Automatic driving decision planning method fusing behavior constraint and potential field constraint

Abstract

This invention discloses an autonomous driving decision-making and planning method that integrates behavioral constraints and potential field constraints, comprising the following steps: collecting vehicle motion data, constructing vehicle state vectors and vehicle temporal state vectors, inputting the vehicle temporal state vectors into a multi-vehicle interaction network to obtain a global interaction feature vector; based on the vehicle's temporal state vectors, using a behavioral consistency constraint unit, extracting the vehicle's behavioral feature vectors, and outputting the behavioral consistency constraint loss; inputting the vehicle's state vectors, global interaction feature vectors, and vehicle behavioral feature vectors into a decision network, and the decision network outputs autonomous driving decision commands; constructing a static potential field and a dynamic potential field, and generating an initial reference trajectory within a safe passage area generated by the coupling of the static and dynamic potential fields according to the autonomous driving decision commands; based on the initial reference trajectory, performing parameterized fitting, constructing a trajectory optimization model, determining the objective function and constraints, solving the trajectory optimization model, and outputting the autonomous driving trajectory.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and more specifically, to an autonomous driving decision-making and planning method that integrates behavioral consistency constraints and potential field safety constraints. Background Technology

[0002] In dynamic and complex multi-vehicle traffic flow environments, the safe and stable operation of autonomous vehicles highly depends on the environmental perception capabilities and multi-objective equilibrium optimization capabilities of the decision-making and planning system. Existing autonomous driving decision-making and planning technologies generally suffer from the following engineering shortcomings: traditional decision-making methods mostly use instantaneous traffic state input at a single moment, lacking the ability to mine the temporal correlation of continuous historical movement trajectories of surrounding vehicles, failing to fully characterize the dynamic interaction and evolution relationship between the vehicle and multiple adjacent vehicles, and having insufficient environmental anti-interference capabilities when faced with disturbances such as sudden lane changes or close-range cut-in by vehicles behind.

[0003] Furthermore, conventional decision-making outputs lack temporal consistency constraints, leading to frequent changes in driving strategies at different times. This results in repeated adjustments to the vehicle's lateral control actions, causing problems such as driving behavior shuddering and frequent steering fluctuations, leading to a poor human-like driving experience and reduced ride comfort. On the other hand, traditional trajectory planning algorithms lack a global continuous risk quantification constraint mechanism, relying solely on discrete safety distance thresholds for obstacle avoidance judgments, resulting in rigid safety boundary characterization. Simultaneously, they struggle to simultaneously balance trajectory safety and obstacle avoidance, driving smoothness, and real-time computational efficiency, leading to lengthy iterative solutions that fail to meet the low-latency operation requirements of in-vehicle platforms. Summary of the Invention

[0004] The purpose of this invention is to overcome the defects and shortcomings of the prior art and provide an autonomous driving decision-making and planning method that integrates behavioral constraints and potential field constraints. This method can effectively suppress behavioral decision fluctuations, reduce trajectory iteration time, and balance driving safety, comfort and traffic efficiency in multiple dimensions, making it applicable to the field of autonomous driving decision-making and planning.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] An autonomous driving decision-making and planning method that integrates behavioral constraints and potential field constraints includes the following steps:

[0007] S1. Collect vehicle motion data, construct vehicle state vector and vehicle temporal state vector, input the temporal state vectors of the vehicle and neighboring vehicles into the multi-vehicle interaction network to obtain the global interaction feature vector;

[0008] S2. Based on the vehicle's temporal state vector, a behavior consistency constraint unit is used to extract the vehicle's behavior feature vector and output the behavior consistency constraint loss to constrain the fluctuation of the vehicle's behavior.

[0009] S3. Input the vehicle state vector, global interaction feature vector, and vehicle behavior feature vector into the decision network, and the decision network outputs autonomous driving decision instructions;

[0010] S4. Construct static and dynamic potential fields, and generate an initial reference trajectory within the safe passage area generated by the coupling of static and dynamic potential fields according to the autonomous driving decision instructions;

[0011] S5. Based on the initial reference trajectory, perform parametric fitting, construct a trajectory optimization model, determine the objective function and constraints, solve the trajectory optimization model, and output the autonomous driving trajectory.

[0012] Further, in step S1, motion data of the vehicle and surrounding vehicles are collected, including real-time vehicle speed, lateral and longitudinal relative position, heading angle, and relative acceleration; based on the vehicle motion data, a vehicle state vector and a vehicle temporal state vector are constructed.

[0013] Define time The vehicle's state vector is ,time The The state vectors of the neighboring vehicles are: The global state vector of the vehicle at time t is ;

[0014] The time-series state vector of the vehicle itself and the time-series state vector of the neighboring vehicle are represented as follows:

[0015] ;

[0016] ;

[0017] In the formula, The length of the historical time window; Represents the vehicle's time-series state vector at time t; This represents the temporal state vector of the neighboring vehicle at time t; For a moment The vehicle's state vector; For a moment The vehicle's state vector; For a moment The The state vectors of the neighboring vehicles; For a moment The The state vectors of the neighboring vehicles;

[0018] By concatenating the time-series state vectors of the vehicle and neighboring vehicles at consecutive time points, the global time-series state vector of the vehicle is obtained. Global temporal state vector Includes historical moments The continuous vehicle state vector up to the current time t is represented as:

[0019] ;

[0020] Global temporal state vector Inputting the multi-vehicle interaction network yields a global interaction feature vector representing the vehicle interaction relationships. .

[0021] Furthermore, the temporal feature encoding layer of the multi-vehicle interaction network performs high-dimensional linear mappings on the temporal state vectors of the vehicle itself and the temporal state vectors of neighboring vehicles, respectively, to generate the query vectors required by the attention mechanism. Key vector AND value vector Among them, the vehicle's time-series state vector is obtained after being encoded by a time-series feature coding layer. , , The temporal state vector of the adjacent vehicle is obtained after being encoded by the temporal feature coding layer. , , ;

[0022] The attention mechanism calculates the association weights between different vehicles at different time points by measuring the similarity between the query vector and the key vector. The weight calculation formula is as follows:

[0023] ;

[0024] In the formula, The dimension of the key vector;

[0025] After passing through the attention mechanism, the output is the attention interaction feature vector. Then, using residual structure and layer normalization, preliminary interactive feature vectors are obtained. The calculation formula is:

[0026] ;

[0027] Finally, the feedforward connection network is used to process the initial interactive feature vectors. Perform deep nonlinear feature enhancement mapping to obtain the global interactive feature vector. The calculation formula is:

[0028] ;

[0029] In the formula, and It is a linear transformation matrix. and This is a bias term.

[0030] Furthermore, in step S2, the behavior consistency constraint unit adopts a two-layer stacked Transformer encoding network, and the input vehicle temporal state vector is... Output the vehicle behavior feature vector And behavioral consistency loss ;

[0031] The behavioral consistency loss is defined as follows:

[0032] ;

[0033] In the formula, , They are time points , The feature vector of the vehicle's behavior.

[0034] Furthermore, in step S3, the decision network uses the SAC algorithm framework. The policy network and value network of the SAC algorithm take multi-source feature fusion as input to realize joint decision-making of interactive perception, behavioral constraints and vehicle state.

[0035] First, the vehicle state vector, global interaction feature vector, and vehicle behavior feature vector are concatenated through channels to obtain the fused feature. :

[0036] ;

[0037] In terms of policy network updates, a reparameterization technique is used to sample actions. Then the loss function of the policy network for:

[0038] ;

[0039] In the formula, Indicates the policy network parameters; Represents the mathematical expectation; The regularization coefficient is used. The strategy to be adopted; Actions sampled in the policy network; The state of sampling; Let be the logarithmic probability of this strategy; Indicates the number of value networks used; for Actions in a state The value of the action;

[0040] Overall loss function Combining the loss function of the policy network with the behavior consistency loss, it can be expressed as:

[0041] ;

[0042] In the formula, This is the balance coefficient;

[0043] Ultimately, the policy network outputs autonomous driving decision instructions with the objective of minimizing the overall loss function. .

[0044] Further, in step S4, the static potential field includes the potential field of the lane lines, the potential field of the road boundary lines, and the potential field of the static obstacles. The potential field function of the lane lines is: The potential field function of the road boundary line is The potential field function of the static obstacle is The dynamic potential field dynamically adjusts the repulsive gain in real time based on the relative distance, the speed of passing vehicles, and the lateral angle. The dynamic potential field function is: ;

[0045] Total potential function for:

[0046] ;

[0047] The potential field intensity at each point in the road area is calculated using the overall potential field function, and a potential field intensity threshold is set. Areas with potential field strength greater than the threshold are designated as high-risk restricted areas, where trajectory crossing is prohibited; areas with potential field strength less than or equal to the threshold are designated as safe passage areas, serving as the generation range for the initial reference trajectory; and the initial reference trajectory is generated within the safe passage areas according to the autonomous driving decision instructions.

[0048] Furthermore, the potential field function of the lane line is:

[0049] ;

[0050] In the formula, This refers to the number of lane lines. For different lane line potential fields; Lane line type; Let x be the horizontal coordinate of any point; For the first The lateral coordinates of the lane markings; This is the attenuation coefficient for the road boundary line position;

[0051] The potential field function of the road boundary line is:

[0052] ;

[0053] In the formula, Define the road boundary line. , These are the left and right boundary lines of the road; The position decay function; Let be the horizontal coordinate of any point; For the first The lateral coordinates of the road boundary line;

[0054] The potential field function of a static obstacle is:

[0055] ;

[0056] In the formula, These are the coefficients of the static obstacle potential field function; This is a control coefficient for vehicle length; This is a control coefficient for vehicle width; L represents the center position of the static obstacle; W represents the vehicle length reference; and W represents the vehicle width reference.

[0057] The dynamic potential field function is:

[0058] ;

[0059] In the formula, This represents the dynamic potential field gain coefficient; Represents the distance vector; Represents the velocity vector of the target vehicle; Indicates the risk distribution coefficient based on spatial distance; Represents the elliptic normalized distance; This represents the angle between the distance vector and the velocity direction; Indicates the influence coefficient of the directional angle; Represents a unit vector indicating the direction of velocity; This represents the vector perpendicular to the velocity direction. This represents the attenuation coefficient of the ellipse length; This represents the attenuation coefficient of the ellipse width.

[0060] Further, in step S5, parametric fitting is performed based on the initial reference trajectory, specifically as follows:

[0061] The initial reference trajectory is sampled for neighborhood perturbation, and combined with fifth-order polynomial continuity constraints, the initial reference trajectory is parameterized into a fifth-order polynomial model:

[0062] ;

[0063] In the formula, , The horizontal and vertical trajectories are fifth-degree polynomials with respect to time t; These are the parameters of the fifth-order polynomial model.

[0064] Furthermore, a trajectory optimization model is constructed. The objective functions of the trajectory optimization model include a driving smoothness objective function, a potential field safety objective function, and a traffic efficiency objective function. The trajectory optimization model is solved by a sequential quadratic programming algorithm to output the autonomous driving trajectory.

[0065] Define the objective function for smooth driving. The square integral of the total jerky intensity is expressed as:

[0066] ;

[0067] In the formula, The second derivative of the horizontal position coordinate; The second derivative of the vertical position coordinate;

[0068] Define the objective function for the potential field security The cumulative potential field strength over the entire trajectory is expressed as:

[0069] ;

[0070] In the formula, coordinates The potential field strength;

[0071] Define the traffic efficiency objective function The deviation between the actual travel time and the ideal travel time is expressed as:

[0072] ;

[0073] in, This represents the total duration of the actual trajectory. The starting point is the vertical coordinate. The vertical coordinate of the endpoint For reference driving speed, Velocity in the form of an ideal trajectory;

[0074] The overall objective function is then expressed as:

[0075] ;

[0076] In the formula, For driving smoothing objective function Weighting coefficients; The objective function for the safety of the potential field Weighting coefficients; The objective function for traffic efficiency The weighting coefficients.

[0077] Furthermore, the constraints are as follows:

[0078] ;

[0079] In the formula, Indicates longitudinal velocity. , Indicates the minimum and maximum longitudinal velocities; Indicates lateral velocity. , Indicates the minimum and maximum lateral velocities; Indicates longitudinal acceleration. , Indicates the minimum and maximum longitudinal acceleration; Indicates lateral acceleration. , Indicates the minimum and maximum lateral acceleration; Indicates the horizontal position at time t. , A passable space constrained by a potential field; This represents the curvature of the trajectory at time t. This indicates the maximum curvature.

[0080] Compared with existing technologies, this invention integrates behavioral constraints and potential field constraints in an autonomous driving decision-making and planning method. By extracting global interaction feature vectors through a multi-vehicle interaction network, it effectively improves the adaptability to driving environments and the accuracy of interaction perception under complex traffic flow disturbance conditions. Through behavioral consistency loss, it suppresses high-frequency decision jumps and driving jitter, enhancing the temporal continuity of autonomous driving behavior and the smoothness of human-like driving. It reduces the computational overhead of trajectory iteration optimization, balances driving safety, comfort and traffic efficiency in multiple dimensions, and adapts to the real-time deployment requirements of in-vehicle embedded terminals. Attached Figure Description

[0081] Figure 1 A flowchart for an autonomous driving decision-making and planning method that integrates behavioral constraints and potential field constraints.

[0082] Figure 2 A schematic diagram of an autonomous driving decision-making and planning method that integrates behavioral constraints and potential field constraints.

[0083] Figure 3 This is a diagram showing the vehicle status of the vehicle itself and its neighboring vehicles.

[0084] Figure 4 This is a schematic diagram of the interaction in a multi-vehicle interaction network.

[0085] Figure 5 A schematic diagram of a safe passage area constrained by a potential field. Detailed Implementation

[0086] The following description, in conjunction with the accompanying drawings and specific embodiments, further illustrates the autonomous driving decision-making and planning method of the present invention that integrates behavioral constraints and potential field constraints.

[0087] Please see Figure 1 and Figure 2 This invention discloses an autonomous driving decision-making and planning method that integrates behavioral constraints and potential field constraints, comprising the following steps:

[0088] S1. Collect vehicle motion data, construct vehicle state vector and vehicle temporal state vector, input the temporal state vectors of the vehicle and neighboring vehicles into the multi-vehicle interaction network to obtain the global interaction feature vector;

[0089] S2. Based on the vehicle's temporal state vector, a behavior consistency constraint unit is used to extract the vehicle's behavior feature vector and output the behavior consistency constraint loss to constrain the fluctuation of the vehicle's behavior.

[0090] S3. Input the vehicle state vector, global interaction feature vector, and vehicle behavior feature vector into the decision network, and the decision network outputs autonomous driving decision instructions;

[0091] S4. Construct static and dynamic potential fields, and generate an initial reference trajectory within the safe passage area generated by the coupling of static and dynamic potential fields according to the autonomous driving decision instructions;

[0092] S5. Based on the initial reference trajectory, perform parametric fitting, construct a trajectory optimization model, determine the objective function and constraints, solve the trajectory optimization model, and output the autonomous driving trajectory.

[0093] Step S1: Collect vehicle motion data, construct vehicle state vector and vehicle temporal state vector, input the temporal state vectors of the vehicle and neighboring vehicles into the multi-vehicle interaction network to obtain the global interaction feature vector.

[0094] In the Frenet road coordinate system, with the EV as the center, consider the eight closest traffic participants traveling in the adjacent lanes. .like Figure 3 As shown, motion data from the vehicle and surrounding vehicles is collected, including real-time vehicle speed, lateral and longitudinal relative positions, heading angle, relative acceleration, and other multi-dimensional information. Based on the vehicle motion data, a vehicle state vector and a vehicle temporal state vector are constructed.

[0095] Define the vehicle's state vector at time t as: The first time t The state vectors of the neighboring vehicles are: The global state vector of the vehicle at time t is .

[0096] The time-series state vector of the vehicle itself and the time-series state vector of the neighboring vehicle are represented as follows:

[0097] ;

[0098] ;

[0099] In the formula, The length of the historical time window; Represents the vehicle's time-series state vector at time t; This represents the temporal state vector of the neighboring vehicle at time t; For a moment The vehicle's state vector; For a moment The vehicle's state vector; For a moment The The state vectors of the neighboring vehicles; For a moment The The state vectors of the neighboring vehicles;

[0100] By concatenating the time-series state vectors of the vehicle and neighboring vehicles at consecutive time points, the global time-series state vector of the vehicle is obtained. Global temporal state vector Includes historical moments The continuous vehicle state vector up to the current time t is represented as:

[0101] ;

[0102] like Figure 4 As shown, the temporal feature encoding layer (FC Layer) of the multi-vehicle interaction network performs high-dimensional linear mappings on the vehicle's temporal state vector and the temporal state vector respectively, thereby generating the query vector required by the attention structure. Key vector AND value vector Among them, the vehicle's time-series state vector is obtained after being encoded by a time-series feature coding layer. , , The temporal state vector of the adjacent vehicle is obtained after being encoded by the temporal feature coding layer. , , .

[0103] The attention mechanism calculates the association weights between different vehicles at different time points by measuring the similarity between the query vector and the key vector. The weight calculation formula is as follows:

[0104] ;

[0105] In the formula, is the dimension of the key vector, used for scaling to maintain numerical stability.

[0106] The output after the above attention mechanism is an attention interaction feature vector. Based on this, residual structure and layer normalization (LayerNorm) are used to improve the convergence stability of interactive features, resulting in preliminary interactive feature vectors. The calculation formula is as follows:

[0107] ;

[0108] Finally, the network is fed into a feedforward network (FFN) to perform deep nonlinear feature enhancement mapping on the initial interaction feature vector, further refining the implicit multi-vehicle interaction correlation patterns to obtain the global interaction feature vector. The joint representation of multi-vehicle interaction relationships is completed, and the calculation formula is as follows:

[0109] ;

[0110] In the formula, and It is a linear transformation matrix. and This is a bias term.

[0111] Step S2: Based on the vehicle's temporal state vector, the behavior consistency constraint unit is used to extract the vehicle's behavior feature vector and output the behavior consistency constraint loss to constrain the fluctuation of the vehicle's behavior.

[0112] Specifically, the behavior consistency constraint unit uses a two-layer stacked Transformer encoding network as the backbone feature extraction structure, taking the vehicle's temporal state vector as the core. As a single input, without introducing additional interference features from surrounding vehicles, it focuses on uncovering the temporal evolution correlation patterns of the vehicle's longitudinal and lateral driving behaviors; the behavior consistency constraint unit ultimately outputs a normalized behavior feature vector. Loss of behavioral consistency .

[0113] Behavioral consistency loss is used to constrain the variation amplitude of the vehicle's behavioral feature vector across adjacent time sequences, suppressing high-frequency jumps in decision-making and fluctuations in driving behavior. The behavioral consistency loss is defined as follows:

[0114] ;

[0115] In the formula, , They are time points , The feature vector of the vehicle's behavior.

[0116] Step S3: Input the vehicle state vector, global interaction feature vector, and vehicle behavior feature vector into the decision network, and the decision network outputs autonomous driving decision instructions.

[0117] Specifically, the decision network uses the SAC algorithm as its underlying network framework. The policy network and value network of the SAC algorithm take multi-source feature fusion as input to achieve joint decision-making based on interactive perception, behavioral constraints, and vehicle state. First, the vehicle state vector, global interaction feature vector, and vehicle behavior feature vector are concatenated to obtain the fused features. :

[0118] ;

[0119] This fusion feature integrates the vehicle's own state, the interaction with surrounding vehicles, and the vehicle's behavioral characteristics, providing a comprehensive basis for policy output.

[0120] In terms of policy network updates, a reparameterization technique is used to sample actions. Then the loss function of the policy network for:

[0121] ;

[0122] In the formula, Indicates the policy network parameters; Represents the mathematical expectation; The regularization coefficient is used. The strategy to be adopted; Actions sampled in the policy network; The state of sampling; Let be the logarithmic probability of this strategy; Indicates the number of value networks used; for Actions in a state The value of the action.

[0123] Final overall loss function The following formula is used:

[0124] ;

[0125] In the formula, This is a balancing coefficient used to adjust the weights of strategy optimization and behavioral consistency.

[0126] Ultimately, the policy network outputs autonomous driving decision instructions with the objective of minimizing the overall loss function. .

[0127] Step S4: Construct a static potential field and a dynamic potential field. Based on the autonomous driving decision instructions, generate an initial reference trajectory within the safe passage area generated by the coupling of the static and dynamic potential fields.

[0128] Specifically, the static potential field includes the potential field of the lane lines, the potential field of the road boundary lines, and the potential field of the static obstacles. The potential field function of the lane lines is: The potential field function of the road boundary line is The potential field function of the static obstacle is .

[0129] The potential field function of the lane line is:

[0130] ;

[0131] In the formula, This refers to the number of lane lines. For different lane line potential fields; Lane line type; Let x be the horizontal coordinate of any point; For the first The lateral coordinates of the lane markings; This is the attenuation coefficient for the road boundary line position.

[0132] The potential field function of the road boundary line is:

[0133] ;

[0134] In the formula, Define the road boundary line. , These are the left and right boundary lines of the road; The position decay function; Let be the horizontal coordinate of any point; For the first The lateral coordinates of the road boundary line.

[0135] The potential field function of a static obstacle is:

[0136] ;

[0137] In the formula, These are the coefficients of the static obstacle potential field function; , These correspond to the control coefficients for the vehicle's length and width, respectively. L represents the center position of the static obstacle; L is the vehicle length reference, and W is the vehicle width reference.

[0138] The dynamic potential field dynamically adjusts the repulsive force gain in real time based on relative distance, intersecting vehicle speed, and lateral angle, automatically strengthening safety constraints in high-risk areas at close range. Dynamic potential field function The expression is as follows:

[0139] ;

[0140] In the formula, This represents the dynamic potential field gain coefficient; Represents the distance vector; Represents the velocity vector of the target vehicle; Indicates the risk distribution coefficient based on spatial distance; Represents the elliptic normalized distance; This represents the angle between the distance vector and the velocity direction; Indicates the influence coefficient of the directional angle; Represents a unit vector indicating the direction of velocity; This represents the vector perpendicular to the velocity direction. This represents the attenuation coefficient of the ellipse length; This represents the attenuation coefficient of the ellipse width.

[0141] The overall potential function is:

[0142] ;

[0143] The potential field intensity at each point in the road area is calculated using the overall potential field function, and a potential field intensity threshold is set. Areas with potential field strength greater than a certain threshold are designated as high-risk restricted zones, prohibiting trajectory crossings; areas with potential field strength less than or equal to the threshold are designated as safe, passable zones, serving as the generation range for the initial reference trajectory. Based on the autonomous driving decision instructions, the initial reference trajectory is generated within these safe zones. For example... Figure 5 The diagram shows the safe passage region constrained by the potential field. The red dashed line represents the boundary of the safe passage region, and the shape of the boundary is determined by the potential fields from multiple sources.

[0144] Among them, the gradient of the potential field The direction pointing to the area where the potential field strength increases most rapidly corresponds to the location of the high-risk area. The repulsive force generated by the potential field is defined as... Its direction is exactly in the direction where the potential field strength decreases the fastest, that is, the direction of the low-risk safety zone, which provides a clear obstacle avoidance guiding force for the trajectory.

[0145] The direction of travel of the initial reference trajectory is always opposite to the repulsive force of the potential field. To maintain consistency, the path is iteratively generated by following the direction of the repulsive force point by point, ensuring that the trajectory moves towards a low-risk region at each step. Simultaneously, decision instructions guide the direction of travel, preventing the trajectory from getting trapped in local optima. This guidance mechanism fundamentally avoids high-risk areas, while relying on the boundary constraints of the safety space to ensure that the initial reference trajectory always remains within the legal feasible region, ultimately resulting in an initial reference trajectory based on potential field constraints.

[0146] Step S5: Based on the initial reference trajectory, perform parametric fitting, construct a trajectory optimization model, determine the objective function and constraints, solve the trajectory optimization model, and output the autonomous driving trajectory.

[0147] Specifically, neighborhood perturbation sampling is performed on the initial reference trajectory, and combined with fifth-order polynomial continuity constraints, the initial reference trajectory is parameterized into a fifth-order polynomial model:

[0148] ;

[0149] In the formula, , The horizontal and vertical trajectories are fifth-degree polynomials with respect to time t; These are the parameters of the fifth-order polynomial model.

[0150] A trajectory optimization model is constructed, whose objective functions include driving smoothness objective function, potential field safety objective function, and traffic efficiency objective function. The trajectory optimization model is solved by a sequential quadratic programming algorithm to output the autonomous driving trajectory.

[0151] Jerk is an important indicator of trajectory smoothness, and its magnitude directly determines the vehicle's ride comfort. To ensure global trajectory smoothness, the driving smoothness objective function is defined as the square integral of the entire jerk:

[0152] ;

[0153] In the formula, The second derivative of the horizontal position coordinate; is the second derivative of the vertical position coordinate.

[0154] Based on the potential field, the higher the potential field strength at a trajectory point, the greater the driving risk. The potential field safety objective function is defined as the cumulative potential field strength over the entire trajectory, guiding the trajectory to actively avoid high-risk areas.

[0155] ;

[0156] In the formula, coordinates The potential field function value.

[0157] To ensure the engineering practicality of the trajectory and avoid unnecessary longitudinal deceleration or lateral redundant offset, the traffic efficiency objective function is defined as the deviation between the actual travel time and the ideal travel time of the trajectory:

[0158] ;

[0159] in, This represents the total duration of the actual trajectory. The starting point is the vertical coordinate. The vertical coordinate of the endpoint For reference driving speed, The velocity is in the form of an ideal trajectory.

[0160] The overall objective function is then expressed as:

[0161] ;

[0162] In the formula, For driving smoothing objective function Weighting coefficients; The objective function for the safety of the potential field Weighting coefficients; The objective function for traffic efficiency The weighting coefficients.

[0163] The constraints of the trajectory optimization model are as follows:

[0164] ;

[0165] In the formula, Indicates longitudinal velocity. , Indicates the minimum and maximum longitudinal velocities; Indicates lateral velocity. , Indicates the minimum and maximum lateral velocities; Indicates longitudinal acceleration. , Indicates the minimum and maximum longitudinal acceleration; Indicates lateral acceleration. , Indicates the minimum and maximum lateral acceleration; Indicates the horizontal position at time t. , The passable space is constrained by the potential field. This represents the curvature of the trajectory at time t. This indicates the maximum curvature.

[0166] The trajectory optimization model is solved iteratively using a sequential quadratic programming algorithm, and the final output is an autonomous driving trajectory that takes into account safety, obstacle avoidance, driving comfort, efficient passage and dynamics throughout the process, thus completing the closed-loop decision planning output.

[0167] The CARLA simulator is used to construct a test traffic scenario. To verify the effectiveness of different methods, the proposed decision planning method is compared with three other reinforcement learning methods: PPO (Proximal Policy Optimization), SAC (Soft Actor Critic), and INTSAC (Interaction-based Soft Actor Critic).

[0168] The PPO method, as a fundamental policy gradient method, suppresses policy update amplitude by pruning the objective function, thus improving training stability. The SAC method, as a current high-performance maximum entropy reinforcement learning method, simultaneously optimizes policy entropy and cumulative reward, exhibiting good stability and exploration capabilities in continuous action spaces, and forms the underlying architecture of the proposed framework. The IntSAC method, serving as the ablation experiment baseline, retains the multi-vehicle interaction network on top of SAC while removing the behavior consistency constraint unit to evaluate its contribution to behavior coherence and policy robustness. The test results are summarized and analyzed, as shown in Table 1.

[0169] Table 1

[0170]

[0171] A comprehensive comparison of the present invention was conducted from both quantitative and qualitative perspectives. Experimental results show that the present invention significantly outperforms the comparative methods in core indicators such as task success rate, average round reward, number of lane change penalties, acceleration variance, and average speed. Compared with the SAC method, the present invention improves the task success rate by 9.65%, the average round reward by 54.98%, reduces the number of lane change penalties by 66.67%, reduces acceleration variance by 61.40%, and increases the average driving speed by 6.42%. These data indicate that the present invention can improve the driving efficiency, safety, and behavioral smoothness of the decision-making method while ensuring driving safety.

[0172] The above description is a detailed description of the preferred embodiments of the present invention. However, the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modifications made under the technical spirit disclosed in the present invention should fall within the patent scope covered by the present invention.

Claims

1. An automatic driving decision planning method fusing behavior constraints and potential field constraints, characterized in that, Includes the following steps; S1. Collect vehicle motion data, construct vehicle state vector and vehicle temporal state vector, input the temporal state vectors of the vehicle and neighboring vehicles into the multi-vehicle interaction network to obtain the global interaction feature vector; S2. Based on the vehicle's temporal state vector, a behavior consistency constraint unit is used to extract the vehicle's behavior feature vector and output the behavior consistency constraint loss to constrain the fluctuation of the vehicle's behavior. S3. Input the vehicle state vector, global interaction feature vector, and vehicle behavior feature vector into the decision network, and the decision network outputs autonomous driving decision instructions; S4. Construct static and dynamic potential fields, and generate an initial reference trajectory within the safe passage area generated by the coupling of static and dynamic potential fields according to the autonomous driving decision instructions; S5. Based on the initial reference trajectory, perform parametric fitting, construct a trajectory optimization model, determine the objective function and constraints, solve the trajectory optimization model, and output the autonomous driving trajectory.

2. The method of claim 1, wherein, In step S1, motion data of the vehicle and surrounding vehicles are collected, including real-time vehicle speed, lateral and longitudinal relative position, heading angle, and relative acceleration; based on the vehicle motion data, a vehicle state vector and a vehicle temporal state vector are constructed. Definition of time The self-vehicle state vector at time is The th neighboring vehicle state vector at time is The vehicle global state vector at time t is The time-series state vector of the vehicle itself and the time-series state vector of the neighboring vehicle are represented as follows: ； ； In the formula, The length of the historical time window; Represents the vehicle's time-series state vector at time t; This represents the temporal state vector of the neighboring vehicle at time t; For a moment The vehicle's state vector; For a moment The vehicle's state vector; For a moment The The state vectors of the neighboring vehicles; For a moment The The state vectors of the neighboring vehicles; By concatenating the time-series state vectors of the vehicle and neighboring vehicles at consecutive time points, the global time-series state vector of the vehicle is obtained. Global temporal state vector Includes historical moments The continuous vehicle state vector up to the current time t is represented as: ； Global temporal state vector Inputting the multi-vehicle interaction network yields a global interaction feature vector representing the vehicle interaction relationships. .

3. The autonomous driving decision-making and planning method integrating behavioral constraints and potential field constraints according to claim 2, characterized in that, The temporal feature encoding layer of the multi-vehicle interaction network performs high-dimensional linear mappings on the temporal state vectors of the vehicle itself and the temporal state vectors of neighboring vehicles, thereby generating the query vectors required by the attention mechanism. Key vector AND value vector Among them, the vehicle's time-series state vector is obtained after being encoded by a time-series feature coding layer. , , The temporal state vector of the adjacent vehicle is obtained after being encoded by the temporal feature coding layer. , , ; The attention mechanism calculates the association weights between different vehicles at different time points by measuring the similarity between the query vector and the key vector. The weight calculation formula is as follows: ； In the formula, The dimension of the key vector; After passing through the attention mechanism, the output is the attention interaction feature vector. Then, using residual structure and layer normalization, preliminary interactive feature vectors are obtained. The calculation formula is: ； Finally, the feedforward connection network is used to process the initial interactive feature vectors. Perform deep nonlinear feature enhancement mapping to obtain the global interactive feature vector. The calculation formula is: ； In the formula, and It is a linear transformation matrix. and This is a bias term.

4. The autonomous driving decision-making and planning method integrating behavioral constraints and potential field constraints according to claim 3, characterized in that, In step S2, the behavior consistency constraint unit adopts a two-layer stacked Transformer encoding network, and the input is the vehicle's temporal state vector. Output the vehicle behavior feature vector And behavioral consistency loss ; The behavioral consistency loss is defined as follows: ； In the formula, , They are time points , The feature vector of the vehicle's behavior.

5. The autonomous driving decision-making and planning method integrating behavioral constraints and potential field constraints according to claim 4, characterized in that, In step S3, the decision network uses the SAC algorithm framework. The policy network and value network of the SAC algorithm take multi-source feature fusion as input to realize joint decision-making of interactive perception, behavioral constraints and vehicle state. First, the vehicle state vector, global interaction feature vector, and vehicle behavior feature vector are concatenated through channels to obtain the fused feature. : ； In terms of policy network updates, a reparameterization technique is used to sample actions. Then the loss function of the policy network for: ； In the formula, Indicates the policy network parameters; Represents the mathematical expectation; The regularization coefficient is used. The strategy to be adopted; Actions sampled in the policy network; The state of sampling; Let be the logarithmic probability of this strategy; Indicates the number of value networks used; for Actions in a state The value of the action; Overall loss function Combining the loss function of the policy network with the behavior consistency loss, it can be expressed as: ； In the formula, This is the balance coefficient; Ultimately, the policy network outputs autonomous driving decision instructions with the objective of minimizing the overall loss function. .

6. The autonomous driving decision-making and planning method integrating behavioral constraints and potential field constraints according to claim 1, characterized in that, In step S4, the static potential field includes the potential field of the lane lines, the potential field of the road boundary lines, and the potential field of the static obstacles. The potential field function of the lane lines is: The potential field function of the road boundary line is The potential field function of the static obstacle is The dynamic potential field dynamically adjusts the repulsive gain in real time based on the relative distance, the speed of passing vehicles, and the lateral angle. The dynamic potential field function is: ; Total potential function for: ； The potential field intensity at each point in the road area is calculated using the overall potential field function, and a potential field intensity threshold is set. Areas with potential field strength greater than this threshold are designated as high-risk restricted areas, and traversal of such areas is prohibited. The area less than or equal to this threshold is defined as the safe passage area and used as the range for generating the initial reference trajectory; Based on the autonomous driving decision instructions, an initial reference trajectory is generated within the safe passage area.

7. The autonomous driving decision-making and planning method integrating behavioral constraints and potential field constraints according to claim 6, characterized in that, The potential field function of the lane line is: ； In the formula, This refers to the number of lane lines. For different lane line potential fields; Lane marking type; Let x be the horizontal coordinate of any point; For the first The lateral coordinates of the lane markings; This is the attenuation coefficient for the road boundary line position; The potential field function of the road boundary line is: ； In the formula, Define the road boundary line. , These are the left and right boundary lines of the road; The position decay function; Let be the horizontal coordinate of any point; For the first The lateral coordinates of the road boundary line; The potential field function of a static obstacle is: ； In the formula, These are the coefficients of the static obstacle potential field function; This is a control coefficient for vehicle length; This is a control coefficient for vehicle width; L represents the center position of the static obstacle; W represents the vehicle length reference; and W represents the vehicle width reference. The dynamic potential field function is: ； In the formula, This represents the dynamic potential field gain coefficient; Represents the distance vector; Represents the velocity vector of the target vehicle; Indicates the risk distribution coefficient based on spatial distance; Represents the elliptic normalized distance; This represents the angle between the distance vector and the velocity direction; Indicates the influence coefficient of the directional angle; Represents a unit vector indicating the direction of velocity; This represents the vector perpendicular to the velocity direction; This represents the attenuation coefficient of the ellipse length; This represents the attenuation coefficient of the ellipse width.

8. The autonomous driving decision-making and planning method integrating behavioral constraints and potential field constraints according to claim 1, characterized in that, In step S5, parametric fitting is performed based on the initial reference trajectory, specifically as follows: The initial reference trajectory is sampled for neighborhood perturbation, and combined with fifth-order polynomial continuity constraints, the initial reference trajectory is parameterized into a fifth-order polynomial model: ； In the formula, , The horizontal and vertical trajectories are fifth-degree polynomials with respect to time t; These are the parameters of the fifth-order polynomial model.

9. The autonomous driving decision-making and planning method integrating behavioral constraints and potential field constraints according to claim 8, characterized in that, A trajectory optimization model is constructed, whose objective functions include driving smoothness objective function, potential field safety objective function, and traffic efficiency objective function. The trajectory optimization model is solved by a sequential quadratic programming algorithm to output the autonomous driving trajectory. Define the objective function for smooth driving. The square integral of the total jerk is expressed as: ； In the formula, The second derivative of the horizontal position coordinate; The second derivative of the vertical position coordinate; Define the objective function for the potential field security The cumulative potential field strength over the entire trajectory is expressed as: ； In the formula, coordinates The potential field strength; Define the traffic efficiency objective function The deviation between the actual travel time and the ideal travel time is expressed as: ； in, This represents the total duration of the actual trajectory. The starting point is the vertical coordinate. The vertical coordinate of the endpoint For reference driving speed, Velocity in the form of an ideal trajectory; The overall objective function is then expressed as: ； In the formula, For driving smoothing objective function Weighting coefficients; The objective function for the safety of the potential field Weighting coefficients; The objective function for traffic efficiency The weighting coefficients.

10. The autonomous driving decision-making and planning method integrating behavioral constraints and potential field constraints according to claim 9, characterized in that, The constraints are as follows: ； In the formula, Indicates longitudinal velocity. , Indicates the minimum and maximum longitudinal velocities; Indicates lateral velocity. , Indicates the minimum and maximum lateral velocities; Indicates longitudinal acceleration. , Indicates the minimum and maximum longitudinal acceleration; Indicates lateral acceleration. , Indicates the minimum and maximum lateral acceleration; Indicates the horizontal position at time t. , A passable space constrained by a potential field; This represents the curvature of the trajectory at time t. This indicates the maximum curvature.

Images