A method for switching autonomous driving strategies based on vehicle driving style type recognition
By acquiring and identifying real-time motion data of vehicles around autonomous vehicles and dynamically updating autonomous driving strategies, the problem of the inability to update the driving style recognition of surrounding vehicles online in existing technologies is solved, thereby improving the interactive safety and traffic efficiency of autonomous vehicles.
Patent Information
- Application Number
- CN202611320452.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
AI Technical Summary
In existing technologies, the recognition of driving styles of surrounding vehicles cannot be dynamically updated online and lacks linkage with the selection of autonomous driving strategies, making it difficult for autonomous vehicles to balance interaction safety and traffic efficiency.
By acquiring real-time temporal motion data of the target vehicle, extracting behavioral features, and inputting them into a pre-trained vehicle driving style type recognition network for online recognition, the system selects a matching autonomous driving strategy from the strategy library based on the recognition results and continuously updates the strategy during execution to adapt to changes in driving style.
It enables online recognition of the driving styles of surrounding vehicles and dynamic switching of strategies, improving the interactive safety and traffic efficiency of autonomous vehicles in mixed traffic environments.
Smart Images

Figure CN123104693A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous driving technology, and in particular to an autonomous driving strategy switching method based on vehicle driving style type recognition. Background Technology
[0002] In mixed traffic environments, autonomous vehicles need to interact with surrounding vehicles that have different driving styles. Current technologies mostly rely on offline classification or static judgment to identify the driving styles of surrounding vehicles. They cannot dynamically update the identification results based on real-time observation data during driving, and the identification results fail to effectively link with the selection and switching of autonomous driving strategies. This results in autonomous vehicles adopting fixed driving strategies when faced with surrounding vehicles with different driving styles, making it difficult to balance interaction safety and traffic efficiency.
[0003] Therefore, there is an urgent need for a method that can identify the driving style of surrounding vehicles online and dynamically switch autonomous driving strategies accordingly. Summary of the Invention
[0004] (a) Technical problems to be solved In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides an autonomous driving strategy switching method based on vehicle driving style type recognition, which solves the technical problem in the prior art that the recognition of the driving style of surrounding vehicles cannot be dynamically updated online and lacks effective linkage with the selection of autonomous driving strategy, making it difficult for autonomous vehicles to balance interaction safety and traffic efficiency.
[0005] (II) Technical Solution To achieve the above objectives, the main technical solutions adopted by the present invention include: In a first aspect, embodiments of the present invention provide an autonomous driving strategy switching method based on vehicle driving style type recognition, comprising: acquiring in real time the temporal motion data of each target vehicle within the perception range of the current autonomous driving vehicle, and extracting behavioral features of the corresponding target vehicle based on the temporal motion data of each target vehicle; wherein, the behavioral features include speed change rate, yield probability, trajectory smoothness, reaction delay time, and relative distance change trend with the current autonomous driving vehicle; inputting the behavioral features of each target vehicle into a pre-trained vehicle driving style type recognition network for vehicle driving style type recognition, so as to obtain the vehicle driving style type of each target vehicle and the corresponding confidence level; wherein, the vehicle driving style type... The driving styles of each target vehicle are categorized into four types: assertive, courteous, hesitant, rule-based, and abnormal. Based on the driving style type and corresponding confidence level of each target vehicle, a matching autonomous driving strategy is selected from a pre-built strategy library. Each autonomous driving strategy includes strategy action parameters and safety risk threshold parameters corresponding to the vehicle's driving style type. The matched autonomous driving strategy is executed, and the driving style type and corresponding confidence level of each target vehicle are continuously updated during execution. Furthermore, when the driving style type of any target vehicle changes or its corresponding confidence level falls below a preset threshold, the autonomous driving strategy for the current autonomous vehicle is redefined based on the updated driving style type and confidence level of each target vehicle.
[0006] Optionally, the strategy library stores differentiated autonomous driving strategies corresponding to each vehicle's driving style type. Among them, the defensive yielding strategy for aggressive target vehicles includes increasing the minimum safe distance threshold and reducing the tolerance for conflict zone intrusion; the decisive passage strategy for courteous target vehicles includes reducing the minimum safe distance threshold; the patient waiting and gradual probing strategy for hesitant target vehicles includes increasing the yielding decision time window and increasing the tolerance for conflict zone intrusion; the compliance game strategy for rule-abiding target vehicles includes maintaining the preset standard safety risk threshold parameters; and the safety avoidance and warning strategy for abnormal target vehicles includes increasing the minimum safe distance threshold and reducing the maximum permissible acceleration.
[0007] Optionally, each autonomous driving strategy in the pre-built strategy library is trained separately in a simulation environment containing vehicles with corresponding driving style types through multi-agent reinforcement learning; after deployment in real vehicles, the parameters of the matched autonomous driving strategy are fine-tuned online based on the driving behavior data of the current autonomous driving vehicle and the temporal motion data of each target vehicle collected during actual driving, so as to continuously optimize the autonomous driving strategy.
[0008] Optionally, the minimum safe distance threshold in the standard safety risk threshold parameters is between the minimum safe distance threshold set for a defensive yielding strategy for aggressive target vehicles and the minimum safe distance threshold set for a decisive passage strategy for yielding target vehicles, and the maximum permissible acceleration in the standard safety risk threshold parameters does not exceed a preset regulatory limit.
[0009] Optionally, when there are multiple target vehicles, the safety risk level of each target vehicle is determined based on its driving style type and its motion state relative to the current autonomous vehicle, and the processing priority is determined according to the safety risk level; wherein, the motion state includes the distance and approach speed of the target vehicle relative to the current autonomous vehicle; the corresponding autonomous driving strategy is selected first for the target vehicle with the highest safety risk level.
[0010] Optionally, when the confidence level of any target vehicle is lower than a preset threshold and the autonomous driving strategy needs to be redefined, the vehicle driving style type identification result of the target vehicle is considered invalid and is not used as the basis for strategy selection. A preset conservative autonomous driving strategy is then determined for the current autonomous vehicle. The conservative autonomous driving strategy uses preset conservative safety risk threshold parameters and conservative strategy action parameters. The conservative safety risk threshold parameters include: the minimum safe distance threshold is not less than the maximum value among the minimum safe distance thresholds contained in each autonomous driving strategy in the strategy library; the maximum permissible acceleration is not greater than the minimum value among the maximum permissible accelerations contained in each autonomous driving strategy in the strategy library; and the yield decision time window is not less than the preset safe conservative time window value.
[0011] Optionally, the conditions for triggering the redetering of the current autonomous driving strategy of the autonomous vehicle may include at least one of the following: the target vehicle enters or leaves the perception range of the current autonomous vehicle; the driving state of the current autonomous vehicle changes abruptly; wherein, the abrupt change in driving state includes the current autonomous vehicle's speed change rate exceeding a preset change threshold or the heading angle change exceeding a preset angle threshold.
[0012] Optionally, while the target vehicle is within the perception range of the current autonomous vehicle, behavioral feature data of the target vehicle is continuously collected and stored to form historical behavioral feature data of the target vehicle. The confidence level of the target vehicle is corrected based on the historical behavioral feature data. When the historical behavioral feature data is inconsistent with the currently identified vehicle driving style type, the confidence level of the target vehicle is reduced. When the confidence level is lower than a preset threshold, the vehicle driving style type identification result of the target vehicle is considered invalid.
[0013] Optionally, when multiple target vehicles exist simultaneously within the perception range of the current autonomous vehicle and correspond to different driving style types, the current autonomous vehicle executes a comprehensive autonomous driving strategy. The safety risk threshold parameter of the comprehensive autonomous driving strategy is the most stringent safety risk threshold parameter among the autonomous driving strategies corresponding to each target vehicle. The most stringent safety risk threshold parameter includes the maximum minimum safe distance threshold, the minimum maximum permissible acceleration, and the maximum yield decision time window.
[0014] Optionally, when the driving style of any target vehicle is identified as anomalous, the current autonomous vehicle, while executing safety avoidance and warning strategies, also triggers external warning actions, including issuing warnings to surrounding traffic participants through lights and / or sounds.
[0015] In a second aspect, embodiments of the present invention provide a new energy vehicle, including a control device configured to execute an autonomous driving strategy switching method based on vehicle driving style type recognition as described in any of the first aspects.
[0016] (III) Beneficial Effects The beneficial effects of this invention are: This invention proposes an autonomous driving strategy switching method based on vehicle driving style type recognition. This method acquires real-time temporal motion data of each target vehicle within the perception range of the current autonomous driving vehicle and extracts behavioral features. These behavioral features are then input into a pre-trained vehicle driving style type recognition network for online recognition, obtaining the vehicle driving style type and confidence level of each target vehicle. Based on the identified driving style type and confidence level, a matching autonomous driving strategy is selected from a pre-built strategy library. This strategy includes strategy action parameters and safety risk threshold parameters corresponding to the vehicle driving style type. Furthermore, the type and confidence level of each target vehicle are continuously updated during strategy execution. When the type of any target vehicle changes or its confidence level falls below a preset threshold, the autonomous driving strategy is re-determined based on the updated information. This achieves online recognition of the driving styles of surrounding vehicles and dynamic strategy switching, solving the problems of existing technologies where recognition results cannot be dynamically updated and lack linkage with strategy selection. Attached Figure Description
[0017] Figure 1 A flowchart of an autonomous driving strategy switching method based on vehicle driving style type recognition provided in an embodiment of this application is shown; Figure 2 A schematic diagram of the structure of a vehicle driving style type recognition network provided in an embodiment of this application is shown. Detailed Implementation
[0018] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0019] This invention proposes an autonomous driving strategy switching method based on vehicle driving style type recognition. This method acquires real-time temporal motion data of each target vehicle within the perception range of the current autonomous driving vehicle and extracts behavioral features. These behavioral features are then input into a pre-trained vehicle driving style type recognition network for online identification, obtaining the vehicle driving style type and confidence level of each target vehicle. Based on the identified driving style type and confidence level, a matching autonomous driving strategy is selected from a pre-built strategy library. This strategy includes strategy action parameters and safety risk threshold parameters corresponding to the vehicle driving style type. Furthermore, the type and confidence level of each target vehicle are continuously updated during strategy execution. When the type of any target vehicle changes or its confidence level falls below a preset threshold, the autonomous driving strategy is re-determined based on the updated information. This achieves online identification of the driving styles of surrounding vehicles and dynamic strategy switching, solving the problems of existing technologies where the identification results cannot be dynamically updated and lack linkage with strategy selection.
[0020] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.
[0021] Please see Figure 1 , Figure 1 A flowchart illustrating an autonomous driving strategy switching method based on vehicle driving style type recognition provided in this application embodiment is shown. Specifically, this method can be executed by the control device in the current autonomous vehicle; the method includes: Step S110: Real-time acquisition of temporal motion data of each target vehicle within the perception range of the current autonomous vehicle, and extraction of behavioral features of the corresponding target vehicle based on the temporal motion data of each target vehicle. These behavioral features include speed change rate, yield probability, trajectory smoothness, reaction delay time, and the trend of relative distance change with the current autonomous vehicle.
[0022] Specifically, current autonomous vehicles use their perception modules (such as LiDAR, millimeter-wave radar, cameras, and multi-sensor fusion systems) to collect real-time information about their surrounding environment, acquiring raw temporal motion data such as position, speed, acceleration, and heading angle of each target vehicle within the perception range of the current autonomous vehicle. This temporal motion data is a continuous observation sequence arranged in chronological order, with each frame of observation data bearing a timestamp to reflect the changes in the motion state of the target vehicle over time.
[0023] Furthermore, based on the temporal motion data of each target vehicle, the current autonomous vehicle extracts behavioral features for each target vehicle separately. These behavioral features are used to characterize the driving behavior style of the target vehicle and specifically include the following features: Rate of change of speed (acceleration distribution): Based on the speed sequence of the target vehicle within a continuous time window, calculate its acceleration values and distribution characteristics, such as the mean, variance, and frequency of rapid acceleration or deceleration. Target vehicles with a larger rate of change of speed usually exhibit a more aggressive driving tendency, while target vehicles with a stable rate of change of speed may be more inclined to drive cautiously or in a regular manner.
[0024] Yield probability: This is defined as the conditional probability that, in a scenario where there is a potential conflict between the current autonomous vehicle and the target vehicle (e.g., both vehicles approach an intersection simultaneously or merge into the same lane), the target vehicle will decelerate and yield in a number of similar interactions prior to the current moment. Specifically, it can be estimated by statistically analyzing the frequency of the target vehicle's deceleration behavior when the approach time in the conflict zone is less than a preset threshold. A high yield probability indicates that the target vehicle is more inclined to yield, while a low yield probability may indicate a more aggressive or preemptive tendency.
[0025] Trajectory smoothness: This measures the continuity and stability of a target vehicle's trajectory. It can be calculated by analyzing the time series of the target vehicle's lateral position to determine the frequency and amplitude of lateral swaying. Frequent and large lateral swaying indicates poor trajectory smoothness, potentially reflecting driver hesitation or unstable handling; conversely, good trajectory smoothness usually corresponds to a regular or decisive driving style.
[0026] Reaction delay time: This refers to the time delay required for the target vehicle to make an observable response to the actions of the current autonomous vehicle (such as deceleration, lane change, or activation of turn signals). This time delay can be determined by comparing the time difference between the moment the current autonomous vehicle's action occurs and the moment when the target vehicle's behavior changes significantly (such as a change in speed or heading angle exceeding a preset threshold). A short reaction delay time usually indicates that the target vehicle driver is attentive and makes decisive decisions, while a long reaction delay time may correspond to a hesitant or inattentive driving style.
[0027] The trend of relative distance change between the target vehicle and the current autonomous vehicle: This feature describes how the relative distance between the target vehicle and the current autonomous vehicle changes over time, such as whether the relative distance is continuously decreasing, continuously increasing, or remaining constant. Combined with relative speed information, it can be determined whether the target vehicle is approaching or moving away from the current autonomous vehicle, thus reflecting its interaction intention.
[0028] All of the above behavioral features are extracted online based on real-time acquired temporal motion data and are continuously updated as new observation data arrives, providing dynamic and real-time input information for subsequent vehicle driving style type recognition.
[0029] Step S120: Input the behavioral features of each target vehicle into the pre-trained vehicle driving style type recognition network for vehicle driving style type recognition, so as to obtain the vehicle driving style type and corresponding confidence level of each target vehicle. Among them, the vehicle driving style types include aggressive, courteous, hesitant, rule-based, and abnormal.
[0030] It should be understood that the specific network and its structure of the vehicle driving style type recognition network can be set according to actual needs, and the embodiments of this application are not limited thereto.
[0031] Optionally, such as Figure 2 As shown, the vehicle driving style type recognition network includes four layers: an individual behavior temporal encoding layer, a scene-level interaction graph attention encoding layer, a learnable style prototype library and instance-level matching layer, and a confidence calibration and uncertainty estimation layer. The individual behavior temporal coding layer's output is connected to the scene-level interaction graph attention coding layer's input. Specifically, the latent code vector of each target vehicle's individual behavior and the current state vector of the autonomous vehicle serve as input to the scene-level interaction graph attention coding layer. The scene-level interaction graph attention coding layer's output is connected to the learnable style prototype library and instance-level matching layer's input. Specifically, the enhanced representation vector of each target vehicle is input to the learnable style prototype library and instance-level matching layer. The probability distribution output by the learnable style prototype library and instance-level matching layer is used to determine the vehicle driving style type label and also serves as the processing object for the confidence calibration and uncertainty estimation layer. The confidence calibration and uncertainty estimation layer introduces the Monte Carlo dropout method during the forward propagation process of the aforementioned individual behavior temporal coding layer and scene-level interaction graph attention coding layer. It performs multiple random forward propagations on the same input to obtain multiple sets of probability distributions, statistically calibrates these probability distributions, and finally outputs the calibrated confidence score, paired with the vehicle driving style type label corresponding to the target vehicle.
[0032] The processing procedure of the temporal coding layer for this individual behavior is as follows: Given behavioral characteristics including rate of change of speed, probability of yielding, trajectory smoothness, reaction delay time, and the trend of relative distance changes with the current autonomous vehicle, for the first vehicle within the perception range... i For each target vehicle, its behavioral feature sequences are first aligned along the time dimension to construct a multivariate temporal feature matrix for that target vehicle. X i ∈R (T×5) ,in, i Let i be the sequence number of the target vehicle, i = 1, 2, ..., N, where N is the total number of target vehicles within the perception range; T represents the length of the time window, i.e., the number of time steps contained in the selected behavioral feature sequence; 5 represents the number of channels for the five behavioral features, i.e., speed change rate, yield probability, trajectory smoothness, reaction delay time, and relative distance change trend each occupy one channel; R represents the set of real numbers. X i The input is fed into a shared-weight temporal feature extractor, which consists of a multi-head self-attention mechanism cascaded with a gated recurrent unit, and outputs the first... i Individual behavioral latent code vector of a target vehicle h i ∈R d Where d is the dimension of the latent code vector; the above processing is performed in parallel on N target vehicles within the perception range to obtain the individual behavior latent code set { h 1, h 2, ..., h N Each hidden code uniquely corresponds to a target vehicle.
[0033] Furthermore, the processing procedure of the scene-level interaction graph attention encoding layer is as follows: Using the current state vector of the autonomous vehicle h0 and the set of latent codes for individual behavior { h 1, h 2, ..., h N Together, they form N+1 graph nodes, constructing a dynamic interaction graph for the scene; among them, h0 This represents the current state vector of the autonomous vehicle, with dimensions equal to... h i The same. This vector is encoded from the motion state information of the current autonomous vehicle, which includes at least velocity, acceleration, heading angle, yaw rate, and longitudinal and lateral positions. These state variables can be normalized first, and then mapped to d-dimensional space through a linear mapping layer to form... h0 ,and h0It can also integrate the driving intent or decision context information of the current autonomous vehicles, so that it can participate in message passing and interaction situation modeling as a self-vehicle node in the scene dynamic interaction graph, together with each target vehicle node. Furthermore, the edge weights between graph nodes are dynamically calculated based on the spatial distance between vehicles, relative speed, and lane topology. A multi-head graph attention network is used to perform L-layer message passing on the scene dynamic interaction graph, where L is the preset number of graph attention network layers, and L≥2. In each layer, each target vehicle node aggregates information from all other vehicle nodes in its neighborhood (including self-vehicle nodes and other target vehicle nodes) through an attention mechanism, updating its own node features. After L-layer message passing, the node outputs corresponding to N target vehicles are extracted to obtain the scene perception enhancement representation vector set {z1, z2, ..., z...}. N}, where z i For the first i Each target vehicle incorporates an enhanced representation vector derived from its surrounding interactive situation, and z i With the i Each target vehicle maintains a one-to-one correspondence.
[0034] In the scene-level interaction graph attention encoding process provided in the above embodiments of this application, the edge weights between graph nodes in the scene dynamic interaction graph are dynamically calculated based on the spatial distance between vehicles, relative speed, and lane topology. The advantage of this method is that the calculation is direct and the physical meaning is clear. However, those skilled in the art will understand that vehicles that are spatially close do not necessarily have related interaction intentions. For example, in adjacent lanes, although a straight-going vehicle and a turning vehicle are spatially close, their driving trajectories and intentions do not substantially conflict. If they are still given high graph attention edge weights, it will cause unnecessary mutual influence between the node features of the two vehicles during message transmission, thereby reducing the purity of the enhanced representation vectors of each target vehicle and affecting the accuracy of subsequent vehicle driving style type recognition.
[0035] Therefore, in a preferred embodiment of this application, the edge weights between graph nodes in the scene dynamic interaction graph also include intention. Figure 1 Contagious factors. Intention. Figure 1The consistency factor is calculated based on the degree of consistency between the current decision intention of the autonomous vehicle and the predicted intention of the target vehicle. Specifically, the current decision intention of the autonomous vehicle is determined by the current expected behavior output by its decision planning module, such as a straight-ahead intention, a left-turn intention, a right-turn intention, a lane-changing intention, or a yielding intention. The predicted intention of the target vehicle is obtained through a preset intention inference model. The intention inference model outputs the distribution of the target vehicle's predicted intention within a preset time window based on the target vehicle's driving style type, current motion state (position, speed, heading angle, yaw rate, etc.), and driving environment information (lane lines, traffic signs, intersection topology, etc.).
[0036] Furthermore, when the current decision-making intent of the autonomous vehicle differs from the predicted intent of the target vehicle... Figure 1 When the intention is to proceed straight through an intersection (e.g., the current intention of the autonomous vehicle is to proceed straight through the intersection, and the predicted intention of the target vehicle is also to proceed straight through the intersection), the intention... Figure 1 The consistency factor increases the edge weights, enhancing the information transmission strength between the target vehicle node and the current autonomous vehicle node; when the decision intention of the current autonomous vehicle conflicts with the predicted intention of the target vehicle (e.g., the current autonomous vehicle's intention is to turn left, while the target vehicle's predicted intention is to go straight and overtake), the decision-making factor increases the edge weights, strengthening the information transmission between the target vehicle node and the current autonomous vehicle node; when the current autonomous vehicle's decision intention conflicts with the predicted intention of the target vehicle (e.g., the current autonomous vehicle's intention is to turn left, while the target vehicle's predicted intention is to go straight and overtake), the decision-making factor increases the edge weights, strengthening the information transmission strength between the target vehicle node and the current autonomous vehicle node; Figure 1 The consistency factor reduces edge weights, weakening the information transmission strength between the target vehicle node and the current autonomous vehicle node, thus preventing the interaction information of vehicles with conflicting intentions from excessively interfering with the representation of the current autonomous vehicle node. Through this method, the graph attention network will focus its attention more on target vehicles relevant at the game-theoretic level during message transmission, rather than just vehicles physically close at the game-theoretic level.
[0037] Furthermore, in another preferred embodiment of this application, it is intended that... Figure 1 The consistency factor is quantified as follows: the degree of consistency between the current decision intention of the autonomous vehicle and the predicted intention of the target vehicle is quantified using a coordination game model in game theory. Specifically, the intention options of the current autonomous vehicle and the target vehicle are constructed as a game strategy space. Assume the set of possible intentions for the current autonomous vehicle is... = {a1, a2, ..., a m}, where a1, a2, ..., a m Let m represent the possible intent options that the autonomous vehicle may take, and m be the total number of intent options; the set of possible intents for the target vehicle is... = {b1, b2, ..., b n}, where b1, b2, ..., b nLet R represent the possible intention options of the target vehicle, where n is the total number of intention options for the target vehicle. The payoff matrix R ∈ R for each strategy combination adopted by both parties is given. m×n Based on this, the elements in the payoff matrix Indicates the current intention of the autonomous vehicle. Target vehicle's intention The payoff matrix represents the combined value of the game payoffs for both parties, where i and j are indices of the intended options. The payoff matrix can be pre-set based on traffic rules, right-of-way priorities, and collision risk assessments. For example, in a no-signal-light intersection scenario, the risk is highest and the payoff is lowest when the current autonomous vehicle proceeds straight and the target vehicle proceeds straight simultaneously; conversely, the risk is lowest and the payoff is highest when the current autonomous vehicle yields and the target vehicle proceeds.
[0038] Based on this, the Nash equilibrium distance is calculated for the strategy combination of the current autonomous vehicle's decision intention and the target vehicle's predicted intention. Specifically, the Nash equilibrium distance is the normalized distance between the payoff of the current strategy combination and the payoff of the optimal Nash equilibrium strategy combination among all possible strategy combinations, denoted as [missing information]. The smaller the Nash equilibrium distance, the stronger the intention of both sides. Figure 1 The higher the consistency, the more severe the conflict of intentions between the two parties. The Nash equilibrium distance is converted into an intentional distance using a pre-defined monotonically decreasing mapping function. Figure 1 Infectious factors, such as ,in That is, meaning Figure 1 Pathogenic factors, The preset proportional coefficient is used to control the Nash equilibrium distance relative to the mean. Figure 1 The decay rate of the causative agent; Let be the distance to the Nash equilibrium mentioned above; exp represents the natural exponential function. Figure 1 The consistency factor ranges from (0,1), with a larger value indicating greater consistency between the current autonomous vehicle and the target vehicle. Figure 1 The higher the consistency.
[0039] Ultimately, the edge weight between the i-th node and the j-th node in the scene dynamic interaction graph is the sum of the physical feature weight and the intention weight. Figure 1 The combination of virulence factors, i.e. .in, This represents the edge weight between the i-th node and the j-th node in the scene dynamic interaction graph. The nodes include the current autonomous vehicle node and each target vehicle node, and i and j are node indices. This represents the weights of physical features calculated based on spatial distance, relative speed, and lane topology. This indicates the relationship between the i-th node and the j-th node. Figure 1 Positive factors; The preset weight fusion coefficient, ranging from 0 to 1, is used to adjust the weight of physical features and intention. Figure 1 The relative proportion of consistency factors in edge weights. When When the edge weights are large, they place more emphasis on physical spatial relationships; when When the edge weights are smaller, the edge weights are more focused on the game's intentions. Figure 1 To the point of being responsive.
[0040] After obtaining the edge weights between nodes in the scene dynamic interaction graph, these edge weights are directly used in the message passing process of the multi-head graph attention network. In each layer of message passing, each target vehicle node aggregates information from all other vehicle nodes in its neighborhood (including the current autonomous vehicle node and other target vehicle nodes) through the attention mechanism based on the edge weights between it and other nodes, and updates its own node features. When the edge weight between a target vehicle node and the current autonomous vehicle node is large, the feature information of that target vehicle is given a higher attention weight during the aggregation process, and its contribution to the feature update of the current node is also greater; when the edge weight between the two is small, the feature information of that target vehicle is suppressed during the aggregation process. After multiple layers of message passing, the output corresponding to each target vehicle node is extracted from the scene dynamic interaction graph, which is the scene perception enhancement representation vector. These enhancement representation vectors are then fed into the learnable style prototype library and instance-level matching layer for type discrimination.
[0041] Therefore, it can be seen that... Figure 1 The consistency factor is not an independent computational component, but is embedded as part of the edge weight calculation in the scene-level interaction graph attention encoding layer. By adjusting the edge weights, it directly participates in the message passing of the multi-head graph attention network, ultimately affecting the output of the enhanced representation vectors of each target vehicle, and indirectly affecting the output of vehicle driving style type labels and confidence through subsequent layers.
[0042] Therefore, through the above method, graph attention networks can distinguish between two different interaction scenarios during message transmission: spatial proximity but game-independent and spatial proximity and game-related. This allows the enhanced representation vectors of each target vehicle to more accurately reflect the real game relationship between it and the current autonomous vehicle, thereby improving the accuracy of subsequent vehicle driving style type recognition and autonomous driving strategy selection.
[0043] Furthermore, the processing procedure for this learnable style prototype library and the instance-level matching layer is as follows: For the i Enhanced characterization of the target vehicle z i The projection vector is obtained by projecting it onto the style prototype embedding space through a trainable bilinear mapping matrix W. iWhere W is a trainable bilinear mapping matrix used to map the augmented representation vectors to a space with the same dimensions as the style prototype vectors. The style prototype embedding space is a learnable style prototype embedding matrix P∈R constructed during the network training phase. (5×d) Define a matrix where P is a learnable style prototype embedding matrix with 5 rows and d columns; the vector in the k-th row is p. k Let k be the learnable center representation in the feature space corresponding to the k-th of the five vehicle driving style types, k=1,2,...,5, corresponding to assertive, courteous, hesitant, rule-based, and aberrant driving styles, respectively. Calculate... i With each prototype vector p in the style prototype embedding matrix P k The cosine similarity is used to obtain the first... i The similarity vectors s of the target vehicles i =[s i,1 , s i,2 , ..., s i,5 ], where s i,k Indicates the first i Projection vectors of the target vehicles i With the k-th style prototype vector p k The cosine similarity between them. The similarity vector s i Input to the Softmax function, output the first... i The probability distribution q of each target vehicle for each driving style type i =[q i,1 , q i,2 , ..., q i,5 ], where q i,k Indicates the first i The predicted probability of a target vehicle belonging to the k-th driving style type is given, and the sum of the predicted probabilities of the five driving style types is equal to 1. The type with the highest probability is then output as the driving style type of the target vehicle.
[0044] Furthermore, the process for confidence level calibration and uncertainty estimation layer is as follows: To quantify the uncertainty of the recognition results, Monte Carlo dropout is introduced during the forward propagation process in both the training and inference phases of the temporal feature extractor and the graph attention network. This involves randomly discarding a portion of neurons during each forward propagation. For input data of the same target vehicle at the same time, M random forward propagations are performed, where M is a preset number of random forward propagations, and M ≥ 10, resulting in M probability distributions. , , ..., },in The m-th random forward propagation yields the first... i Given the probability distributions of M target vehicles, m = 1, 2, ..., M, calculate the mean of each of the M probability distributions for each category. and variance ,in Let be the mean vector of the M probability distributions of the i-th target vehicle across the five style categories. Let M be the variance vector of the probability distribution of the i-th target vehicle across the five style categories; with mean... The highest probability value is used as the base confidence level, while the sum of variances across the category dimension is used as a measure of identification uncertainty. A penalized calibration is applied to the base confidence level based on the uncertainty, using the following calibration formula: ; In the formula, This represents the confidence level of the i-th target vehicle after calibration; For the first i The average probability vector of a target vehicle for each vehicle driving style type contains five elements, which correspond to the average probability of the target vehicle belonging to the assertive, courteous, hesitant, rule-abiding, and abnormal types, respectively. This represents the maximum element value in the five-dimensional average probability vector, which is the largest value among the average probabilities corresponding to the five vehicle driving style types. This represents the preset penalty coefficient, used to control the intensity of the penalty imposed on the confidence level by uncertainty; This represents the prediction variance of the i-th target vehicle in the k-th style category. The final output is the calibrated confidence score of the i-th target vehicle. It is output in pairs with the vehicle driving style type label mentioned above.
[0045] Through the above process, the identification network first extracts the individual behavior latent code from the behavioral feature sequence of each target vehicle, then integrates the multi-vehicle interaction situation through scene dynamic interaction graph, and then realizes type discrimination based on style prototype matching. Finally, it outputs a reliable confidence score through uncertainty calibration, providing accurate type labels and confidence score basis for the selection of autonomous driving strategies in the subsequent strategy library.
[0046] It should be noted that the above network structure is only a preferred embodiment of the vehicle driving style type recognition network in this invention, and is not the only limitation on the specific architecture of the recognition network. Provided that the input is the behavioral characteristics of each target vehicle and the output is the vehicle driving style type and corresponding confidence level, this vehicle driving style type recognition network can also be implemented using other existing sequence modeling or classification network architectures, such as recurrent neural networks based on Long Short-Term Memory (LSTM) or Gated Recurrent Units (GRU), temporal convolutional networks based on one-dimensional convolution, temporal classification networks based on Transformer encoders, or any combination of the above networks. Any method that can achieve temporal modeling of the behavioral characteristics of each target vehicle and output the vehicle driving style type and confidence level can be used as a specific implementation of the vehicle driving style type recognition network.
[0047] To facilitate understanding of step S120, a specific embodiment will be described below.
[0048] Specifically, the behavioral features of each target vehicle extracted in step S110 are input into a pre-trained vehicle driving style type recognition network for vehicle driving style type recognition, so as to obtain the vehicle driving style type and corresponding confidence level of each target vehicle. The pre-trained vehicle driving style type recognition network refers to a neural network model that has been trained using a large amount of sample data labeled with vehicle driving style types. It can automatically output a vehicle driving style type label and a confidence level representing the credibility of that type based on the input behavioral features.
[0049] For each target vehicle within the perception range of the current autonomous vehicle, the identification process is as follows: First, the behavioral feature sequence of the i-th target vehicle is aligned along the time dimension to construct a multivariate temporal feature matrix for that target vehicle. The number of rows in this matrix corresponds to the length of the time window, and the number of columns represents the five behavioral features. This multivariate temporal feature matrix is then input into a shared-weight temporal feature extractor in the recognition network. This extractor consists of a multi-head self-attention mechanism cascaded with a gated recurrent unit, and outputs the individual behavioral latent code vector of the i-th target vehicle. The above operation is performed on all N target vehicles within the perception range to obtain the individual behavioral latent code vector for each target vehicle.
[0050] Subsequently, a dynamic interaction graph of the scene is constructed using the current state vector of the autonomous vehicle and the individual behavioral latent code vectors of each target vehicle as nodes. The edge weights between nodes in the graph are dynamically calculated based on the spatial distance between vehicles, relative speed, and lane topology. A multi-head graph attention network is used to perform multi-layer message passing on the dynamic interaction graph of the scene, enabling each target vehicle node to aggregate information from other vehicle nodes in its neighborhood and update its own node features. After multi-layer message passing, the node outputs corresponding to each target vehicle are extracted to obtain an enhanced representation vector for each target vehicle that incorporates the surrounding interaction situation.
[0051] Next, for each target vehicle's enhanced representation vector, it is projected into a style prototype embedding space, which pre-contains learnable style prototype vectors corresponding to five different vehicle driving style types. The cosine similarity between the projected vector and each style prototype vector is calculated to obtain a similarity vector. This similarity vector is then input into a Softmax function to obtain the probability distribution of the target vehicle belonging to each driving style type, and the type with the highest probability is output as the vehicle driving style type label for that target vehicle. Simultaneously, to measure the uncertainty of the recognition results, a Monte Carlo dropout method is introduced during the forward propagation process of the temporal feature extractor and graph attention network. Multiple random forward propagations are performed on the same input to obtain multiple sets of probability distributions. The mean and variance of these probability distributions are calculated, and the highest probability among the means is used as the base confidence score. The base confidence score is then penalized and calibrated based on the variance, finally outputting the calibrated confidence score of the target vehicle, paired with its corresponding vehicle driving style type label.
[0052] Through the above process, the recognition network completes the online identification of the driving style type of each target vehicle, providing type labels and confidence basis for the subsequent selection of matching autonomous driving strategies from the strategy library.
[0053] Step S130: Based on the driving style type and corresponding confidence level of each target vehicle, select a matching autonomous driving strategy from a pre-built strategy library. The autonomous driving strategy includes strategy action parameters and safety risk threshold parameters corresponding to the vehicle's driving style type.
[0054] Specifically, a strategy library can be constructed first. This library stores differentiated autonomous driving strategies corresponding one-to-one with each vehicle's driving style type, including aggressive, courteous, hesitant, rule-based, and abnormal driving styles. Each autonomous driving strategy includes strategy action parameters and safety risk threshold parameters. The strategy action parameters determine the specific driving actions of the current autonomous vehicle, such as target speed, acceleration, following distance, lane change timing, and yielding or proceeding decision tendency. The safety risk threshold parameters constrain the boundaries of the strategy actions, including the minimum safe distance threshold, maximum permissible acceleration, conflict zone intrusion tolerance, and yielding decision time window. The minimum safe distance threshold refers to the minimum distance that the current autonomous vehicle is allowed to maintain with the target vehicle. If the distance is lower than this, deceleration or avoidance measures must be taken. The maximum permissible acceleration refers to the maximum acceleration value that the current autonomous vehicle is allowed to reach when executing the strategy action. The conflict zone intrusion tolerance refers to the degree to which the current autonomous vehicle allows itself to enter the area where there is a potential conflict with the target vehicle. The lower the tolerance, the more conservative it is. The yield decision time window refers to the longest permissible time that the current autonomous vehicle must make a yield or pass decision after recognizing a conflict. The larger the window, the more relaxed the decision is.
[0055] For different vehicle driving styles, the corresponding autonomous driving policies in the policy library have different safety risk threshold parameter settings. Specifically, for aggressive target vehicles, the policy library configures a defensive yielding strategy. This strategy increases the minimum safe distance threshold and reduces the conflict zone intrusion tolerance, requiring the current autonomous vehicle to maintain a greater safe distance from the aggressive target vehicle and avoid entering potential conflict zones as much as possible. For example, under default conditions, the minimum safe distance threshold is 5 meters. When an aggressive target vehicle is identified, this strategy increases the minimum safe distance threshold to 8 meters, meaning the current autonomous vehicle must maintain a distance of at least 8 meters from the aggressive target vehicle. Simultaneously, the conflict zone intrusion tolerance is adjusted from the default "allow entry into conflict zones" to "disallow active entry into conflict zones," prioritizing slowing down and yielding even if the vehicle has the right-of-way, avoiding direct competition for road rights with the aggressive target vehicle.
[0056] Furthermore, for yielding vehicles, the policy library configures a decisive passage strategy. This strategy lowers the minimum safe distance threshold. When a yielding vehicle shows a tendency to give way, the current autonomous vehicle can appropriately reduce the safe distance to improve traffic efficiency, provided it is safe to do so. For example, the default minimum safe distance threshold is 5 meters. When a yielding vehicle is identified and it has already shown signs of slowing down to give way (e.g., a yielding probability exceeding 0.7), the strategy lowers the minimum safe distance threshold to 3 meters. This allows the current autonomous vehicle to quickly pass through the conflict zone with a smaller distance after the yielding vehicle has given way, avoiding excessive conservatism that could reduce traffic efficiency.
[0057] Furthermore, for hesitant target vehicles, the strategy library is configured with a patient waiting and gradual probing strategy. This strategy increases the yield decision time window and improves the conflict zone intrusion tolerance, giving hesitant target vehicles more response time and gradually probing their intentions to avoid misjudgment due to repeated decisions by the other party. For example, the default yield decision time window is 2 seconds. When a hesitant target vehicle is identified, this strategy increases the yield decision time window to 4 seconds, giving the hesitant target vehicle more time to react and make decisions. At the same time, it improves the conflict zone intrusion tolerance, allowing the current autonomous vehicle to gradually approach the conflict zone at a low speed, and judge the final intention of the hesitant target vehicle through small probing actions (such as slowly moving forward, slightly decelerating and observing the other party's reaction), avoiding dangerous actions such as sudden braking or sudden lane changes due to repeated decisions by the other party.
[0058] Furthermore, for rule-based target vehicles, the strategy library configures compliant game-theoretic strategies. These strategies maintain preset standard safety risk threshold parameters, executing driving behavior with conventional safety parameters that comply with traffic regulations. For example, the minimum safe distance threshold remains at the default 5 meters, the maximum permissible acceleration remains at the default 2 meters per square second, and the yield decision time window remains at the default 2 seconds. Currently, the autonomous vehicle is driving normally according to the traffic rules stipulated by traffic regulations, without tightening or loosening safety parameters based on the type of other vehicle.
[0059] Furthermore, for anomalous target vehicles, the policy library configures safety avoidance and warning strategies. These strategies increase the minimum safe distance threshold and decrease the maximum permissible acceleration. By maintaining a greater safe distance and limiting the vehicle's own acceleration, the risk of collision with unpredictable target vehicles is reduced. For example, the default minimum safe distance threshold is 5 meters, and the default maximum permissible acceleration is 2 meters per square second. When an anomalous target vehicle is identified, the strategy increases the minimum safe distance threshold to 10 meters and decreases the maximum permissible acceleration to 1 meter per square second. This requires the current autonomous vehicle to maintain a sufficiently large safe distance from the anomalous target vehicle while limiting the vehicle's own acceleration to avoid the need for sudden acceleration or deceleration due to the unpredictable behavior of the anomalous target vehicle, thereby reducing the risk of collision.
[0060] Furthermore, for compliance game strategies targeting rule-based vehicles, the minimum safe distance threshold among the standard safety risk threshold parameters is between the minimum safe distance threshold set by the defensive yielding strategy for aggressive target vehicles and the minimum safe distance threshold set by the decisive passage strategy for yielding target vehicles, and the maximum permissible acceleration does not exceed the preset regulatory limit. For example, if the defensive yielding strategy for aggressive target vehicles sets the minimum safe distance threshold to 8 meters, and the decisive passage strategy for yielding target vehicles sets the minimum safe distance threshold to 3 meters, then the standard minimum safe distance threshold for the compliance game strategy for rule-based target vehicles can be set to 5 meters, falling between the two; simultaneously, its maximum permissible acceleration does not exceed the limit stipulated by local traffic regulations, such as not exceeding 2 meters per square second. The preset regulatory limit can be determined according to the traffic regulations of the region where the target vehicle is located.
[0061] Furthermore, each autonomous driving strategy in the policy library is trained separately in a simulation environment containing target vehicles of the corresponding driving style type through multi-agent reinforcement learning. Specifically, during the training phase, simulation scenarios are constructed, each containing target vehicles with different driving styles: assertive, courteous, hesitant, rule-abiding, and abnormal. In each scenario, the target vehicle generates behavior according to the corresponding driving style type. The current autonomous driving vehicle continuously learns the optimal strategy action and optimal safety risk threshold parameters for that driving style type through trial and error in the interaction with these target vehicles using the multi-agent reinforcement learning algorithm, thereby obtaining the corresponding differentiated autonomous driving strategy in the policy library. Multi-agent reinforcement learning refers to a learning method in which multiple agents interact in the same environment, each optimizing its own strategy through feedback from interactions with the environment and other agents.
[0062] After deployment in real vehicles, the autonomous vehicles continuously optimize their driving strategies by fine-tuning the parameters online based on real-time driving behavior data and target vehicle motion data collected during actual driving. Online fine-tuning involves incrementally updating the deployed strategy parameters using newly collected real-world interaction data during vehicle operation, making the strategy more adaptable to the current traffic environment. The data used for fine-tuning includes driving behavior data generated by the autonomous vehicle itself after executing the strategy (e.g., actual speed, acceleration, steering angle) and perceived time-series motion data of each target vehicle. Fine-tuning can be implemented using online reinforcement learning or online gradient updates, which calculate the gradient information of the strategy performance using real-world interaction data and update the strategy parameters in small steps accordingly. For example, suppose the autonomous vehicle identifies a target vehicle as aggressive and selects a defensive yielding strategy from the strategy library, initially setting the minimum safe distance threshold to 8 meters. During actual driving, the current autonomous vehicle follows an 8-meter threshold for yielding, but repeatedly encounters aggressive target vehicles that cut in or abruptly decelerate to approach even at a distance of 8 meters, causing frequent emergency braking. At this point, based on actual driving behavior data of the current autonomous vehicle (e.g., frequent deceleration exceeding the comfort threshold, increased emergency braking) and temporal motion data of each target vehicle (e.g., rapidly decreasing relative distance, sudden changes in relative speed), the system determines that the current 8-meter minimum safe distance threshold is too small and insufficient to handle aggressive behavior from aggressive target vehicles in real-world scenarios. Therefore, the online fine-tuning process utilizes this actual interaction data to calculate the direction in which the minimum safe distance threshold should be increased through online gradient updates, and slightly increases this parameter from 8 meters to 8.5 meters. If the number of emergency braking decreases significantly in subsequent driving, it indicates that the fine-tuning direction is correct and the strategy has been optimized; if 8.5 meters is still insufficient, it continues to increase in small steps until it adapts to the current traffic environment. Conversely, if the defensive yielding strategy is found to be too conservative and frequently causes unnecessary stops, the minimum safe distance threshold can be appropriately reduced using actual interaction data to balance traffic efficiency.
[0063] Furthermore, after identifying the driving style type of each target vehicle and obtaining the corresponding confidence level, the current autonomous driving vehicle selects a matching autonomous driving strategy from the strategy library based on the driving style type and corresponding confidence level of each target vehicle. Specifically, for each target vehicle within the perception range, a differentiated autonomous driving strategy corresponding to the identified driving style type is searched in the strategy library. When multiple target vehicles correspond to different driving style types, the current autonomous driving vehicle makes a comprehensive judgment based on the confidence level of each target vehicle and its actual motion state to determine the final autonomous driving strategy to be executed. The confidence level is used to measure the reliability of the identification results of each target vehicle type. When the confidence level of a target vehicle is lower than a preset threshold, the identification result of that target vehicle type is considered unreliable and is not used as the main basis for strategy selection. In this case, the current autonomous driving vehicle adopts a preset conservative autonomous driving strategy to ensure interaction safety.
[0064] For example, suppose an autonomous vehicle is driving to an intersection without traffic lights, and three target vehicles appear simultaneously within its perception range: Target vehicle A is in the left lane and is approaching the intersection at a relatively high speed, and the recognition result is "aggressive," with a confidence level of 0.92; Target vehicle B is in the right lane, has clearly slowed down and turned on its turn signal, and the recognition result is "yielding," with a confidence level of 0.85; Target vehicle C is directly in front and is slowly approaching, with obvious speed fluctuations, and the recognition result is "hesitant," but the corresponding confidence level is only 0.45, which is lower than the preset threshold of 0.6.
[0065] At this point, target vehicle A has a high confidence level and is classified as a "hardline" type. Its defensive yielding strategy requires increasing the minimum safe distance threshold and reducing the conflict zone intrusion tolerance. Target vehicle B has a high confidence level and is classified as a "yielding" type. Its decisive passage strategy allows it to pass with a smaller safe distance after confirming yielding. Target vehicle C, however, has a confidence level below the preset threshold, and its type identification result is considered unreliable and not used as the primary basis for strategy selection. The current autonomous vehicle, considering the motion states of each target vehicle, determines that: target vehicle A is close to the intersection and has a high approach speed, posing the highest safety risk; target vehicle B is far away and is already slowing down to yield, posing a lower risk; target vehicle C is ahead but moving slowly, and its type is unreliable. After comprehensive judgment, the current autonomous vehicle prioritizes a defensive yielding strategy for target vehicle A, increasing the minimum safe distance threshold from the default 5 meters to 8 meters and reducing the conflict zone intrusion tolerance. Simultaneously, it adopts a preset conservative autonomous driving strategy in the direction of target vehicle C, maintaining a larger safe distance and limiting acceleration, waiting for further confirmation of its intentions before making a final decision. This avoids competing for right-of-way with aggressive vehicles and prevents misjudgments due to insufficient confidence from hesitant vehicles, thus ensuring safe interaction.
[0066] Step S140: Execute the matched autonomous driving strategy, and continuously update the vehicle driving style type and corresponding confidence level of each target vehicle during the execution process. When the vehicle driving style type of any target vehicle changes or the confidence level of the target vehicle is lower than a preset threshold, the autonomous driving strategy of the current autonomous vehicle is re-determined based on the updated vehicle driving style type and confidence level of each target vehicle.
[0067] Specifically, during the execution of the matched autonomous driving strategy, the current autonomous vehicle does not remain fixed after a single strategy selection. Instead, it continuously collects temporal motion data of each target vehicle within its perception range during execution. Based on the newly observed data, behavioral features are re-extracted and input into the vehicle driving style type recognition network, thereby continuously updating the vehicle driving style type and corresponding confidence level of each target vehicle. The purpose of this continuous updating is to promptly detect changes in the driving behavior style of target vehicles and changes in the reliability of the recognition results. When the driving style type of any target vehicle changes, or the confidence level of that target vehicle falls below a preset threshold, the autonomous driving strategy of the current autonomous vehicle is redefined based on the updated target vehicle types and confidence levels. Redefining the strategy means selecting a matching autonomous driving strategy from the strategy library based on the new type label and confidence level, and replacing the currently executed strategy accordingly. For example, a target vehicle is initially identified as a yielding vehicle, and the current autonomous vehicle executes a decisive passage strategy; as it approaches the intersection, the target vehicle suddenly accelerates and cuts in front of others. The newly observed behavioral characteristics cause the recognition network to update its vehicle driving style type to aggressive. At this point, the current autonomous vehicle immediately selects a defensive yielding strategy from the policy library to cope with the change in the target vehicle's behavioral style.
[0068] Furthermore, in scenarios where multiple target vehicles exist simultaneously, selecting a strategy for only a single target vehicle is often insufficient. The current autonomous vehicle needs to determine the safety risk level of each target vehicle based on its driving style and its motion state relative to the current autonomous vehicle, and then prioritize its handling according to the safety risk level. The motion state includes the distance and approach speed of the target vehicle relative to the current autonomous vehicle. The safety risk level measures the degree of potential threat posed by each target vehicle to the current autonomous vehicle; the higher the risk, the higher the priority for response. Specifically, the closer the target vehicle and the higher its approach speed, the higher its safety risk level; when the distance and approach speed are similar, aggressive or abnormal target vehicles have a higher safety risk level than yielding or rule-abiding target vehicles. The current autonomous vehicle prioritizes selecting the corresponding autonomous driving strategy for the target vehicle with the highest safety risk level. For example, if two target vehicles exist simultaneously within the perception range, target vehicle A is aggressive, 15 meters away from the current autonomous vehicle, and approaching at a speed of 5 meters per second; target vehicle B is hesitant, 8 meters away from the current autonomous vehicle, but approaching at only 0.5 meters per second. Although target vehicle B is closer, target vehicle A approaches at a faster speed and is more aggressive. After comprehensive assessment, target vehicle A has a higher safety risk level. Therefore, the current autonomous vehicle prioritizes the defensive yielding strategy corresponding to the tough type to ensure that it maintains a sufficient safe distance from target vehicle A.
[0069] Furthermore, when the confidence level of a target vehicle falls below a preset threshold and the autonomous driving strategy needs to be redefined, the vehicle driving style type identification result for that target vehicle is considered invalid and will not be used as the basis for strategy selection. The current autonomous driving vehicle will determine a preset conservative autonomous driving strategy for that target vehicle. The conservative autonomous driving strategy uses preset conservative safety risk threshold parameters and conservative strategy action parameters. The conservative safety risk threshold parameters include: the minimum safe distance threshold is not less than the maximum value among the minimum safe distance thresholds included in each autonomous driving strategy in the strategy library; the maximum permissible acceleration is not greater than the minimum value among the maximum permissible accelerations included in each autonomous driving strategy in the strategy library; and the yield decision time window is not less than the preset conservative safety time window value. This setting ensures that when type identification is unreliable, the current autonomous driving vehicle operates with the most conservative safety parameters, reducing the risk of misjudgment. For example, if the minimum safe distance threshold for the defensive yielding strategy in the strategy library is 8 meters, the minimum safe distance threshold for the decisive passage strategy is 3 meters, and other strategies fall between the two, then the minimum safe distance threshold for the conservative autonomous driving strategy should be no less than 8 meters. Similarly, if the minimum maximum permissible acceleration among all strategies is 1 meter per square second, then the maximum permissible acceleration of the conservative strategy should not exceed this value. By taking the most stringent parameters among all strategies, the conservative strategy can provide the maximum safety margin when the type is unknown or unreliable.
[0070] In addition to changes in the target vehicle type and a confidence level below a preset threshold, the conditions that trigger a re-determination of the current autonomous driving strategy for the autonomous vehicle also include at least one of the following: the target vehicle enters or leaves the current autonomous driving vehicle's perception range, or the current autonomous driving vehicle's driving state undergoes a sudden change. The target vehicle entering the perception range signifies the emergence of a new interaction object, requiring its inclusion in the strategy decision-making process. For example, if only yielding target vehicles were initially within the perception range, and the current autonomous driving vehicle was executing a decisive passage strategy, and then an aggressive target vehicle cuts into the perception range from an adjacent lane, the system needs to immediately re-determine the strategy to deal with the newly appearing aggressive object. The target vehicle leaving the perception range indicates that the object no longer affects the vehicle, and the response strategy can be released. For example, if a defensive yielding strategy was initially executed due to an aggressive target vehicle approaching from behind, and the target vehicle changes lanes and leaves the perception range, the system can re-determine the strategy and revert to a normal passage strategy. A sudden change in the current autonomous driving vehicle's driving state includes the current autonomous driving vehicle's speed change rate exceeding a preset change threshold or its heading angle change exceeding a preset angle threshold. For example, if a current autonomous vehicle brakes suddenly due to a sudden situation ahead, its speed drops dramatically in a very short time, exceeding the preset change threshold. The original following strategy and safety parameters may no longer be suitable for the current situation, and the strategy needs to be redefined to ensure safety.
[0071] Furthermore, while the target vehicle is within the perception range of the current autonomous vehicle, the current autonomous vehicle continuously collects and stores the target vehicle's behavioral characteristic data, forming the target vehicle's historical behavioral characteristic data. Historical behavioral characteristic data is a collection of behavioral characteristics accumulated over a period of time relative to the current moment, such as the target vehicle's speed change rate sequence, yielding behavior records, and trajectory sway amplitude changes over the past 30 seconds. This historical behavioral characteristic data can be used to correct the confidence level corresponding to the target vehicle. When the historical behavioral characteristic data is inconsistent with the currently identified vehicle driving style type, it indicates a deviation between the target vehicle's actual behavior pattern and the identification result; in this case, the confidence level corresponding to the target vehicle is reduced. If the confidence level falls below a preset threshold due to the reduction, the vehicle driving style type identification result for the target vehicle is considered invalid, and a preset conservative autonomous driving strategy is adopted. For example, if a target vehicle continuously slows down and yields within the first 20 seconds after entering the perception range, with a smooth trajectory, the recognition network identifies it as a yielding vehicle with a confidence level of 0.88. However, within the next 10 seconds, the target vehicle suddenly accelerates frequently and forcibly changes lanes multiple times. The early behavior in the historical behavioral feature data is significantly contradictory to the recent behavior, which is inconsistent with the currently identified yielding vehicle label. Based on this, the system reduces its confidence level from 0.88 to 0.52, which is lower than the preset threshold of 0.6. Therefore, the type identification result of the target vehicle is considered invalid, and the current autonomous driving vehicle adopts a conservative autonomous driving strategy for it, no longer treating it as a yielding vehicle.
[0072] Furthermore, when multiple target vehicles with different driving styles exist simultaneously within the perception range of the current autonomous vehicle, the current autonomous vehicle executes a comprehensive autonomous driving strategy. The safety risk threshold parameters of the comprehensive autonomous driving strategy are taken from the most stringent safety risk threshold parameters among the autonomous driving strategies corresponding to each target vehicle, namely the maximum minimum safe distance threshold, the minimum maximum permissible acceleration, and the maximum yield decision time window. The reason for choosing the most stringent parameters is that in a multi-target interaction environment, the autonomous vehicle must simultaneously meet the safety constraints of each target vehicle. Therefore, the most conservative parameters among the strategies for each target vehicle are used to ensure that the corresponding safety boundary is not exceeded for any target vehicle. For example, if there is a yielding target vehicle and an aggressive target vehicle within the perception range, the minimum safe distance threshold for the yielding type's decisive passage strategy is 3 meters and the maximum allowable acceleration is 2 meters per square second, while the minimum safe distance threshold for the aggressive type's defensive yielding strategy is 8 meters and the maximum allowable acceleration is 1.5 meters per square second. In this case, the comprehensive autonomous driving strategy takes the minimum safe distance threshold of 8 meters and the maximum allowable acceleration of 1.5 meters per square second, that is, adopts the more stringent parameters of the two to ensure that the vehicle maintains a sufficient safe distance from both target vehicles at the same time.
[0073] Furthermore, when the driving style of any target vehicle is identified as anomalous, the current autonomous vehicle, while executing safety avoidance and warning strategies, also triggers external warning actions. These external warning actions include alerting surrounding road users through lights and / or sounds. Anomalous target vehicles behave unpredictably, potentially changing lanes suddenly, braking abruptly, or violating traffic rules. Therefore, relying solely on the vehicle's own avoidance may not be sufficient to completely eliminate the risk. External warnings can alert other vehicles and road users to the presence and potential dangers of anomalous target vehicles, thus creating more comprehensive safety protection. For example, when an anomalous target vehicle ahead is identified as traveling on an unstable trajectory, the current autonomous vehicle, while maintaining a greater safe distance and limiting its own acceleration, warns surrounding vehicles by flashing its headlights or honking its horn, reminding drivers to avoid the anomalous target vehicle and reducing the risk of potential conflict.
[0074] In summary, this application acquires real-time temporal motion data of each target vehicle within the perception range of the current autonomous vehicle and extracts behavioral features. These behavioral features are then input into a pre-trained vehicle driving style type recognition network for online identification, yielding the driving style type and corresponding confidence level of each target vehicle. Based on the identification results, a matching autonomous driving strategy is selected from a pre-built strategy library. This strategy includes strategy action parameters and safety risk threshold parameters corresponding to the vehicle's driving style type. During the execution of the matched autonomous driving strategy, the driving style type and confidence level of each target vehicle are continuously updated. When the type of any target vehicle changes or its confidence level falls below a preset threshold, the autonomous driving strategy is redefined based on the updated information. Thus, this application achieves online real-time identification and dynamic updating of the driving styles of surrounding target vehicles, overcoming the problems of static identification results and the inability to dynamically adjust with the interaction process in existing technologies. Furthermore, by forming a closed-loop mapping between the identification results and differentiated autonomous driving strategies in the strategy library, the current autonomous vehicle can adaptively adjust the strategy actions and safety risk threshold parameters according to the actual driving style type of each target vehicle, avoiding the incompatibility of fixed strategies in different interaction scenarios. Furthermore, the strategy switching mechanism and multi-objective integrated processing method ensure that current autonomous vehicles can promptly redetermine or comprehensively select appropriate autonomous driving strategies in complex situations such as changes in target vehicle type, insufficient confidence, or multi-vehicle interaction, thereby improving the interaction safety and traffic efficiency in mixed traffic environments.
[0075] It should be understood that the above-described method for switching autonomous driving strategies based on vehicle driving style type recognition is merely exemplary. Those skilled in the art can make various modifications based on the above content, and the modified solutions also fall within the protection scope of this application.
[0076] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0077] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.
[0078] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.
[0079] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0080] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0081] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.
Claims
1. A method for automatic driving strategy switching based on vehicle driving style type recognition, characterized in that, include: The system acquires real-time temporal motion data of each target vehicle within the perception range of the current autonomous vehicle, and extracts behavioral features of the corresponding target vehicles based on the temporal motion data of each target vehicle; wherein, the behavioral features include speed change rate, yield probability, trajectory smoothness, reaction delay time, and relative distance change trend with the current autonomous vehicle. The behavioral characteristics of each target vehicle are input into a pre-trained vehicle driving style type recognition network for vehicle driving style type recognition, so as to obtain the vehicle driving style type and corresponding confidence level of each target vehicle; wherein, the vehicle driving style type includes aggressive, courteous, hesitant, rule-based and abnormal. Based on the vehicle driving style type and corresponding confidence level of each target vehicle, a matching autonomous driving strategy is selected from a pre-built strategy library; wherein, the autonomous driving strategy includes strategy action parameters and safety risk threshold parameters corresponding to the vehicle driving style type. The matched autonomous driving strategy is executed, and the vehicle driving style type and corresponding confidence level of each target vehicle are continuously updated during the execution process. When the vehicle driving style type of any target vehicle changes or the confidence level of the target vehicle is lower than the preset threshold, the autonomous driving strategy of the current autonomous driving vehicle is re-determined based on the updated vehicle driving style type and confidence level of each target vehicle.
2. The automatic driving policy switching method of claim 1, wherein, The strategy library stores differentiated autonomous driving strategies corresponding to each of the vehicle driving style types. Among them, the defensive yielding strategy for the aggressive target vehicle includes increasing the minimum safe distance threshold and reducing the conflict zone intrusion tolerance; the decisive passage strategy for the courteous target vehicle includes reducing the minimum safe distance threshold; the patient waiting and gradual probing strategy for the hesitant target vehicle includes increasing the yielding decision time window and increasing the conflict zone intrusion tolerance; the compliance game strategy for the rule-abiding target vehicle includes maintaining the preset standard safety risk threshold parameter; and the safety avoidance and warning strategy for the abnormal target vehicle includes increasing the minimum safe distance threshold and reducing the maximum permissible acceleration.
3. The autonomous driving strategy switching method according to claim 1 or 2, characterized in that, Each autonomous driving strategy in the pre-built strategy library is trained through multi-agent reinforcement learning in a simulation environment containing vehicles with corresponding driving style types. After deployment in real vehicles, the parameters of the matched autonomous driving strategy are fine-tuned online based on the driving behavior data of the current autonomous driving vehicle and the temporal motion data of each target vehicle collected during actual driving, so as to continuously optimize the autonomous driving strategy.
4. The autonomous driving strategy switching method according to claim 2, characterized in that, The minimum safe distance threshold in the standard safety risk threshold parameter is between the minimum safe distance threshold set for the defensive yielding strategy for the aggressive target vehicle and the minimum safe distance threshold set for the decisive passage strategy for the yielding target vehicle, and the maximum permissible acceleration in the standard safety risk threshold parameter does not exceed the preset regulatory limit.
5. The autonomous driving strategy switching method according to claim 1, characterized in that, When multiple target vehicles exist, the safety risk level of each target vehicle is determined based on its driving style type and its motion state relative to the current autonomous driving vehicle, and the processing priority is determined according to the safety risk level; wherein, the motion state includes the distance and approach speed of the target vehicle relative to the current autonomous driving vehicle; the corresponding autonomous driving strategy is selected first for the target vehicle with the highest safety risk level.
6. The autonomous driving strategy switching method according to claim 1, characterized in that, When the confidence level of any target vehicle is lower than the preset threshold and the autonomous driving strategy needs to be redefined, the vehicle driving style type identification result of the target vehicle is considered invalid and is not used as the basis for strategy selection. A preset conservative autonomous driving strategy is then determined for the current autonomous driving vehicle. The conservative autonomous driving strategy adopts preset conservative safety risk threshold parameters and conservative strategy action parameters. The conservative safety risk threshold parameters include: the minimum safe distance threshold is not less than the maximum value of the minimum safe distance thresholds contained in each autonomous driving strategy in the strategy library; the maximum permissible acceleration is not greater than the minimum value of the maximum permissible acceleration contained in each autonomous driving strategy in the strategy library; and the yield decision time window is not less than the preset safe conservative time window value.
7. The autonomous driving strategy switching method according to claim 1, characterized in that, The conditions for triggering a redetering of the autonomous driving strategy for the current autonomous vehicle also include at least one of the following: The target vehicle enters or leaves the perception range of the currently autonomous vehicle; The driving state of the current autonomous vehicle undergoes a sudden change; wherein, the sudden change in driving state includes the rate of change of the current autonomous vehicle speed exceeding a preset change threshold or the change in heading angle exceeding a preset angle threshold.
8. The autonomous driving strategy switching method according to claim 1, characterized in that, While the target vehicle is within the perception range of the current autonomous vehicle, behavioral characteristic data of the target vehicle is continuously collected and stored to form historical behavioral characteristic data of the target vehicle. The confidence level of the target vehicle is adjusted based on the historical behavioral feature data. When the historical behavioral feature data is inconsistent with the currently identified vehicle driving style type, the confidence level of the target vehicle is reduced. When the confidence level is lower than the preset threshold, the vehicle driving style type identification result of the target vehicle is considered invalid.
9. The autonomous driving strategy switching method according to claim 1, characterized in that, When multiple target vehicles with different driving styles exist simultaneously within the perception range of the current autonomous vehicle, the current autonomous vehicle executes a comprehensive autonomous driving strategy. The safety risk threshold parameter of the comprehensive autonomous driving strategy is the most stringent safety risk threshold parameter among the autonomous driving strategies corresponding to each target vehicle. The most stringent safety risk threshold parameter includes the maximum minimum safe distance threshold, the minimum maximum allowable acceleration, and the maximum yield decision time window.
10. The autonomous driving strategy switching method according to claim 2, characterized in that, When the driving style of any target vehicle is identified as the anomalous type, the current autonomous vehicle, while executing the safety avoidance and warning strategy, also triggers an external warning action, which includes issuing a warning to surrounding traffic participants through lights and / or sounds.
