An automatic driving behavior decision-making method based on double heuristic tree search

Abstract

The application discloses an automatic driving behavior decision-making method based on double heuristic tree search, and belongs to the technical field of automatic driving.The method is as follows: actions are abstracted as a combination of lateral lane-changing decision and longitudinal acceleration-deceleration decision, a decision time domain and a decision tree depth are set, and an initial behavior decision tree is constructed; a heuristic search strategy is designed by fusing field knowledge priori and data-driven priori, a reference action sequence is generated to guide the search process; a decision tree pruning mechanism based on field knowledge is proposed, and the tree structure is dynamically reconstructed with the last period optimal decision as a root node; the comprehensive cost of the candidate decision sequence after pruning is evaluated, and the sequence with the minimum cost is selected as the optimal decision sequence.The application effectively solves the technical problems of high computational complexity, poor real-time performance and insufficient decision rationality of traditional methods in a long decision time domain by constructing a semantic action decision tree and fusing a double heuristic guide and a field knowledge pruning mechanism.

Description

Technical Field

[0001] This invention relates to an autonomous driving behavior decision-making method based on dual heuristic tree search, belonging to the field of autonomous driving technology. Background Technology

[0002] In autonomous driving technology, behavioral decision-making, as the core link connecting environmental perception and vehicle control, directly determines the applicability of the autonomous driving system in dynamic traffic environments through its efficiency and rationality. Behavioral decision-making needs to be based on real-time environmental perception data to quickly plan the vehicle's driving strategy (such as lane changing, acceleration, and deceleration) in complex traffic scenarios in order to achieve safe and efficient driving goals.

[0003] Traditional autonomous driving behavior decision-making methods often employ state-space search or reinforcement learning, but in dynamic traffic environments, they face two core challenges:

[0004] First, after the action space is discretized, the number of decision sequences still grows exponentially with the decision time domain, which leads to a surge in computational complexity during long-term decision-making, making it difficult to meet real-time requirements.

[0005] Secondly, the lack of effective integration of human driving experience or domain knowledge makes the search process prone to ineffective path exploration, reducing decision-making efficiency.

[0006] Although some methods reduce the action space through semantic discretization (such as abstracting actions into semantic-level actions such as lane changing and acceleration / deceleration), in long decision-making time domains (such as 5 seconds and above), the curse of dimensionality still limits long-term decision-making capabilities, leading to decision delays or unreasonable strategies, which affects driving efficiency and safety. Summary of the Invention

[0007] To address the problems existing in the background technology, the present invention provides an autonomous driving behavior decision-making method based on dual heuristic tree search.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: an autonomous driving behavior decision-making method based on dual heuristic tree search, the method comprising the following steps:

[0009] S1: Construct a behavior decision tree based on semantic actions;

[0010] S2: Design a heuristic search strategy;

[0011] S3: Implement a domain knowledge pruning mechanism;

[0012] S4: Evaluate and optimize the cost of the decision sequence to generate an efficient and reasonable decision sequence for autonomous driving behavior.

[0013] Furthermore, step S1 includes the following steps:

[0014] S101: Define the set of semantic actions:

[0015] Autonomous driving behavior decisions are abstracted into semantic-level actions, including lateral and longitudinal decisions. Lateral decisions include left lane change, right lane change, and lane keeping; longitudinal decisions include acceleration, deceleration, and speed maintenance, forming a discrete action space. :

[0016] (1)

[0017] In formula (1):

[0018] For lateral movement;

[0019] It is a vertical movement;

[0020] S102: Define the decision time domain and decision tree depth, and construct the initial behavioral decision tree:

[0021] S10201: Decision time domain can be freely set according to onboard computing resources. ;

[0022] S10202: From the decision-making time domain Divide by time step The number of decision steps is calculated, i.e., the depth of the decision tree;

[0023] S10203: Constructing the initial action decision tree ,

[0024] in:

[0025] Decision tree nodes represent discrete action spaces. Internal decision-making actions;

[0026] The depth of a decision tree corresponds to the number of decision steps within the decision-making time domain;

[0027] The directed edges of a decision tree represent the sequential execution order of decision actions in time.

[0028] Furthermore, step S2 includes the following steps:

[0029] S201: Constructing Domain Knowledge Priors: Extracting Heuristic Constraints Based on Human Driving Rules;

[0030] S201 includes the following steps:

[0031] S20101: Extract priority constraints and define priority constraint scoring items. :

[0032] (2)

[0033] In formula (2):

[0034] for Decision-making actions at any given moment;

[0035] for Lateral movement at any moment;

[0036] Lane keeping in lateral decision-making;

[0037] For left lane changes in lateral decision-making;

[0038] For right lane changes in lateral decision-making;

[0039] S20102: Extract stationarity constraints and define stationarity constraint scoring terms. :

[0040] (3)

[0041] In formula (3):

[0042] for The rate of change of acceleration at any given moment;

[0043] Human acceleration comfort threshold set by humans;

[0044] S20103: Extract timing constraints and define timing constraint scoring items. :

[0045] (4)

[0046] In equation (4):

[0047] For continuous time steps The number of lateral movement changes within a given time period;

[0048] S20104: Combining the above constraints, the domain knowledge heuristic scoring function is obtained. :

[0049] (5)

[0050] In equation (5):

[0051] These are the weight parameters for priority constraints;

[0052] The weighting parameters for the stationarity constraint;

[0053] These are the weight parameters for the time series constraints.

[0054] S202: Constructing Data-Driven Priors:

[0055] Based on the deployment of a lightweight prediction network, historical traffic environment data is input, and the output is a reference action sequence of human drivers in the same scenario. Data-driven heuristics are constructed based on reference action sequences. :

[0056] (6)

[0057] In formula (6):

[0058] Action semantic distance;

[0059] S203: Heuristic Value of Computation Nodes

[0060] In the behavioral decision tree traversal, domain knowledge priors and data-driven priors are integrated to calculate the heuristic value of each node. The search process prioritizes expansion. High-value nodes:

[0061] (7)

[0062] In equation (7):

[0063] It inspires value for domain knowledge;

[0064] Data-driven insights and value;

[0065] The weighted parameter is used to weigh the influence of domain knowledge inspiration and data-driven inspiration on the path exploration in the search process.

[0066] Furthermore, step S3 includes the following steps:

[0067] S301: Reconstructing the Decision Tree Dynamics:

[0068] The best decision sequence in the previous period The last action Reconstruct the decision tree for the root node;

[0069] S302: Action Sequence Expansion Rules:

[0070] Calculate the number of action changes Define the constraint on the number of action changes as follows: :

[0071] (8)

[0072] In equation (8):

[0073] This is an indicator function; the value is 1 if the decision action is not simultaneous.

[0074] This represents the total number of decision actions in the decision sequence.

[0075] For decision sequence;

[0076] The number of action changes in the decision tree is... Pruning the decision sequence;

[0077] S303: Optimization of Analysis Complexity:

[0078] Unpruned decision tree The number of layer nodes grows exponentially. After pruning the decision tree based on domain knowledge, the root node expands to have 9 actions, and each subsequent layer is allowed only one change. Therefore, the number of nodes is... Complexity increases exponentially Reduced to polynomial growth .

[0079] Furthermore, step S4 includes the following steps:

[0080] S401: Calculate the total cost:

[0081] For the pruned candidate sequence Calculate the costs of safety, efficiency, and comfort.

[0082] S401 includes the following steps:

[0083] S40101: Based on minimum distance to obstacles Calculate security costs :

[0084] (9)

[0085] In equation (9):

[0086] A safety distance threshold set manually;

[0087] S40102: Based on deviation from desired speed Computational efficiency cost :

[0088] (10)

[0089] In formula (10):

[0090] for The vehicle's speed at all times;

[0091] For the desired speed;

[0092] S40103: Based on longitudinal acceleration rate of change Calculate the cost of comfort :

[0093] (11)

[0094] In equation (11):

[0095] for The longitudinal acceleration of the vehicle at all times;

[0096] The maximum acceleration rate set by the user;

[0097] S40104: Calculate the total cost:

[0098] (12)

[0099] In equation (12):

[0100] The hyperparameters for balancing the weights.

[0101] S402: Choosing the optimal decision:

[0102] The optimal decision sequence is the one with the minimum total cost obtained through tree search. :

[0103] (13)

[0104] In equation (13):

[0105] A decision tree pruned based on domain knowledge.

[0106] Compared with the prior art, the beneficial effects of the present invention are:

[0107] This invention employs a dual heuristic approach, combining domain knowledge priors (such as constraints from human driving rules) and data-driven priors (reference action sequences output by a lightweight prediction network), to provide high-value path guidance for the search process, significantly reducing ineffective path exploration and improving decision-making efficiency. By restricting high-frequency, unreasonable action changes through a domain knowledge pruning mechanism, the complexity is reduced from exponential to polynomial level, significantly decreasing computational load and meeting real-time requirements. Simultaneously, by integrating human driving experience with data-driven references, it balances decision rationality and safety, ultimately achieving high efficiency in long-term continuous decision-making. This effectively enhances the reliability and applicability of autonomous driving systems in complex and dynamic traffic scenarios, solving the technical problems of high computational complexity, poor real-time performance, and insufficient decision rationality in traditional methods over long decision-making time domains. Attached Figure Description

[0108] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0109] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0110] An autonomous driving behavior decision-making method based on dual heuristic tree search, the method comprising the following steps:

[0111] S1: Construct a behavior decision tree based on semantic actions;

[0112] S101: Define the set of semantic actions:

[0113] Autonomous driving behavior decisions are abstracted into semantic-level actions, including lateral and longitudinal decisions. Lateral decisions include left lane change, right lane change, and lane keeping, totaling three types. Longitudinal decisions include acceleration, deceleration, and speed maintenance, also totaling three types, forming a discrete action space. , by lateral movement and vertical movement Combining various steps, there are a total of 9 single-step movements:

[0114] (1)

[0115] In formula (1):

[0116] For lateral movement;

[0117] It is a vertical movement;

[0118] S102: Define the decision time domain and decision tree depth, and construct the initial behavioral decision tree:

[0119] S10201: Decision time domain can be freely set according to onboard computing resources. Preferred decision time domain Second;

[0120] S10202: From the decision-making time domain Divide by time step The number of decision steps is calculated, i.e., the depth of the decision tree, preferably the time step. Second;

[0121] S10203: Constructing the initial action decision tree ,

[0122] in:

[0123] Decision tree nodes represent discrete action spaces. The decision-making actions within the decision tree are the horizontal content carriers of the decision tree;

[0124] The depth of a decision tree corresponds to the number of decision steps within the decision time domain and is a vertical hierarchy indicator of the tree.

[0125] The directed edges of a decision tree represent the sequential execution order of decision actions in time, i.e., the... Layer nodes Pointing to the Layer nodes , indicating that the action starts from the first Step to the first The temporal relationships between steps represent the vertical connections within a decision tree. In the unpruned state, the number of nodes in the decision tree grows exponentially with the tree's depth, requiring optimization through subsequent pruning mechanisms.

[0126] S2: Design a heuristic search strategy;

[0127] S201: Constructing Domain Knowledge Priors: Extracting heuristic constraints based on human driving rules, including priority constraints, stability constraints, and temporal constraints;

[0128] S20101: Refine priority constraints so that lane keeping actions have higher priority than lane changing actions, and define priority constraint scoring items. :

[0129] (2)

[0130] In formula (2):

[0131] for The decision-making actions at any given moment consist of lateral decision-making and vertical decision-making;

[0132] for Lateral movement at any moment;

[0133] Lane keeping in lateral decision-making;

[0134] For left lane changes in lateral decision-making;

[0135] For right lane changes in lateral decision-making;

[0136] S20102: Extract stationarity constraints to ensure that the rate of change of longitudinal acceleration in decision-making does not exceed the human acceleration comfort threshold. And define the stationarity constraint scoring term. :

[0137] (3)

[0138] In formula (3):

[0139] for The rate of change of acceleration at any given moment;

[0140] Human acceleration comfort threshold set by humans;

[0141] S20103: Extract timing constraints and define timing constraint scoring items. :

[0142] (4)

[0143] In equation (4):

[0144] For continuous time steps Number of lateral action changes within a given time period; Number of lateral action changes within a short-term decision-making period No more than once, i.e., continuous time step Do not change lanes repeatedly within the specified time;

[0145] S20104: Combining the above constraints, the domain knowledge heuristic scoring function is obtained. :

[0146] (5)

[0147] In equation (5):

[0148] These are the weight parameters for priority constraints;

[0149] The weighting parameters for the stationarity constraint;

[0150] These are the weight parameters for the time series constraints.

[0151] Preferred Scoring function range ;

[0152] S202: Constructing Data-Driven Priors:

[0153] Based on the deployment of lightweight prediction networks (such as trajectory prediction models based on convolutional neural networks or recurrent neural networks), historical traffic environment data (such as surrounding vehicle trajectories and lane line information) are input, and the output is a reference action sequence of a human driver in the same scenario. (Consistent with the decision tree depth), a data-driven heuristic function is constructed based on the reference action sequence. :

[0154] (6)

[0155] In formula (6):

[0156] The semantic distance of the action is 1, such as the distance between left lane change + acceleration and lane keeping + acceleration, and the distance between left lane change + acceleration and right lane change + deceleration is 3. This is used to guide the search to focus on human-like actions.

[0157] S203: Heuristic Value of Computation Nodes

[0158] In the behavioral decision tree traversal, domain knowledge priors and data-driven priors are integrated to calculate the heuristic value of each node. The search process prioritizes expansion. High-value nodes:

[0159] (7)

[0160] In equation (7):

[0161] It inspires value for domain knowledge;

[0162] Data-driven insights and value;

[0163] The weighted parameter is used to weigh the influence of domain knowledge inspiration and data-driven inspiration on the path exploration in the search process.

[0164] S3: Implement a domain knowledge pruning mechanism;

[0165] S301: Reconstructing the Decision Tree Dynamics:

[0166] The best decision sequence in the previous period The last action Reconstruct the decision tree for the root node to ensure temporal continuity;

[0167] S302: Action Sequence Expansion Rules:

[0168] Calculate the number of action changes Define the constraint on the number of action changes as follows: :

[0169] (8)

[0170] In equation (8):

[0171] This is an indicator function; the value is 1 if the decision action is not simultaneous.

[0172] This represents the total number of decision actions in the decision sequence.

[0173] For decision sequence;

[0174] The number of action changes in the decision tree is... Pruning the decision sequence;

[0175] S303: Optimization of Analysis Complexity:

[0176] Unpruned decision tree The number of layer nodes grows exponentially. After pruning the decision tree based on domain knowledge, the root node expands to have 9 actions, and each subsequent layer is allowed only one change. Therefore, the number of nodes is... Complexity increases exponentially Reduced to polynomial growth .

[0177] S4: Evaluate and optimize the cost of the decision sequence to generate an efficient and reasonable decision sequence for autonomous driving behavior.

[0178] S401: Calculate the total cost:

[0179] For the pruned candidate sequence Calculate the costs of safety, efficiency, and comfort.

[0180] S40101: Based on minimum distance to obstacles Calculate security costs :

[0181] (9)

[0182] In equation (9):

[0183] A safety distance threshold set manually;

[0184] S40102: Based on deviation from desired speed Computational efficiency cost :

[0185] (10)

[0186] In formula (10):

[0187] for The vehicle's speed at all times;

[0188] For the desired speed;

[0189] S40103: Based on longitudinal acceleration rate of change Calculate the cost of comfort :

[0190] (11)

[0191] In equation (11):

[0192] for The longitudinal acceleration of the vehicle at all times;

[0193] The maximum acceleration rate set by the user;

[0194] S40104: Calculate the total cost:

[0195] (12)

[0196] In equation (12):

[0197] The hyperparameters for balancing the weights.

[0198] S402: Choosing the optimal decision:

[0199] The optimal decision sequence is the one with the minimum total cost obtained through tree search. :

[0200] (13)

[0201] In equation (13):

[0202] A decision tree pruned based on domain knowledge.

[0203] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0204] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An autonomous driving behavior decision-making method based on dual heuristic tree search, characterized in that: The method includes the following steps: S1: Construct a behavior decision tree based on semantic actions; S1 includes the following steps: S101: Define the set of semantic actions: Autonomous driving behavior decisions are abstracted into semantic-level actions, including lateral and longitudinal decisions. Lateral decisions include left lane change, right lane change, and lane keeping; longitudinal decisions include acceleration, deceleration, and speed maintenance, forming a discrete action space. : (1) In formula (1): For lateral movement; It is a vertical movement; S102: Define the decision time domain and decision tree depth, and construct the initial behavioral decision tree: S10201: Decision time domain can be freely set according to onboard computing resources. ; S10202: From the decision-making time domain Divide by time step The number of decision steps is calculated, i.e., the depth of the decision tree; S10203: Constructing the initial action decision tree , in: Decision tree nodes represent discrete action spaces. Internal decision-making actions; The depth of a decision tree corresponds to the number of decision steps within the decision-making time domain; The directed edges of a decision tree represent the sequential execution order of decision actions in time. S2: Design a heuristic search strategy; S2 includes the following steps: S201: Constructing Domain Knowledge Priors: Extracting Heuristic Constraints Based on Human Driving Rules; S202: Constructing Data-Driven Priors: Based on the deployment of a lightweight prediction network, historical traffic environment data is input, and the output is a reference action sequence of human drivers in the same scenario. Data-driven heuristics are constructed based on reference action sequences. : (6) In formula (6): Action semantic distance; S203: Heuristic Value of Computation Nodes In the behavioral decision tree traversal, domain knowledge priors and data-driven priors are integrated to calculate the heuristic value of each node. The search process prioritizes expansion. High-value nodes: (7) In equation (7): It inspires value for domain knowledge; Data-driven insights and value; The weighted parameter for heuristic value is used to weigh the impact of domain knowledge heuristics and data-driven heuristics on path exploration in the search process; S3: Implement a domain knowledge pruning mechanism; S3 includes the following steps: S301: Reconstructing the Decision Tree Dynamics: The best decision sequence in the previous period The last action Reconstruct the decision tree for the root node; S302: Action Sequence Expansion Rules: Calculate the number of action changes Define the constraint on the number of action changes as follows: : (8) In equation (8): This is an indicator function; the value is 1 if the decision action is not simultaneous. This represents the total number of decision actions in the decision sequence. For decision sequence; The number of action changes in the decision tree is... Pruning the decision sequence; S303: Optimization of Analysis Complexity: Unpruned decision tree The number of layer nodes grows exponentially. After pruning the decision tree based on domain knowledge, the root node expands to have 9 actions, and each subsequent layer is allowed only one change. Therefore, the number of nodes is... Complexity increases exponentially Reduced to polynomial growth ; S4: Evaluate and optimize the cost of the decision sequence to generate an efficient and reasonable autonomous driving behavior decision sequence; S4 includes the following steps: S401: Calculate the total cost: For the pruned candidate sequence Calculate the costs of safety, efficiency, and comfort. S402: Choosing the optimal decision: The optimal decision sequence is the one with the minimum total cost obtained through tree search. : (13) In equation (13): A decision tree pruned based on domain knowledge.

2. The autonomous driving behavior decision-making method based on dual heuristic tree search according to claim 1, characterized in that: S201 includes the following steps: S20101: Extract priority constraints and define priority constraint scoring items. : (2) In formula (2): for Decision-making actions at any given moment; for Lateral movement at any moment; Lane keeping in lateral decision-making; For left lane changes in lateral decision-making; For right lane changes in lateral decision-making; S20102: Extract stationarity constraints and define stationarity constraint scoring terms. : (3) In formula (3): for The rate of change of acceleration at any given moment; Human acceleration comfort threshold set by humans; S20103: Extract timing constraints and define timing constraint scoring items. : (4) In equation (4): For continuous time steps The number of lateral movement changes within a given time period; S20104: Combining the above constraints, the domain knowledge heuristic scoring function is obtained. : (5) In equation (5): These are the weight parameters for priority constraints; The weighting parameters for the stationarity constraint; These are the weight parameters for the time series constraints.

3. The autonomous driving behavior decision-making method based on dual heuristic tree search according to claim 2, characterized in that: S401 includes the following steps: S40101: Based on minimum distance to obstacles Calculate security costs : (9) In equation (9): A safety distance threshold set manually; S40102: Based on deviation from desired speed Computational efficiency cost : (10) In formula (10): for The vehicle's speed at all times; For the desired speed; S40103: Based on longitudinal acceleration rate of change Calculate the cost of comfort : (11) In equation (11): for The longitudinal acceleration of the vehicle at all times; The maximum acceleration rate set by the user; S40104: Calculate the total cost: (12) In equation (12): The hyperparameters for balancing the weights.

Images