predicted distribution, as shown in the following expression:

[0100] ,

[0101] In the formula, This is the output set of multimodal trajectory prediction; For the predicted M possible trajectories; Let X be the probability distribution of the future trajectory Y given input features X.

[0102] Furthermore,

[0103] The interactive vehicle trajectory prediction error and intent classification loss are optimized using the following expression to output the multimodal trajectory of the interactive vehicle:

[0104] ,

[0105] In the formula, The joint loss for trajectory prediction and intent classification; The loss is negative log-likelihood. Cross-entropy loss; To balance the hyperparameters; Y is the true trajectory label; represents the output set of multimodal trajectory prediction; e represents the true intent label. The predicted intent label.

[0106] Based on the above technical solution, the method described in this invention has the following beneficial effects:

[0107] 1. The method described in this invention effectively improves the accuracy of lane change data extraction through a dynamic threshold sliding sampling window algorithm; the bimodal adaptive graph attention network can fuse spatiotemporal interaction information between vehicles, greatly improving the accuracy of driver intent recognition.

[0108] 2. In the method described in this invention, the differentiated spatiotemporal feature extraction architecture avoids feature confusion; the intent probability modeling and spatiotemporal interaction feature extraction improve the ability to handle the uncertainty of driving behavior and spatiotemporal interaction characteristics; the intent feature fusion mechanism realizes multimodal trajectory prediction, improves the accuracy and reliability of trajectory prediction, and has obvious advantages in long-term prediction tasks.

[0109] 3. The table below shows a comparison of the performance metrics of current mainstream models and the method described in this invention on the NGSIM dataset. As can be seen from the relevant performance test data in the table, the method described in this invention has significantly lower errors on the NGSIM dataset than other models, demonstrating better intent recognition and trajectory prediction capabilities.

[0110] Attached Figure Description

[0111] Figure 1 This is a schematic diagram of the overall process of the method described in the embodiment;

[0112] Figure 2 This is a schematic diagram illustrating the multimodal trajectory output effect described in the embodiment.

[0113] Figure 3 This is a schematic diagram of the attention encoder structure described in the embodiment. Detailed Implementation

[0114] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this disclosure or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without creative effort are within the scope of protection of this disclosure.

[0115] Unless otherwise defined, the technical or scientific terms used in this disclosure shall have the ordinary meaning as understood by one of ordinary skill in the art to which this disclosure pertains.

[0116] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0117] In all examples shown and discussed herein, any specific value should be interpreted as merely exemplary and not as a limitation; therefore, other examples of exemplary embodiments may have different values.

[0118] The target vehicle described in the specification and claims refers to the vehicle whose trajectory needs to be predicted; the driving environment refers to the driving environment in which the autonomous vehicle is in a stationary state and / or a moving state; the interacting vehicle refers to the surrounding vehicles other than the autonomous vehicle in the driving environment.

[0119] like Figures 1 to 3 As shown, this embodiment describes a method for vehicle driving intention recognition and trajectory prediction for autonomous driving. Figure 1 A flowchart illustrating the method is shown, which specifically includes the following steps.

[0120] S1. Construct a dynamic threshold sliding sampling window algorithm. This algorithm dynamically adjusts the sliding window length in real time based on the complexity of road conditions, extracts the lateral displacement changes of the interacting vehicle within the sliding window, and outputs the driving intention of the interacting vehicle. The specific steps are as follows:

[0121] S101, the road condition complexity of the driving environment is calculated using the following formula:

[0122] ,

[0123] In the formula,

[0124] T represents the complexity of road conditions in the driving environment;

[0125] The standard deviation of the lateral acceleration of the interactive vehicle is calculated based on the triaxial acceleration acquired in real time by the IMU inertial measurement unit on the vehicle.

[0126] k is the lane curvature of the driving environment. The density of interactive vehicles is calculated based on the relative position sequence obtained from millimeter-wave radar and visual fusion positioning.

[0127] They are respectively The corresponding empirical weighting coefficients, where, , .

[0128] S102, Define the base window length Based on the complexity of road conditions in the driving environment, the adaptive adjustment rules for the dynamic window length are set as follows:

[0129] set up For complex road conditions, the dynamic window length ;

[0130] set up For moderate road conditions, the dynamic window length is... ;

[0131] set up For simple road conditions, the dynamic window length is... .

[0132] S103: Centered on the current frame, extract continuous frame data according to the dynamic window length, perform sliding sampling of lane change data, and smooth the lane change data through Kalman filtering;

[0133] in:

[0134] ① Lane change data includes continuous changes in the lateral position, lateral velocity, lateral acceleration, and heading angle of the target vehicle and interacting vehicles during the lane change process. It also includes the relative distance, relative velocity, and lane line position between the target vehicle and vehicles in adjacent lanes during the lane change process.

[0135] ② The current frame and consecutive frame data are spatiotemporal fusion data collected synchronously by multiple sensors (such as cameras, vehicle GPS / IMU, millimeter-wave radar, etc.); the road scene ahead is captured by a camera (1920×1080 resolution, 25Hz frame rate) installed on the windshield of the target vehicle. Each frame contains visual information such as lane lines, surrounding vehicles, and traffic signs, which serve as the basis for lane boundary recognition, vehicle positioning, and visual judgment of lane change behavior.

[0136] ③ The real-time position (latitude and longitude, altitude), speed, acceleration and heading angle of the target vehicle are obtained through the vehicle-mounted GPS / IMU integrated navigation system, with a sampling frequency of 10Hz;

[0137] ④ The relative distance, speed, and azimuth of surrounding vehicles are acquired using millimeter-wave radar (detection range 150m, angular resolution 1°), with a sampling frequency of 5Hz. Image frames and kinematic data are hard-synchronized with time stamps (based on the vehicle's CAN bus clock, synchronization error <1ms). Each image frame is associated with the corresponding GPS position, IMU motion parameters, and radar detection results, forming a "visual-motion-interaction" multimodal fusion frame data.

[0138] S104, through formula Calculate the lateral displacement change of the interactive vehicle within the sliding window, and determine the driving intention of the interactive vehicle based on the following conditions:

[0139] when meter and longitudinal acceleration The driving intention of the interactive vehicle is to maintain a straight course;

[0140] when Meters and lane change direction angle The driving intention of the interactive vehicle is to change lanes to the left;

[0141] when Meters and lane change direction angle The driving intention of the interactive vehicle is to change lanes to the right.

[0142] In the above formula or expression,

[0143] longitudinal acceleration Acquired through measurements taken by the onboard inertial measurement unit (IMU) mounted on the target vehicle;

[0144] Lane change direction angle Obtained through calculation or measurement of the change in heading angle;

[0145] The horizontal position coordinates of the last frame (frame L) within the sliding window;

[0146] These are the horizontal coordinates of the first frame within the sliding window.

[0147] S105, the driving intention of the interactive vehicle is labeled, and the label for keeping straight is defined as [1,0,0], the label for changing lanes to the left is [0,1,0], and the label for changing lanes to the right is [0,0,1].

[0148] The preferred method is to use a time window voting method to correct the labels for lane change data. If a certain frame is determined to be a lane change intention, and at least two frames within each of the three frames before and after it meet the same intention condition, then the label is confirmed to be valid. Otherwise, it is marked as "uncertain" and removed, ensuring that the label accuracy rate exceeds 95%.

[0149] As a preferred option:

[0150] ① The lane change data is horizontally flipped, the time series is randomly pruned (keeping at least 80% of the original length), and Gaussian noise is added (standard deviation = 0.1). The sample size of the three data categories "keep straight", "left lane change", and "right lane change" after labeling is balanced to 1:1:1 to avoid class imbalance.

[0151] ② Extract motion features, such as velocity, acceleration, and heading angle, from frame data fused from multiple sensors using formulas. Z-score standardization is performed, where and These are the mean and standard deviation of the features in the training set (composed of lane change data), respectively, to ensure that the feature distribution of the input "bimodal adaptive graph attention network" described below is consistent.

[0152] ③ Adaptively adjust data augmentation strategies based on the characteristics of different time-series datasets (including training, validation, and test sets) composed of lane-change data. Data augmentation strategies include adding longitudinal motion noise to simulate minute speed fluctuations during following in highway scenarios (achieved by superimposing Gaussian noise onto the longitudinal speed); and randomly adding noise to the heading angle sequence in intersection scenarios. Offset amount enhances the robustness of steering intention recognition.

[0153] S2. To address the problem that traditional methods cannot fully integrate spatiotemporal interaction information between vehicles and to obtain more accurate representations of vehicle motion features, this step constructs a bimodal adaptive graph attention network based on the driving intentions of the interactive vehicles obtained in step S1.

[0154] The bimodal adaptive graph attention network constructs a physical interaction graph at its bottom layer using vehicle kinematic parameters (including position coordinates, velocity vector, acceleration, and heading angle), and a semantic interaction graph at its top layer using an intent reasoning network. A time decay gating mechanism is used to fuse the physical and semantic interaction graphs, outputting the final node of the bimodal computation, which in turn outputs the motion feature vector of the interacting vehicle. .

[0155] Specifically, the following steps are included:

[0156] Specifically, the following steps are included:

[0157] S201, Construct the physical interaction graph of the bimodal adaptive graph attention network, and output the comprehensive edge weights of the physical interaction graph. The steps are as follows:

[0158] S201-1, construct the spatiotemporal graph structure of the physical interaction graph using the following expression:

[0159] ,

[0160] In the formula, It is a holistic representation of the physical interaction graph, including a set of nodes and a set of edges; A set of nodes in a physical interaction graph; It is the set of edges of the physical interaction graph.

[0161] S201-2, Calculate the comprehensive edge weights of the physical interaction graph based on spatiotemporal distance attenuation and the influence of relative motion, as shown in the following expression:

[0162] ,

[0163] In the formula, The combined edge weights between node i and node j in the physical interaction graph; For spatiotemporal distance attenuation; The influence of relative motion; For the Sigmoid function; Let i be the node feature vector of node i; Let be the node feature vector of node j; This is a learnable weight matrix used to calculate the relationships between nodes; This is a bias term used to calculate the correlation between nodes.

[0164] In this formula:

[0165] ① The expression for the node feature vector is as follows:

[0166] ,

[0167] In the formula: Let i be the position coordinates of the interactive vehicle; Let J be the position coordinates of the interactive vehicle j; Let i be the three-dimensional velocity vector of the interactive vehicle i; Let be the three-dimensional velocity vector of the interactive vehicle j; Let i be the acceleration of the interactive vehicle. Let J be the acceleration of the interactive vehicle j; Let i be the heading angle of the interactive vehicle. Let be the heading angle of the interactive vehicle j.

[0168] ②The expression for spatiotemporal distance decay is as follows:

[0169] ,

[0170] In the formula, For spatiotemporal distance attenuation; Let i be the three-dimensional velocity vector of the interactive vehicle i; Let be the three-dimensional velocity vector of the interactive vehicle j; Let i be the position coordinates of the interactive vehicle; Let J be the position coordinates of the interactive vehicle j; For the observation timestamp of interactive vehicle i; For the observation timestamp of interactive vehicle j;

[0171] This is a hyperparameter of the spatiotemporal distance attenuation coefficient, used to control the attenuation strength of the spatiotemporal distance between interacting vehicles on the interaction weight. It can filter out invalid long-range interference and retain valid short-range interactions. Its value is [value missing]. ;

[0172] This is a hyperparameter representing the spatial weighting in spatiotemporal distance calculation. It indicates the proportion of spatial distance in the overall spatiotemporal distance, emphasizing spatial dominance and adapting to the physical constraints of vehicle motion. Its value is... .

[0173] ③ Effects of relative motion The expression is as follows

[0174] ,

[0175] In the formula, The influence of relative motion; Let i be the three-dimensional velocity vector of the interactive vehicle i; Let be the three-dimensional velocity vector of the interactive vehicle j; This is the attenuation coefficient hyperparameter for the effect of relative motion, used to control the intensity of the effect of velocity difference, and its value is [value missing]. .

[0176] S202, Construct the semantic interaction graph of the bimodal adaptive graph attention network, and output the comprehensive edge weights of the semantic interaction graph. The steps are as follows:

[0177] S202-1, Constructing a semantic interaction graph structure based on a GNN-based intent reasoning network, the expression is:

[0178] ,

[0179] In the formula: A set of nodes for a semantic interaction graph; It is the set of edges of the semantic interaction graph.

[0180] S202-2, Calculate the comprehensive edge weights of the semantic interaction graph based on intent similarity and intent influence, as shown in the following expression:

[0181] ,

[0182] In the formula: Let i be the three-dimensional velocity vector of the interactive vehicle i; Let be the three-dimensional velocity vector of the interactive vehicle j; Let i be the position coordinates of the interactive vehicle; Let J be the position coordinates of the interactive vehicle j; For the observation timestamp of interactive vehicle i; For the observation timestamp of interactive vehicle j; This is a hyperparameter representing the attenuation coefficient of spatial distance on intentional interaction. It controls the attenuation effect of spatial distance on intentional interaction and takes a value of [value missing]. .

[0183] In this formula:

[0184] ①Intent similarity The expression is as follows

[0185] ,

[0186] ②Intentional Influence The expression is as follows

[0187] ,

[0188] In the formula: For the Sigmoid function; Let i be the feature vector of the driving intention of the interactive vehicle i. Let be the feature vector of the driving intention of the interactive vehicle j; To calculate the learnable weight matrix for the influence of intent; This is a bias term used to calculate the degree of influence of the intention.

[0189] S203 dynamically adjusts the weights for fusion of bimodal information from the physical and semantic interaction graphs using a time factor, and employs a time decay gating mechanism to fuse bimodal features, outputting the final node based on a graph attention network structure. The specific steps are as follows:

[0190] S203-1, the dual-modal information fusion weights are calculated using the following expression:

[0191] ,

[0192] In the formula, The fusion gating coefficients for the physical interaction graph and the semantic interaction graph; The learnable weight matrix for the gating mechanism; Use the Sigmoid activation function; The feature representation of node i in the physical interaction graph; The feature representation of node i in the semantic interaction graph; For learnable bias terms of the gating mechanism; For gating mechanisms, the expression is: ; This is the attenuation coefficient, with a value of 0.2; This is the current time step; t A historical time step, relative to the current time step. .

[0193] S203-2, Historical trajectory sequence Input a bidirectional GRU and output an interactive vehicle driving intention feature vector. Semantic associations between nodes are calculated through self-attention, as shown in the following expression.

[0194] ,

[0195] In the formula, For semantic relationships between nodes; This is the learnable weight matrix for the query (Q) in the attention mechanism; Let K be the learnable weight matrix of the key (K) in the attention mechanism; Let i be the feature vector of the driving intention of the interactive vehicle i. Let be the feature vector of the driving intention of the interactive vehicle j; Let be the driving intention feature vector of the interactive vehicle k; n is the index of the neighboring node of node i. Let i be the set of neighbors of node i.

[0196] S203-3 employs a time decay gating mechanism and outputs the final node of the bimodal computation by fusing bimodal features, thus outputting the motion feature vector of the interactive vehicle. .

[0197] The fusion of dual-modal feature expressions is as follows

[0198] ,

[0199] In the formula: The motion feature vector of the interactive vehicle; The fusion gating coefficients for the physical interaction graph and the semantic interaction graph; The feature representation of node i in the physical interaction graph; The feature representation of node i in the semantic interaction graph; This is a gating mechanism.

[0200] The expression for the final node in bimodal computation is as follows:

[0201] ,

[0202] In the formula: For graph attention network Feature representation of nodes in the layer; GAT is a graph attention network structure; This represents the feature representation of node i in the l-th layer of the graph attention network; This is the feature representation of node j in the l-th layer of the graph attention network; This refers to the semantic relationships between nodes.

[0203] As a preferred option, to optimize computational efficiency:

[0204] ① The graph attention network uses 4-head attention, with an accuracy loss of <0.5%;

[0205] ② The intent reasoning network uses depthwise separable convolutions, reducing the number of parameters by 43%;

[0206] ③ Gated fusion module fixed-point quantization (FP16) reduces memory usage by 60%.

[0207] S3. Calculate the probability distribution of driving intentions for the final node output in step S2, and output the probability distribution of the driving intentions of the interactive vehicle, thereby making an accurate judgment on the driving intentions of the interactive vehicle and providing a basis for predicting the driving trajectory of the interactive vehicle.

[0208] The specific steps are as follows:

[0209] A multilayer perceptron is used to perform multilayer nonlinear transformations on the final node, as expressed below.

[0210] ,

[0211] In the formula, This is the output feature vector of the first layer of the multilayer perceptron; This is the output feature vector of the second layer of the multilayer perceptron; ,N Number of intent categories; This is the comprehensive feature vector after dimensional transformation; This refers to the bias term of the second layer of the multilayer perceptron; This refers to the bias term of the third layer of the multilayer perceptron.

[0212] Combining the driving intention features output by the Softmax function, the expression is as follows:

[0213] ,

[0214] In the formula: This is a characteristic of driving intent; This is the output feature vector of the second layer of the multilayer perceptron; is the c-th element in the output feature vector of the second layer of the multilayer perceptron; c is the intention category index (1 for keeping straight, 2 for changing lanes to the left, 3 for changing lanes to the right); N is the total number of intention categories (N=3).

[0215] As a preferred option:

[0216] ① Construct a conditional probability table and obtain trajectory distribution patterns under different intentions through historical data statistics, such as Figure 2 As shown, the probability distribution of driving intention is fused with the input features of the trajectory prediction model.

[0217] ② Construct a sliding window weighted average mechanism to improve the reliability of probability distribution calculation. The expression is:

[0218] ,

[0219] In the formula: The length of the window; The maximum probability value for the i-th frame; The mean is the value within the window; F is the confidence level, with a value of [value missing]. ,when The intent re-detection mechanism is triggered at any time.

[0220] ③ A combination of cross-entropy loss and focus loss is used to jointly optimize the objective and improve the performance of driving intention classification.

[0221] The expression for cross-entropy loss is: ;

[0222] The expression for focus loss is: ;

[0223] Through expression Joint optimization objectives The weighting coefficient has a value of [value]. .

[0224] In the above formula:

[0225] c is the intent category index (1 for stay straight, 2 for change lanes left, 3 for change lanes right); N is the total number of intent categories (N=3). This is a real label; , represents the driving intention feature; p is the probability value of the true label in the intention probability vector predicted by the model;

[0226] The hyperparameter of the spatiotemporal distance attenuation coefficient, This is a hyperparameter representing the attenuation coefficient due to relative motion, used to mitigate class imbalance. Its value is [value missing]. ;

[0227] It acts as a balancing factor, used to balance the importance of different categories and solve the problem of category imbalance. At that time, the model can balance the weights of different categories well;

[0228] It acts as a modulation factor, used to adjust the contribution of easily classified and difficult-to-classify samples to the loss function. At that time, the model performed best in balancing the learning of easy-to-classify and hard-to-classify samples.

[0229] S4, to address the technical problem of feature confusion between the target vehicle and surrounding vehicles in traditional methods, and to achieve the beneficial effect of obtaining more discriminative spatiotemporal features, the motion feature vector of the interactive vehicle output in step S2 is used as the basis for the output. The probability distribution of the interactive vehicle's driving intentions, output from step S3, is used to further extract the differentiated features between the target vehicle and surrounding vehicles, providing more accurate input for trajectory prediction. This step constructs a differentiated spatiotemporal feature extraction architecture, specifically:

[0230] ① Construct a long-cycle attention encoder to capture the global motion trend of the target vehicle over a longer time span;

[0231] ② Construct a gating local feature extraction module to enhance the identification of key frame mutation behavior of surrounding vehicles and accurately capture local and sudden motion changes of surrounding vehicles.

[0232] In this step, "long period" refers to a relatively long time span (e.g., 10 seconds, corresponding to 50 frames of data); "relatively long time" specifically refers to 10 seconds; "surrounding vehicles" refers to other interacting vehicles around the target vehicle, such as vehicles following in adjacent lanes or vehicles traveling in the opposite direction.

[0233] The specific steps are as follows:

[0234] S401, such as Figure 3 As shown, the long-term attention encoder for the target vehicle is constructed as follows: a multi-head self-attention mechanism is used to model the motion dependencies of the target vehicle over long time, and the feature representation is enhanced through global temporal context, as detailed below:

[0235] S401-1, Input the historical trajectory sequence of the target vehicle, as shown in the following expression:

[0236] ,

[0237] In the formula, D is the time step, and D is the feature dimension.

[0238] S401-2, the attention weights are calculated using the following expression:

[0239] ,

[0240] In the formula, , These are the query, key, and value matrices for the attention mechanism. Learnable weights for generating the query matrix, Let be the feature vector of the target vehicle in frame t; For the key vector dimension, multi-head attention is achieved by concatenating 8 single-head outputs; For time step.

[0241] S401-3 outputs the global trend characteristics of the target vehicle using the following expression:

[0242] ,

[0243] In the formula, As a trend feature; Layer normalization is used to accelerate model training convergence and alleviate the gradient vanishing problem. Regularization enables the model to learn more robust and generalizable features; It employs multi-head self-attention to uncover the correlations between different positions in the kinematic parameter sequence of a target vehicle, thereby enhancing the feature extraction effect.

[0244] S402, the surrounding vehicle gate local feature extraction module is constructed, and the output local features are as follows:

[0245] S402-1, Input surrounding vehicle trajectories One-dimensional causal convolution is used to extract local temporal features, expressed as follows:

[0246] ,

[0247] In the formula, The output features of a one-dimensional causal convolution. The trajectory feature sequence of surrounding vehicles;

[0248] S402-2 enhances the characteristic response at the moment of abrupt change by using a gating unit. The gating weight expression is as follows:

[0249] ,

[0250] In the formula, Let be the gating weight for frame t; For the Sigmoid function; The learnable weight matrix for the gating mechanism; The output features of a one-dimensional causal convolution; The difference features between adjacent frames; This is a learnable bias term for the gating mechanism.

[0251] S402-3 outputs the local mutation characteristics of surrounding vehicles using the following expression:

[0252] ,

[0253] In the formula, This represents the local mutation feature vector of the surrounding vehicles; For the first t Frame gating weights; This represents the output feature of a one-dimensional causal convolution.

[0254] S403 adaptively fuses the global trend features of the target vehicle with the local abrupt change features of surrounding vehicles and outputs the spatiotemporal interaction features, expressed as:

[0255] ,

[0256] In the formula, MLP is used to coordinate the feature scales of different modalities to ensure spatiotemporal consistency; Features of spatiotemporal interaction; This represents the local mutation feature vector of the surrounding vehicles; This represents a trend characteristic.

[0257] S5, Construct an intent feature fusion mechanism to integrate the motion features of the interactive vehicle obtained in step S2. The driving intention feature p obtained in step S3 and the spatiotemporal interaction feature obtained in step S4. The system integrates and predicts the future trajectory of interactive vehicles based on multimodal analysis, specifically as follows:

[0258] S501, construct a gated fusion module to dynamically adjust the contribution weights of multimodal features, wherein the feature importance weights are calculated using the following expression:

[0259] ,

[0260] In the formula, These represent the weights of motion, interaction, and intent features, respectively. This is a learnable matrix used to calculate the feature importance weights; Features of spatiotemporal interaction; The motion feature vector of the interactive vehicle; This is a characteristic of driving intention.

[0261] Weighted feature fusion is calculated using the following expression:

[0262] ,

[0263] In the formula, For weighted feature fusion; The weights for the motion features; The weights of the interaction features; The weights of the intent features; Features of spatiotemporal interaction; The motion feature vector of the interactive vehicle; This is a characteristic of driving intention.

[0264] S502, Multimodal Trajectory Prediction: Probabilistic prediction head, using a hybrid density network (MDN) to generate a multimodal future trajectory distribution, where each trajectory corresponds to a possible intent, as expressed below:

[0265] ,

[0266] In the formula, Given input features X, let X be the probability distribution of the future trajectory Y. The mixing coefficient for the intent category K; Let be the mean and covariance of the Gaussian distribution.

[0267] Multimodal trajectory generation: Sample the M most likely trajectories from the predicted distribution as the final output, as shown in the following expression:

[0268] ,

[0269] In the formula, This is the output set of multimodal trajectory prediction; For the predicted M possible trajectories; Let X be the probability distribution of the future trajectory Y given input features X.

[0270] S503 optimizes the interactive vehicle trajectory prediction error and intent classification loss using the following expression, and outputs the multimodal trajectory of the interactive vehicle:

[0271] ,

[0272] In the formula, Loss is the joint loss for trajectory prediction and intent classification; NLL is the negative log-likelihood loss; CE is the cross-entropy loss. , To balance the hyperparameters; Y is the true trajectory label; represents the output set of multimodal trajectory prediction; e represents the true intent label. The predicted intent label.

[0273] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A vehicle driving intention recognition and trajectory prediction method for autonomous driving, characterized by comprising the following steps: S1, a dynamic threshold sliding sampling window is constructed, and its window length is dynamically adjusted according to road condition complexity; the lateral displacement change amount of an interactive vehicle in the dynamic threshold sliding sampling window is extracted, and the driving intention of the interactive vehicle is output; S2, based on the driving intention of the interactive vehicle, a dual-modal adaptive graph attention network is constructed, a bottom-layer physical interaction graph and an upper-layer semantic interaction graph of the dual-modal adaptive graph attention network are set, the physical interaction graph and the semantic interaction graph are fused, and a final node and a motion feature vector of the interactive vehicle are output; wherein, I, a space-time graph structure of the physical interaction graph is constructed through the following expression: The comprehensive edge weight of the physical interaction graph is calculated through the following expression: ， wherein is a global representation of the physical interaction graph; is a set of nodes of the physical interaction graph; is a set of edges of the physical interaction graph; The space-time distance decay is calculated through the following expression: ， wherein, is the integrated edge weight between nodes and nodes ; is the spatio-temporal distance decay; is the relative motion influence; is the Sigmoid function; is the node feature vector of node ; is the node feature vector of node ; is the learnable weight matrix for computing the association between nodes; is the bias term for computing the association between nodes; The relative motion influence is calculated through the following expression: ， In the formula, For spatiotemporal distance attenuation; For interactive vehicles Position coordinates; For interactive vehicles Position coordinates; For interactive vehicles Observation timestamp; For interactive vehicles Observation timestamp; This is a hyperparameter of the spatiotemporal distance attenuation coefficient; This is a hyperparameter representing the spatial weighting in spatiotemporal distance calculation; The comprehensive edge weight of the semantic interaction graph is calculated through the following expression: ， wherein is the relative motion influence; is the three-dimensional velocity vector of the interacting vehicle is the three-dimensional velocity vector of the interacting vehicle is the three-dimensional velocity vector of the interacting vehicle is the three-dimensional velocity vector of the interacting vehicle is the decay coefficient hyperparameter for the relative motion influence; The semantic interaction graph structure is constructed by the following expression: ， wherein: is an overall representation of the semantic interaction graph; is a set of nodes of the semantic interaction graph; is a set of edges of the semantic interaction graph; S3, driving intention probability distribution calculation is performed on the final node, and the probability distribution of the interactive vehicle driving intention feature is output; ， wherein, is a comprehensive edge weight; is an intention similarity; is an intention influence degree; is a position coordinate of the interacting vehicle ; is a position coordinate of the interacting vehicle ; is an observation timestamp of the interacting vehicle ; is an observation timestamp of the interacting vehicle ; is a spatial distance-to-intention interaction decay coefficient hyperparameter; intention similarity is calculated by the following expression: ， intention influence degree is calculated by the following expression: ， wherein: is a Sigmoid function; is a driving intention feature vector of the interacting vehicle is a driving intention feature vector of the interacting vehicle is a driving intention feature vector of the interacting vehicle is a driving intention feature vector of the interacting vehicle is a learnable weight matrix for computing intention influence degree; is a bias term for computing intention influence degree; S4, based on the motion feature vector of the interactive vehicle and the probability distribution of the interactive vehicle driving intention, a long-period attention encoder is constructed to capture the global trend feature of the target vehicle; a gated local feature extraction module is constructed to identify the key frame mutation behavior of the surrounding vehicle and capture the motion change of the surrounding vehicle; and a space-time interaction feature is output; S5, an intention feature fusion mechanism is constructed to integrate the motion feature vector of the interactive vehicle, the interactive vehicle driving intention feature and the space-time interaction feature, and predict the future trajectory of the interactive vehicle. 2.The vehicle driving intention recognition and trajectory prediction method for autonomous driving according to claim 1, characterized by outputting the motion feature vector of the interactive vehicle through the following steps: Wherein, the expression of the dual-modal calculation final node is as follows: The road condition complexity is calculated by the following equation: ， In the formula, is a road condition complexity index; is a lateral acceleration standard deviation of the interacting vehicle; is a lane curvature; is an interacting vehicle density; , , are respectively , , corresponding empirical weight coefficients; The window length of the dynamic threshold sliding sampling window is set by the following rule : Setting for complex road conditions, ; setting for medium complex road conditions, ; Setting For simple road conditions, ; wherein, is the base window length; The lateral displacement change amount of the interacting vehicle is calculated by the formula and the following condition is set to output the driving intention of the interacting vehicle: When the longitudinal acceleration of the target vehicle is greater than 0.5 m / s2 , the driving intention of the interactive vehicle is output as keeping straight; when the longitudinal acceleration of the target vehicle is greater than 0.5 m / s2 , the driving intention of the interactive vehicle is output as left lane changing; when the longitudinal acceleration of the target vehicle is greater than 0.5 m / s2 , the driving intention of the interactive vehicle is output as right lane changing; In the formula, is the lateral displacement change amount of the interactive vehicle; is the lateral position coordinate of the first frame in the dynamic threshold sliding sampling window. is the lateral position coordinate of the first frame in the dynamic threshold sliding sampling window. is the lateral position coordinate of the first frame in the dynamic threshold sliding sampling window.

3. The vehicle driving intention recognition and trajectory prediction method for autonomous driving according to claim 1, characterized in that, 4.The vehicle driving intention recognition and trajectory prediction method for autonomous driving according to claim 1, characterized by performing multi-layer nonlinear transformation on the final node through the following expression: The physical interaction graph and the semantic interaction graph are fused by the following expression: ， wherein, is the fusion gating coefficient of the physical interaction graph and the semantic interaction graph; is the learnable weight matrix of the gating mechanism; is the Sigmoid activation function; is the feature representation of the node in the physical interaction graph; is the feature representation of the node in the semantic interaction graph; is the learnable bias term of the gating mechanism; is the time decay coefficient; The historical trajectory sequence is input into the bidirectional GRU, and an interactive vehicle driving intention feature vector is output. The semantic correlation between nodes is calculated by the following expression: ， In the formula, For semantic relationships between nodes; This represents the learnable weight matrix for the query in the attention mechanism; This represents the learnable weight matrix of the keys in the attention mechanism; For interactive vehicles The feature vector of driving intention; For interactive vehicles The feature vector of driving intention; For interactive vehicles The feature vector of driving intention; For nodes The neighbor node index; For nodes The set of neighbors; The motion feature vector of the interacting vehicle is output by the following expression: ， In the formula, is a motion feature vector of the interactive vehicle; is a fusion gating coefficient of the physical interaction graph and the semantic interaction graph; is a feature representation of a node in the physical interaction graph; is a feature representation of a node in the semantic interaction graph; is a time decay coefficient; The driving intention feature is output through the following expression: ， wherein, is the feature representation of the node in the i-th layer of the graph attention network; is the feature representation of the node in the i-th layer of the graph attention network; is the feature representation of the node in the i-th layer of the graph attention network; is the graph attention network structure; is the feature representation of the node in the i-th layer of the graph attention network; is the feature representation of the node in the i-th layer of the graph attention network; is the feature representation of the node in the i-th layer of the graph attention network; is the feature representation of the node in the i-th layer of the graph attention network; is the feature representation of the node in the i-th layer of the graph attention network; is the feature representation of the node in the i-th layer of the graph attention network; is the semantic association between nodes. 5.The vehicle driving intention recognition and trajectory prediction method for autonomous driving according to claim 1, characterized by constructing the long-period attention encoder of the target vehicle through the following steps: 6.The vehicle driving intention recognition and trajectory prediction method for autonomous driving according to claim 5, characterized by constructing the gated local feature extraction module through the following steps: ， In the formula, is an output feature vector of the first layer of the multilayer perception machine; is an output feature vector of the second layer of the multilayer perception machine; , N is the number of intent categories; is a comprehensive feature vector after dimension transformation; is a bias term of the second layer of the multilayer perception machine; is a bias term of the third layer of the multilayer perception machine; The global trend feature of the target vehicle and the local mutation feature of the surrounding vehicle are adaptively fused to output the space-time interaction feature, and the expression is as follows: ， wherein is a driving intent feature; is an output feature vector of a second layer of a multilayer perceptron; is an element of the output feature vector of the second layer of the multilayer perceptron; is an element of the output feature vector of the second layer of the multilayer perceptron; is an intent class index; is a total number of intent classes. The gated fusion module is constructed to dynamically adjust the contribution weight of the multi-modal feature, wherein the feature importance weight is calculated through the following expression: The weighted feature fusion is calculated through the following expression: , input a historical trajectory sequence of a target vehicle , is a time step, is a feature dimension, the attention weight is calculated by the following expression: ， wherein, , are query, key, value matrices of attention mechanism, respectively, is a learnable weight for generating the query matrix, is a feature vector of the target vehicle at the t-th frame; is the dimension of the key vector; is the time step. The global trend feature of the target vehicle is output by the following expression: ， In the formula, is a global trend feature; is layer normalization; is regularization; is multi-head self-attention. 8.The vehicle driving intention recognition and trajectory prediction method for autonomous driving according to claim 7, characterized by performing multi-modal trajectory prediction distribution of the interactive vehicle through the following expression: ​ , input surrounding vehicle trajectories , is the time step, is the feature dimension, the local temporal feature is extracted by the following expression: ， In the formula, is a one-dimensional causal convolution output feature, is a trajectory feature sequence of surrounding vehicles; The characteristic response at the moment of mutation is calculated by the following expression: ， wherein, is the gating weight for the t-th frame; is a Sigmoid function; is a learnable weight matrix for the gating mechanism; is the output feature of one-dimensional causal convolution; is the adjacent frame difference feature; is a learnable bias term for the gating mechanism; The local abruptness feature of the surrounding vehicle is output by the following expression: ， In the formula, is a local mutation feature vector of the surrounding vehicle; is the gating weight of the t-th frame; is the output feature of one-dimensional causal convolution; ​ ， In the formula, MLP is used to coordinate the feature scales of different modalities, ensuring spatio-temporal consistency. is a spatio-temporal interaction feature; is a local mutation feature vector of surrounding vehicles; is a global trend feature.

7. The vehicle driving intention recognition and trajectory prediction method for autonomous driving according to claim 1, characterized in that, ​ ， wherein, represent weights of motion, interaction, and intention features, respectively; is a learnable matrix for computing feature importance weights; is a spatio-temporal interaction feature; is a motion feature vector of the interacting vehicle; is a driving intention feature; ​ ， wherein, is a weighted feature fusion; is a weight of a motion feature; is a weight of an interaction feature; is a weight of an intention feature; is a spatio-temporal interaction feature; is a motion feature vector of an interacting vehicle; is a driving intention feature. ​ ​ ， wherein is the probability distribution of future trajectory Y given input features X; is the mixture coefficient for intent class is the mixture coefficient for intent class is the mean and covariance of the Gaussian distribution. Sample multiple possible prediction trajectories from the predictive distribution as the final output, expressed as follows: ， wherein is an output set for multi-modal trajectory prediction; is M possible trajectories predicted; is a probability distribution of future trajectories Y given input features X. 9.The vehicle driving intention recognition and trajectory prediction method for autonomous driving according to claim 8, wherein, The interactive vehicle trajectory prediction error and intention classification loss are optimized by the following expression, and the multi-modal trajectory of the interactive vehicle is output: ， wherein is a joint loss for trajectory prediction and intent classification; is a negative log-likelihood loss; is a cross-entropy loss; , is a balancing hyperparameter; is a true trajectory label; is a set of outputs for multi-modal trajectory prediction; is a true intent label; is a predicted intent label.

Images