Automatic driving decision control method and system and related equipment

Abstract

The invention discloses an automatic driving decision control method and system and related equipment, and belongs to the technical field of automatic driving. The method comprises the following steps: generating an initial behavior decision tree by adopting a Markov algorithm according to a vehicle state; pruning and reconstructing the initial behavior decision tree according to human driving characteristics to obtain a pre-screened behavior decision tree; pruning and rejecting the unsafe behavior decision sequence through an RSS model, and iteratively generating a vehicle candidate trajectory in combination with a vehicle kinematic model and a behavior decision tree; performing trajectory generation on the vehicle candidate trajectory sequence to obtain a predicted trajectory; performing safety speed check and collision check on the predicted trajectory, and rejecting behavior sequences which do not meet requirements to obtain a pre-selected trajectory; and calculating the total cost of each pre-selected track, and taking the behavior sequence with the minimum total cost as an optimal vehicle decision result and outputting the optimal vehicle decision result. According to the invention, through multi-level screening and optimization of the decision sequence, the safety and rationality of the automatic driving decision are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, in particular to an automatic driving decision control method and system and related equipment. BACKGROUND

[0002] With the rapid development of automatic driving technology, the decision control system of automatic driving vehicles as a core component directly affects the safety and reliability of automatic driving. At present, the automatic driving decision control system is mainly responsible for making appropriate driving decisions, such as lane keeping, lane changing, acceleration and deceleration, etc., according to the environmental information obtained by the perception system and the vehicle state.

[0003] The decision algorithm of automatic driving is a key link to ensure driving safety and efficiency. It usually includes three levels of global path planning, behavior decision and motion planning. Current decision algorithms face challenges in dealing with complex dynamic traffic scenarios, especially in considering multiple safety mechanisms.

[0004] Global path planning is the first step in the decision-making process, which generates a global path based on map information and driving destination. This step is the basis for subsequent behavior decision and motion planning. The automatic driving system needs to make specific behavior decisions according to the perceived environmental information (including other vehicles, pedestrians, obstacles and traffic rules), such as deciding whether to change lanes or follow the front vehicle. Behavior decision needs to consider multi-agent interaction and uncertainty, such as partially observable Markov decision process can be used to handle such uncertainty. According to the results of behavior decision, the motion planning layer is responsible for generating a specific trajectory, which needs to meet the vehicle dynamics constraints, avoid collisions and consider passenger comfort. Motion planning usually involves path planning and speed planning, which can use sampling-based methods, graph search methods or numerical optimization methods. The automatic driving system needs to be able to evaluate driving risk, which includes comparative analysis of time, acceleration and distance indicators. Probabilistic evaluation methods use probability models to describe the motion of traffic participants, and then evaluate the risk. In addition, potential field-based evaluation methods are also used to describe collision risk. Current decision algorithms may not fully consider multiple safety mechanisms, which may lead to safety risks in the planning process.

[0005] The safety challenges of automatic driving include the uncertainty of perception and positioning systems, the complexity of decision and planning systems, and the reliability of navigation and control systems. Future research directions may include improving the ability to recognize and handle complex environments, enhancing the ability to cooperate with human drivers, and building an integrated safety framework for overall cooperation. The decision algorithm of automatic driving is a complex system that needs to evolve continuously to adapt to changing traffic environments and ensure safe and efficient decisions in all situations.

[0006] In the field of automatic driving decision control, Markov decision process is a commonly used decision model. However, there are still some problems in existing automatic driving decision control technology: first, existing decision algorithms often lack multiple safety mechanisms, which pose safety risks to planning. Although some methods use a safety distance model to determine safety, they lack accurate calculation of safety distance in different scenarios and multi-level safety check mechanisms, which may lead to safety hazards in complex traffic environments. Second, the decision algorithm is not efficient and cannot effectively eliminate unreasonable decision results. Existing decision tree construction methods often lack consideration of human driving characteristics, resulting in decision trees that contain a large number of decision sequences that do not conform to actual driving habits, increasing computational complexity and potentially leading to unreasonable decision results. Finally, existing decision algorithms fail to adequately simulate the thinking patterns of human drivers, resulting in decision results that are not natural and safe. Although some methods consider vehicle interaction behavior and driving style, they lack systematic modeling and application of human driving characteristics, resulting in differences between the behavior of autonomous vehicles and human drivers, affecting the acceptability and safety of autonomous driving systems. SUMMARY

[0007] To solve the problems of existing automatic driving decision algorithms lacking multiple safety mechanisms, low decision efficiency, and failing to simulate the thinking patterns of human drivers, and to achieve the effects of improving decision safety, algorithm running efficiency, and driving experience, the present application provides an automatic driving decision control method, system, and related equipment. This method can consider safety, efficiency, and human driving characteristics to improve the safety, reliability, and naturalness of autonomous driving systems.

[0008] The technical solution adopted by the present application to solve its technical problems is: In a first aspect, the present application provides an automatic driving decision control method, comprising: According to the vehicle state of the positioning input, an initial behavior decision tree is generated using a Markov algorithm; According to human driving characteristics, the sequence of the initial behavior decision tree is pruned and reconstructed to obtain a pre-screened behavior decision tree; Through the RSS model, unsafe behavior decision sequences in the pre-screened behavior decision tree are pruned and removed, and in combination with the vehicle kinematics model and the behavior decision tree, the vehicle candidate trajectory corresponding to each behavior sequence in the vehicle decision tree is iteratively generated; The sequence of vehicle candidate trajectories is generated to obtain a predicted trajectory; The predicted trajectory is checked for safety speed and collision, and behavior sequences that do not meet the safety speed and collision checks are removed to obtain a pre-selected trajectory; The overall cost of each pre-selected trajectory is calculated, and the behavior sequence with the smallest overall cost is selected as the optimal vehicle decision result and output.

[0009] As a further improvement of the application, the vehicle state according to the positioning input is used to generate an initial behavior decision tree by Markov algorithm, including: Obtaining the vehicle state of the vehicle; Generating an initial behavior decision tree by Markov method based on the vehicle state, and constructing the action set of the vehicle decision process as A={LLC, LK, RLC}, wherein LLC represents left lane change, LK represents keeping the original lane, and RLC represents right lane change.

[0010] As a further improvement of the application, the sequence of the initial behavior decision tree is pruned and reconstructed according to the human driving characteristics to obtain a pre-screened behavior decision tree, including: Pruning the behavior sequence of the vehicle change behavior by the human driving characteristics, and pruning all unreasonable behavior sequences in the total behavior decision tree; the unreasonable behavior includes left and right lane changes at the same time and sequences of multiple changes of vehicle behavior; After driving for one decision cycle in the optimal decision sequence, reconstructing the behavior decision tree as a pre-screened behavior decision tree; Among them, the human driving characteristics include: Heuristic rule one: two behavior changes are not allowed in a single decision time domain; Heuristic rule two: left lane change to right lane change or right lane change to left lane change is not allowed in a single decision time domain.

[0011] As a further improvement of the application, the RSS model is divided into longitudinal safety distance and lateral safety distance; Longitudinal safety distance d min It is the distance that can still avoid collision in the worst case, specifically:

[0012] Among them, v r is the speed of the rear vehicle, ρ is the reaction time, a max is the maximum acceleration, β min is the comfortable deceleration, v f is the speed of the front vehicle, β max Maximum deceleration; Lateral safety distance: if the two vehicles c1, c2 running at lateral velocities v1, v2, and if the two vehicles exert the maximum lateral acceleration to each other within the time interval [0, ρ], and then the two vehicles will exert the minimum lateral deceleration to brake until the relative lateral velocity is zero, in this case, the collision can be avoided, and the lateral safety distance is specifically:

[0013] wherein v1 is the right side vehicle speed, v2 is the left side vehicle speed, ρ is the time interval, β 1,lat,min is the minimum lateral deceleration of the right side vehicle, β 2,lat,min is the minimum lateral deceleration of the left side vehicle, a 1,max is the maximum lateral acceleration of the right side vehicle, a 2,max is the maximum lateral acceleration of the left side vehicle, µ is the lateral distance between the last two vehicles.

[0014] As a further improvement of the present application, the trajectory generation on the sequence of vehicle candidate trajectories to obtain the predicted trajectory, comprising: Based on the sequence of vehicle candidate trajectories, the pure tracking model, the vehicle kinematics model and the intelligent driver IDM model are used to generate the predicted trajectory.

[0015] As a further improvement of the present application, the safety speed check on the predicted trajectory, comprising: According to the safety distance of the vehicle and the front vehicle, the upper limit of the speed of the vehicle is obtained; according to the safety distance of the vehicle and the rear vehicle, the lower limit of the speed of the vehicle is obtained; If the vehicle speed exceeds the safety speed upper limit, adjust the vehicle speed to the safety speed interval, reduce the speed of each trajectory point in the candidate trajectory corresponding to the behavior sequence; If the vehicle speed is lower than the safety speed lower limit, adjust the vehicle speed to the safety speed interval, increase the speed of each trajectory point in the candidate trajectory corresponding to the behavior sequence.

[0016] As a further improvement of the present application, the collision check, comprising: After obtaining the trajectory prediction of the surrounding important vehicles and the candidate trajectory of the vehicle according to the input of the prediction module, the collision safety check is performed by judging whether there will be a collision risk between the vehicle and the surrounding important vehicles at each time in the decision time domain, marking the behavior sequence that does not pass the collision safety detection as an unfeasible behavior sequence, and eliminating the unfeasible behavior sequence from the behavior decision tree.

[0017] As a further improvement of the present application, the function of the total cost is represented as:

[0018] Je (τ) is that the vehicle needs to travel as much as possible according to the global optimal path, and there is a lane changing cost when the vehicle needs to change lanes in special circumstances; Js (τ) is the consistency of high decision, which encourages the current behavior sequence to be similar to the behavior sequence in the last decision time domain; Jn (τ) is the safety of vehicle travel, which is the most basic and important requirement; λ1, λ2 and λ3 are weight coefficients.

[0019] As a further improvement of the application, the behavior sequence with the minimum overall cost is taken as the optimal vehicle decision result and output. By calculating the overall cost of each behavior sequence, the behavior sequence with the minimum overall cost is taken as the optimal vehicle behavior sequence.

[0020] In the formula, J total (τ) is the overall cost.

[0021] In a second aspect, the application provides an automatic driving decision control system, comprising: A decision tree generation module is configured to generate an initial behavior decision tree by using a Markov algorithm according to a positioning input vehicle state. A sequence pruning module is configured to prune and reconstruct the sequences of the initial behavior decision tree according to human driving characteristics to obtain a pre-screened behavior decision tree. A sequence elimination module is configured to prune and eliminate unsafe behavior decision sequences in the pre-screened behavior decision tree by using an RSS model, and then combine a vehicle kinematics model and the behavior decision tree to iteratively generate vehicle candidate trajectories corresponding to each behavior sequence in the vehicle decision tree. A trajectory generation module is configured to generate trajectories for the sequences of the vehicle candidate trajectories to obtain predicted trajectories. The sequence elimination module is configured to perform safety speed checking and collision checking on the predicted trajectories, and eliminate behavior sequences that do not meet the safety speed checking and collision checking to obtain preselected trajectories. A decision output module is configured to calculate the overall cost of each preselected trajectory, and take the behavior sequence with the minimum overall cost as the optimal vehicle decision result and output.

[0022] In a third aspect, the application provides an electronic device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the automatic driving decision control method when executing the computer program.

[0023] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the automatic driving decision control method.

[0024] In a fifth aspect, the present application provides a computer program product, which comprises computer instructions, and the computer instructions instruct a computer to execute the automatic driving decision control method.

[0025] The present application has the beneficial effects that: by introducing human-like thinking for decision tree pruning, the unreasonable sequence is reduced, and the algorithm running efficiency is greatly improved; by introducing the RSS safety evaluation model for evaluation, and performing speed and collision checking on the result, the safety and reliability of the decision are improved; through the multiple safety checking mechanism (RSS model evaluation, speed checking, collision checking), the safety and reliability of the automatic driving decision are ensured; by simulating the thinking way of human drivers, the decision result is more natural, and the driving experience is improved; through the design of the cost function, the global path driving, decision consistency and driving safety are balanced, so that the decision is more reasonable. Compared with the prior art, the automatic driving decision control method provided by the present application can effectively solve the problems of lack of multiple safety mechanisms in the existing decision algorithm, low decision efficiency and failure to simulate the thinking way of human drivers, and realize safer, more efficient and more natural automatic driving decision control. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is the flowchart of the automatic driving decision control method in the embodiment of the present application; Figure 2 is the MDP decision flowchart in the embodiment of the present application; Figure 3 is the pruning and reconstruction flowchart in the embodiment of the present application; Figure 4 is the overall decision flowchart in the embodiment of the present application; Figure 5 is an automatic driving decision control system provided by the present application; Figure 6 is a schematic diagram of an electronic device provided by the present application. DETAILED DESCRIPTION

[0027] The technical solutions of the present application will be described clearly and completely below by means of embodiments in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0028] Embodiment one As Figure 1 shown, the embodiment of the application provides an automatic driving decision control method, comprising: S101, according to the vehicle state input by positioning, an initial behavior decision tree is generated by using Markov algorithm; Based on Markov decision process, the vehicle state (position, speed, environment perception, etc.) is abstracted as state space, and all possible driving behavior sequences (such as lane changing, accelerating, decelerating, etc.) are generated through state transition probability to form an initial decision tree. This method has the advantages of systematic exhaustive driving possibilities, avoiding missing key scenarios, and providing complete decision space for subsequent optimization.

[0029] Among them, Markov decision process (Markov Decision Process, MDP) is a core mathematical framework in reinforcement learning and dynamic decision problem, which is used to model the optimization problem of agent (Agent) achieving the goal through sequential decision making in uncertain environment.

[0030] S102, according to the sequence of the initial behavior decision tree, the behavior decision tree is pruned and reconstructed according to the characteristics of human driving, and a pre-screened behavior decision tree is obtained; Through pruning and reconstruction, the behavior sequence that does not conform to the driving habits of human beings (such as frequent emergency braking and unreasonable lane changing) is removed, and the decision path that conforms to the traffic rules and driving common sense is retained. This step improves the humanization of decision making and reduces the risk caused by the inconsistency between machine decision and human expectation (such as rear-end collision or misunderstanding of pedestrians).

[0031] S103, through the RSS model, the unsafe behavior decision sequence in the pre-screened behavior decision tree is pruned and removed, and then combined with the vehicle kinematics model and the behavior decision tree, the vehicle candidate trajectory corresponding to each behavior sequence in the vehicle decision tree is iteratively generated; RSS (Responsibility - Sensitive Safety) model is a model used to evaluate the safety of vehicle driving, which defines the safety distance that the vehicle should maintain under various conditions. The RSS model is divided into longitudinal safety distance and lateral safety distance.

[0032] RSS safety model: define safety distance threshold (such as longitudinal / lateral safety distance), remove behavior sequence that may cause collision. Vehicle kinematics model: combined with vehicle dynamics parameters (such as acceleration limit, turning radius), generate physically feasible candidate trajectory. Double protection of safety, both avoid dangerous decision in logic, and ensure that the trajectory conforms to the physical characteristics of the vehicle.

[0033] S104, trajectory generation is performed on the sequence of vehicle candidate trajectories to obtain predicted trajectories; S105, safety speed check and collision check are performed on the predicted trajectory, and the behavior sequence that does not meet the safety speed check and collision check is removed, to obtain a pre-selected trajectory; Based on the candidate behavior sequence, multiple predicted trajectories (such as different speed curves and path planning) are generated. The pre-selected trajectories are screened through speed check (such as overspeed identification) and collision check (such as time collision risk assessment). Redundant solutions can be provided to enhance system fault tolerance and adapt to complex dynamic environments.

[0034] S106, the overall cost of each pre-selected trajectory is calculated, and the behavior sequence with the minimum overall cost is taken as the optimal vehicle decision result and output.

[0035] A multi-objective cost function (such as time efficiency, energy consumption, comfort, and safety weight) is defined, the overall cost of each pre-selected trajectory is calculated, and the trajectory with the minimum cost is selected. Multi-objective balance can be achieved to avoid suboptimal decisions caused by single indicator optimization (such as excessive pursuit of speed at the expense of safety).

[0036] Therefore, the method of the present application significantly reduces the risk of accidents through double filtering of RSS model and collision check. Combined with the kinematic model, dangerous trajectories that are physically infeasible are avoided. Pruning and reconstruction make the decision consistent with human driving habits, improving passenger trust and road cooperation. From decision tree generation to optimal output, a closed loop is formed, covering the entire driving chain logic. The cost function is flexible to adapt to different scene requirements (such as high-speed priority efficiency and urban priority safety). The multi-trajectory generation mechanism enhances the ability to respond to sudden situations (such as sudden appearance of obstacles). Through multi-level screening and optimization of decision sequences, the safety and rationality of automatic driving decisions are improved.

[0037] Therefore, the method of the present application balances rational decision-making of the automatic driving system and perceptual cognition of human driving through a hierarchical optimization mechanism, providing technical support for safe and efficient driving in complex scenarios. Application scenarios include: Urban congestion section: the system preferentially selects a smooth following trajectory to avoid frequent start-stop causing passenger discomfort.

[0038] Highway overtaking: the RSS model ensures the overtaking distance, and the kinematic model generates a smooth lane-changing trajectory.

[0039] Emergency obstacle avoidance: multiple avoidance trajectories are quickly generated, and the path with the shortest time and safety is selected through the cost function.

[0040] Specifically, as shown in Figure 4 each step of the present application is described in detail.

[0041] S101, according to the vehicle state input by positioning, a Markov algorithm is used to generate an initial behavior decision tree, including: obtaining a vehicle state of a vehicle; generating an initial behavior decision tree based on the vehicle state by a Markov method, as shown in Figure 2 and constructing an action set of the vehicle decision process as A={LLC, LK, RLC}, where LLC represents changing lanes to the left, LK represents keeping the original lane, and RLC represents changing lanes to the right.

[0042] In this embodiment, the vehicle state includes information such as the position, speed, acceleration, heading angle, etc. of the vehicle. These information can be obtained by various sensors on the vehicle, such as GPS (Global Positioning System), IMU (Inertial Measurement Unit), wheel encoder, etc. The vehicle state information is the basic data for the automatic driving system to make decisions.

[0043] The Markov algorithm is a decision-making method based on a probability model, which assumes that the next state of the system is only related to the current state, and has nothing to do with the previous state. In the automatic driving decision control, the Markov algorithm can predict the possible behavior sequence in the future according to the current vehicle state, and build a decision tree.

[0044] The initial behavior decision tree is a multi-level tree structure, each node represents a possible behavior decision, and each path from the root node to the leaf node represents a complete behavior sequence. In this embodiment, the depth of the behavior decision tree is 3, i.e. considering the behavior sequence of 3 decision periods in the future. The length of each decision period is 1 second, so the entire decision time domain is 3 seconds.

[0045] The action set A={LLC, LK, RLC} defines three basic behaviors that the vehicle can take in each decision period: changing lanes to the left (LLC), keeping the original lane (LK), and changing lanes to the right (RLC). Based on these three basic behaviors, a complete behavior decision tree can be constructed. For example, for a decision tree with a depth of 3, there are a total of 27 possible behavior sequences, such as {LLC, LLC, LLC}, {LLC, LLC, LK}, {LLC, LLC, RLC}, etc.

[0046] S102, pruning and reconstructing the sequence of the initial behavior decision tree according to human driving characteristics to obtain a pre-screened behavior decision tree, as shown in Figure 3 includes: Pruning the behavior sequence of changing vehicle behavior by human driving characteristics, pruning all unreasonable behavior sequences in the total behavior decision tree; unreasonable behaviors include simultaneous left and right lane changes and multiple changes in vehicle behavior sequences; After driving one decision period with the optimal decision sequence, the behavior decision tree is reconstructed as the pre-screened behavior decision tree; wherein the human driving characteristics include: Heuristic rule one: no two behavior changes are allowed in a single decision time domain; Heuristic rule two: no left-lane change to right-lane change or right-lane change to left-lane change change is allowed in a single decision time domain.

[0047] In real-world driving, human drivers usually follow some driving habits and rules, which can be used to optimize the decision-making process of autonomous driving systems. By simulating human driving characteristics, some behavior sequences that do not conform to actual driving habits can be pruned, thereby reducing computational load and improving the rationality of decision-making.

[0048] Heuristic rule one specifies that no two behavior changes are allowed in a single decision time domain. For example, the behavior sequence {LLC, LK, LLC} contains two behavior changes (from left-lane change to keep the original lane, and then from keeping the original lane to left-lane change) in a single decision time domain, which will be pruned. This is because frequent behavior changes not only increase driving difficulty, but also may cause vehicle control instability and increase the risk of accidents.

[0049] Heuristic rule two specifies that no left-lane change to right-lane change or right-lane change to left-lane change change is allowed in a single decision time domain. For example, the behavior sequence {LLC, RLC, LK} contains a behavior change from left-lane change to right-lane change directly, which will also be pruned. This is because in real-world driving, vehicles usually need to return to the original lane first, and then perform another direction of lane change. Directly changing from left-lane change to right-lane change (or from right-lane change to left-lane change) does not conform to safe driving standards.

[0050] By applying these two heuristic rules, the initial 27 behavior sequences can be reduced to fewer reasonable behavior sequences. For example, the behavior sequences {LK, LK, LK}, {LLC, LLC, LLC}, {RLC, RLC, RLC}, {LK, LLC, LLC}, {LK, RLC, RLC}, etc. that conform to human driving characteristics may be left.

[0051] After driving one decision period with the optimal decision sequence, the system will reconstruct the behavior decision tree. This is because as the vehicle state changes, the optimal decision sequence also needs to be updated continuously. By periodically reconstructing the behavior decision tree, the real-time and adaptability of decision-making can be ensured.

[0052] S103, pruning and removing unsafe behavior decision sequence in the pre-screened behavior decision tree through the RSS model, and then combining the vehicle kinematics model and the behavior decision tree to iteratively generate vehicle candidate trajectories corresponding to each behavior sequence in the vehicle decision tree.

[0053] Longitudinal safety distance d min The distance that can avoid collision in the worst case, specifically:

[0054] Wherein, v r is the speed of the rear vehicle, ρ is the reaction time, a max is the maximum acceleration, β min is the comfortable deceleration, v f is the speed of the front vehicle, β max maximum deceleration; Lateral safety distance: if the two vehicles c1 and c2 traveling at lateral velocities v1 and v2 apply the maximum lateral acceleration to each other within the time interval [0, ρ], and then the two vehicles will brake with the minimum lateral deceleration until the relative lateral velocity is zero, in this case, the lateral safety distance can avoid collision, and the lateral safety distance is specifically:

[0055] Wherein, v1 is the speed of the right vehicle, v2 is the speed of the left vehicle, ρ is the time interval, β 1,lat,min is the minimum lateral deceleration of the right vehicle, β 2,lat,min is the minimum lateral deceleration of the left vehicle, a 1,max is the maximum lateral acceleration of the right vehicle, a 2,max is the maximum lateral acceleration of the left vehicle, µ is the lateral distance between the last two vehicles.

[0056] Through the RSS model, the safety of each behavior sequence can be evaluated. If a behavior sequence cannot meet the safety distance requirement at any time, the behavior sequence will be pruned. For example, if the behavior sequence {LK, LLC, LLC} will cause the distance between the vehicle and the front vehicle to be less than the longitudinal safety distance, or the distance between the adjacent lane vehicle to be less than the lateral safety distance during execution, the behavior sequence will be pruned.

[0057] After pruning the unsafe behavior sequences, the system generates a corresponding vehicle candidate trajectory for each remaining behavior sequence, in combination with a vehicle kinematic model. The vehicle kinematic model describes the motion characteristics of the vehicle, including the relationship between parameters such as speed, acceleration, steering angle, etc. Through the vehicle kinematic model, the motion trajectory of the vehicle can be predicted according to the behavior sequence.

[0058] The vehicle kinematic model adopts a bicycle model, which simplifies the vehicle as a two-dimensional model with a front wheel and a rear wheel. The state variables of the model include the position (x, y) of the vehicle, the heading angle θ, the speed v, and the steering angle δ. The state update equations are as follows: x(t+1) = x(t) + v(t)·cos(θ(t))·Δt y(t+1) = y(t) + v(t)·sin(θ(t))·Δt θ(t+1) = θ(t) + v(t)·tan(δ(t)) / L·Δt v(t+1) = v(t) + a(t)·Δt where L is the wheelbase of the vehicle, a(t) is the acceleration, and Δt is the time step.

[0059] For each behavior sequence, the system iteratively calculates the state of the vehicle at each time period based on the vehicle kinematic model, thereby generating a complete candidate trajectory. For example, for the behavior sequence {LK, LK, LK}, the system generates a trajectory that remains in the current lane; for the behavior sequence {LLC, LLC, LLC}, the system generates a trajectory that changes lanes to the left.

[0060] S104, trajectory generation is performed on the sequence of vehicle candidate trajectories to obtain a predicted trajectory, including: Based on the sequence of vehicle candidate trajectories, a pure pursuit model and a vehicle kinematic model and IDM are used to generate a predicted trajectory.

[0061] After generating the vehicle candidate trajectories, the system needs to further refine these trajectories to generate more accurate predicted trajectories. This process uses a combination of a pure pursuit model, a vehicle kinematic model, and an intelligent driver model (IDM, Intelligent Driver Model).

[0062] The pure pursuit model is a simple and effective path following controller that calculates the required steering angle based on the lateral deviation between the current position of the vehicle and the target path. The control law of the pure pursuit model is as follows: δ = arctan(2·L·e / (Kp·v 2 )) where δ is the steering angle, L is the wheelbase of the vehicle, e is the lateral deviation, Kp is the control gain, and v is the vehicle speed.

[0063] In this embodiment, the control gain Kp of the pure pursuit model is set to 1.0, and the wheelbase L of the vehicle is 2.8 meters.

[0064] IDM is a model used to simulate the longitudinal motion of a vehicle, which takes into account the relative distance and relative speed between the vehicle and the preceding vehicle to calculate the required acceleration. The acceleration calculation formula of IDM is as follows: a = a max ·[1 - (v / v0) 4 - (s* / s) 2 ] where a is the acceleration, a max is the maximum acceleration, v is the current speed, v0 is the desired speed, s is the actual distance to the preceding vehicle, and s* is the desired safety distance. The calculation formula of the desired safety distance s* is as follows: s* = s0 + v·T + v·Δv / (2· ) where s0 is the minimum safety distance, T is the safety time interval, and Δv is the relative speed to the preceding vehicle.

[0065] In this embodiment, the parameters of IDM are set as follows: the maximum acceleration a max is 3.0 m / s 2 , the desired speed v0 is determined according to the road speed limit, the minimum safety distance s0 is 2.0 meters, and the safety time interval T is 1.5 seconds.

[0066] Through the combination of the pure pursuit model, the vehicle kinematics model, and IDM, the system can generate accurate prediction trajectories that take into account path tracking, vehicle dynamics, and traffic flow characteristics. These prediction trajectories will be used for subsequent safety checks and decision optimization.

[0067] S105, safety speed checking is performed on the prediction trajectories, including: According to the safety distance between the vehicle and the preceding vehicle, the upper limit of the speed at which the vehicle travels is obtained; according to the safety distance between the vehicle and the following vehicle, the lower limit of the speed at which the vehicle travels is obtained; If the vehicle speed exceeds the upper limit of the safety speed, adjust the vehicle speed to be within the safety speed interval, and reduce the speed of each trajectory point in the candidate trajectory corresponding to the behavior sequence; If the vehicle speed is lower than the lower limit of the safety speed, adjust the vehicle speed to be within the safety speed interval, and increase the speed of each trajectory point in the candidate trajectory corresponding to the behavior sequence.

[0068] Safety speed check is an important step to ensure the safety of vehicle driving. By calculating the upper and lower limits of safety speed, the system can determine whether the current speed is within the safe range and adjust it if necessary.

[0069] The upper limit of safety speed is calculated based on the safety distance between the vehicle and the front vehicle. If the distance between the vehicle and the front vehicle is d front , then the upper limit of safety speed v max can be calculated by the following formula: v max =

[0070] where βmin is the comfortable deceleration and s0 is the minimum safety distance.

[0071] The lower limit of safety speed is calculated based on the safety distance between the vehicle and the rear vehicle. If the distance between the vehicle and the rear vehicle is d rear , then the lower limit of safety speed v min can be calculated by the following formula: v min = v rear

[0072] where v rear is the speed of the rear vehicle, βmax is the maximum deceleration, and s0 is the minimum safety distance.

[0073] In this embodiment, the minimum safety distance s0 is set to 2.0 meters, the comfortable deceleration βmin is 2.0 m / s 2 , and the maximum deceleration βmax is 8.0 m / s 2 .

[0074] If the speed in the predicted trajectory of the vehicle exceeds the upper limit of safety speed, the system will reduce the speed at each point in the trajectory so that it does not exceed the upper limit of safety speed. For example, if the speed of a certain trajectory point is 25 m / s, and the upper limit of safety speed is 20 m / s, the system will adjust the speed of this point to 20 m / s.

[0075] Similarly, if the speed in the predicted trajectory of the vehicle is lower than the lower limit of safety speed, the system will increase the speed at each point in the trajectory so that it does not fall below the lower limit of safety speed. For example, if the speed of a certain trajectory point is 10 m / s, and the lower limit of safety speed is 15 m / s, the system will adjust the speed of this point to 15 m / s.

[0076] Through safety speed check, the system can ensure that the speed of the vehicle is always within the safe range, avoiding the risk of collision caused by improper speed.

[0077] In S105, collision check, including: After obtaining the trajectory predictions of important surrounding vehicles and the candidate trajectories of the vehicle based on the input of the prediction module, a collision safety check is performed by judging whether there is a collision risk between the vehicle and the important surrounding vehicles at each moment in the decision time domain. Behavior sequences that fail the collision safety check are marked as infeasible behavior sequences and are eliminated from the behavior decision tree.

[0078] Collision detection is another important step in ensuring driving safety. By comparing the vehicle's predicted trajectory with those of surrounding vehicles, the system can determine whether there is a collision risk and eliminate behavior sequences that may lead to a collision.

[0079] In this embodiment, collision detection uses a rectangular bounding box approach. The vehicle is simplified to a rectangle with a length of 4.8 meters and a width of 1.8 meters. At each time step, the system calculates the position of the vehicle and surrounding significant vehicles and checks whether their bounding boxes overlap. If overlap occurs at any time step, a collision risk is considered present.

[0080] Specifically, let the coordinates of the four vertices of the rectangular bounding box of vehicle A be (x A1 , y A1 )、(x A2 , y A2 )、(x A3 ,y A3 )、(x A4 , y A4 ), the coordinates of the four vertices of the rectangular bounding box of vehicle B are (x B1 , y B1 )、(x B2 , y B2 )、(xB3, y B3 ), (xB4, yB4). Two rectangular bounding boxes are considered non-overlapping if any of the following conditions is met: 1) The rightmost boundary of vehicle A is smaller than the leftmost boundary of vehicle B: max(x A1 , x A2 , x A3 , x A4 ) <min(x B1 ,x B2 , x B3 , x B4 ) 2) The leftmost boundary of vehicle A is greater than the rightmost boundary of vehicle B: min(x A1 , x A2 , x A3 , x A4 )>max(x B1 ,x B2 , xB3 , x B4 ) 3) The uppermost boundary of vehicle A is less than the lowermost boundary of vehicle B: max(y A1 , y A2 , y A3 , y A4 )<min(y B1 ,y B2 , y B3 , y B 4 ) 4) The lowermost boundary of vehicle A is greater than the uppermost boundary of vehicle B: min(y A1 , y A2 , y A3 , y A4 )>max(y B1 ,y B2 , y B3 , y B4 ) If none of the above four conditions are met, it is considered that the two rectangular bounding boxes overlap, and there is a risk of collision.

[0081] For each predicted trajectory corresponding to a behavior sequence, the system performs a collision check at each time step within the decision-making horizon. If a collision risk is detected at any time step, the behavior sequence is marked as infeasible and is removed from the behavior decision tree.

[0082] Through collision checking, the system can ensure that the selected behavior sequence will not result in a collision with surrounding vehicles, thereby ensuring driving safety.

[0083] S106, calculate the overall cost of each pre-selected trajectory, and output the behavior sequence with the minimum overall cost as the optimal vehicle decision result.

[0084] The function of the overall cost is represented as: J total (τ) = λ1·Je(τ) + λ2·Js(τ) + λ3·Jn(τ) Je(τ) is the cost of the vehicle needing to travel as much as possible along the globally optimal path, and there is also a lane-changing cost when the vehicle needs to change lanes in special circumstances; Js(τ) is the consistency of high decision-making, encouraging the current behavior sequence to be similar to the behavior sequence in the previous decision-making horizon; Jn(τ) is the safety of vehicle travel, which is the most basic and important requirement, and a safety distance model is used to measure the risk of collision; λ1, λ2, λ3 are weight coefficients.

[0085] After the safety speed check and collision check are completed, the system obtains a series of pre-selected trajectories, which are all safe and feasible. Next, the system needs to select the optimal one from these pre-selected trajectories as the final decision result. The selection criterion is the overall cost function, which considers multiple factors.

[0086] Je(τ) is the path cost, which measures the degree of deviation of the vehicle's travel path from the global optimal path. The global optimal path is usually generated by the path planning module, considering factors such as road network, traffic rules, and destination. The calculation formula of the path cost is: Je(τ) = wd·d lat + w θ ·|θ - θ ref | + w c ·c lane where d lat is the lateral deviation of the vehicle from the reference path, θ is the vehicle heading angle, θ ref is the heading angle of the reference path, c lane is the lane changing cost (if the behavior sequence contains lane changing behavior, c lane is 1, otherwise 0), w d , w θ , and wc are weight coefficients.

[0087] In this embodiment, the weight coefficients of the path cost are set as follows: w d = 1.0, w θ = 0.5, and w c = 2.0.

[0088] Js(τ) is the decision consistency cost, which encourages the current behavior sequence to be similar to the behavior sequence in the last decision time domain, avoiding frequent changes in decisions. The calculation formula of the decision consistency cost is: Js(τ) = w s ·d s (τ, τ prev ) where d s (τ, τ prev ) is the difference degree between the current behavior sequence τ and the behavior sequence τ prev in the last decision time domain, and w s is the weight coefficient. The calculation method of the difference degree d s is: d s (τ, τ prev ) = Σi I(τ i ≠ τ previ ) / n where I(·) is an indicator function, 1 if the condition is true, otherwise 0; n is the length of the behavior sequence; τ i and τ previ are the behaviors at the i-th position in the current behavior sequence and the behavior sequence in the last decision time domain, respectively.

[0089] In this embodiment, the weight coefficient ws of the decision consistency cost is set to 3.0.

[0090] Jn(τ) is the safety cost, which measures the safety of the vehicle driving, and uses the safety distance model to calculate the collision risk. The calculation formula of the safety cost is: Jn(τ) = w front ·max(0, d safefront- d front ) + w rear ·max(0, d saferear- d rear ) +w lat ·max(0, d safelat- d lat ) where d safefront , d saferear and d safelat are the safety distances from the front vehicle, the rear vehicle and the side vehicle, respectively, d front , d rear and d lat are the actual distances from the front vehicle, the rear vehicle and the side vehicle, respectively, and w front , w rear and w lat are weight coefficients.

[0091] In this embodiment, the weight coefficients of the safety cost are set as follows: w front = 2.0, w rear = 1.0, and w lat = 1.5.

[0092] The weight coefficients λ1, λ2 and λ3 of the overall cost function are set to 0.3, 0.2 and 0.5, respectively, reflecting the importance of safety in the decision-making process.

[0093] The behavior sequence with the minimum overall cost is taken as the optimal vehicle decision result and output, including: By calculating the overall cost of each behavior sequence, the behavior sequence with the minimum overall cost is taken as the optimal vehicle behavior sequence: τ* =

[0094] where J total (τ) is the overall cost.

[0095] After calculating the total cost of each pre-selected trajectory, the system selects the behavior sequence with the minimum total cost as the optimal vehicle decision result. This process can be expressed as a minimization problem: τ* =

[0096] Among them, τ* is the optimal behavior sequence, J total (τ) is the total cost of the behavior sequence τ.

[0097] For example, suppose that after the safe speed check and collision check, the remaining preselected trajectories correspond to the behavior sequences {LK, LK, LK}, {LLC, LLC, LLC}, and {RLC, RLC, RLC}, with overall costs of 5.2, 6.8, and 4.5, respectively. The system then selects the behavior sequence {RLC, RRC, RRC} with the lowest overall cost as the optimal vehicle decision.

[0098] The optimal vehicle decision result will be output to the control module, which will generate specific control instructions based on this decision result, such as steering angle, acceleration, etc., to control the vehicle to drive according to the decision result.

[0099] In practice, the system re-executes the decision-making process every decision cycle (typically 0.1 seconds), updating the decision results based on the latest vehicle status and environmental information. This rolling optimization approach ensures real-time and adaptable decision-making.

[0100] Example 2 like Figure 5 As shown, an embodiment of the present invention provides an automatic driving decision control system, including: The decision tree generation module 100 is used to generate an initial behavior decision tree using a Markov algorithm according to the vehicle state input by the positioning; A sequence pruning module 200 is used to prune and reconstruct the sequence of the initial behavior decision tree according to human driving characteristics to obtain a pre-screened behavior decision tree; The sequence elimination module 300 is used to prune and eliminate unsafe behavior decision sequences in the pre-screened behavior decision tree using the RSS model, and then iteratively generate candidate vehicle trajectories corresponding to each behavior sequence in the vehicle decision tree by combining the vehicle kinematic model and the behavior decision tree; The trajectory generation module 400 is used to generate a trajectory for a sequence of candidate vehicle trajectories to obtain a predicted trajectory; A sequence elimination module 500 is used to perform a safety speed check and a collision check on the predicted trajectory, and eliminate the behavior sequences that do not meet the safety speed check and the collision check to obtain a preselected trajectory; The decision output module 600 is configured to calculate the overall cost of each preselected trajectory, and output the behavior sequence with the minimum overall cost as the optimal vehicle decision result.

[0101] In the embodiment, the automatic driving decision control system is a software system composed of multiple functional modules, which can be deployed on a vehicle-mounted computing platform, such as a high-performance computer or a dedicated automatic driving computing unit. The modules of the system communicate with each other through a data interface to jointly complete the automatic driving decision control task.

[0102] The decision tree generation module is the starting point of the system, which receives vehicle state information from the positioning module, including position, speed, acceleration, heading angle, etc. Based on this information, the decision tree generation module uses the Markov algorithm to generate an initial behavior decision tree. The Markov algorithm assumes that the next state of the system is only related to the current state, and has nothing to do with the previous state. This feature makes it suitable for automatic driving decision control.

[0103] The initial behavior decision tree is a multi-level tree structure, each node represents a possible behavior decision, and each path from the root node to the leaf node represents a complete behavior sequence. In the embodiment, the depth of the behavior decision tree is 3, i.e. considering the behavior sequence of the next 3 decision periods. The length of each decision period is 1 second, so the entire decision time domain is 3 seconds.

[0104] The action set of the vehicle decision process constructed by the decision tree generation module is A={LLC, LK, RLC}, where LLC represents left lane change, LK represents keeping the original lane, and RLC represents right lane change. Based on these three basic behaviors, the decision tree generation module can construct a complete behavior decision tree. For example, for a decision tree with a depth of 3, there are a total of 27 possible behavior sequences.

[0105] The sequence pruning module receives the initial behavior decision tree from the decision tree generation module and prunes and reconstructs it according to human driving characteristics. Human driving characteristics include two heuristic rules: one is that no two behavior changes are allowed in a single decision time domain, and the other is that left lane change to right lane change or right lane change to left lane change is not allowed in a single decision time domain.

[0106] By applying these two heuristic rules, the sequence pruning module can reduce the initial 27 behavior sequences to fewer reasonable behavior sequences. For example, the behavior sequence {LLC, LK, LLC} contains two behavior changes in a decision time domain and will be pruned; the behavior sequence {LLC, RLC, LK} contains a behavior change from left lane change to right lane change directly, and will also be pruned.

[0107] The sequence pruning module receives the pre-screened behavior decision tree from the sequence pruning module and prunes unsafe behavior decision sequences through the RSS model. The RSS model is divided into longitudinal safety distance and lateral safety distance to evaluate the safety of vehicle driving.

[0108] Through the RSS model, the sequence pruning module can evaluate the safety of each behavior sequence and prune unsafe behavior sequences.

[0109] After pruning unsafe behavior sequences, the sequence pruning module generates corresponding vehicle candidate trajectories for each remaining behavior sequence in combination with the vehicle kinematics model. The vehicle kinematics model uses a bicycle model that simplifies the vehicle into a two-dimensional model with a front wheel and a rear wheel. Through the vehicle kinematics model, the sequence pruning module can predict the motion trajectory of the vehicle according to the behavior sequence.

[0110] The trajectory generation module receives the vehicle candidate trajectories from the sequence pruning module and further refines these trajectories to generate more accurate prediction trajectories. This process uses a combination of pure pursuit model, vehicle kinematics model, and IDM.

[0111] The pure pursuit model is a simple and effective path tracking controller that calculates the required steering angle based on the lateral deviation between the current position of the vehicle and the target path. IDM is a model used to simulate the longitudinal motion of the vehicle, which considers the relative distance and relative speed between the vehicle and the preceding vehicle to calculate the required acceleration.

[0112] Through the combination of pure pursuit model, vehicle kinematics model, and IDM, the trajectory generation module can generate accurate prediction trajectories that take into account path tracking, vehicle dynamics, and traffic flow characteristics.

[0113] The sequence pruning module is also responsible for safety speed checking and collision checking of the prediction trajectories, and pruning behavior sequences that do not meet safety requirements. Safety speed checking includes calculating the upper and lower limits of safe speed and adjusting the speed of each point in the trajectory to be within the safe range. Collision checking is to compare the predicted trajectory of the vehicle with the predicted trajectory of surrounding vehicles to determine whether there is a risk of collision.

[0114] The decision output module receives the pre-selected trajectories from the sequence pruning module, calculates the overall cost of each trajectory, and selects the behavior sequence with the smallest overall cost as the optimal vehicle decision result. The overall cost function considers path cost, decision consistency cost, and safety cost, which correspond to the rationality of the vehicle driving path, the stability of the decision, and the safety of the driving, respectively.

[0115] In practice, the autonomous driving decision-making control system re-executes the decision-making process during each decision cycle, updating the decision results based on the latest vehicle status and environmental information. This rolling optimization approach ensures real-time and adaptable decision-making.

[0116] Example 3 like Figure 6 As shown, an embodiment of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the following autonomous driving decision control method is implemented: According to the vehicle status input by positioning, the Markov algorithm is used to generate the initial behavior decision tree; The sequence of the initial behavior decision tree is pruned and reconstructed according to human driving characteristics to obtain a pre-screened behavior decision tree; The RSS model is used to prune unsafe behavior decision sequences in the pre-screened behavior decision tree. Then, the vehicle kinematic model and the behavior decision tree are combined to iteratively generate candidate vehicle trajectories corresponding to each behavior sequence in the vehicle decision tree. Generate trajectories for the sequence of candidate vehicle trajectories to obtain predicted trajectories; Perform safety speed check and collision check on the predicted trajectory, and eliminate the behavior sequences that do not meet the safety speed check and collision check to obtain the pre-selected trajectory; Calculate the overall cost of each pre-selected trajectory, and output the behavior sequence with the minimum overall cost as the optimal vehicle decision result.

[0117] In this embodiment, the electronic device may be an onboard computer, an autonomous driving control unit, or other computing device capable of executing the autonomous driving decision-making control method. The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor.

[0118] Memory can include various types of storage media, such as read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory, magnetic disks, or optical disks. Memory is used to store operating systems, application programs, and computer programs that implement autonomous driving decision-making and control methods.

[0119] The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, etc. The processor is responsible for executing the computer program stored in the memory, implementing each step of the autonomous driving decision control method.

[0120] The computer program contains a sequence of instructions for implementing the autonomous driving decision control method. When the processor executes these instructions, it will execute each step of the autonomous driving decision control method in the order of the instruction sequence.

[0121] Specifically, when the processor executes the computer program, it first generates an initial behavior decision tree using the Markov algorithm based on the input vehicle state. The vehicle state includes information such as position, speed, acceleration, heading angle, etc., which can be obtained through various sensors on the vehicle. The Markov algorithm predicts possible future behavior sequences based on the current vehicle state and constructs a decision tree.

[0122] The initial behavior decision tree is a multi-level tree structure, with each node representing a possible behavior decision and each path from the root node to the leaf node representing a complete behavior sequence. The depth of the behavior decision tree is 3, i.e., considering behavior sequences for 3 decision periods in the future. Each decision period is 1 second long, so the entire decision time domain is 3 seconds.

[0123] The action set of the vehicle decision process is A = {LLC, LK, RLC}, where LLC represents left lane change, LK represents keeping the original lane, and RLC represents right lane change. Based on these three basic behaviors, a complete behavior decision tree can be constructed. For example, for a decision tree with a depth of 3, there are a total of 27 possible behavior sequences.

[0124] Next, the processor prunes and reconstructs the sequence of the initial behavior decision tree based on human driving characteristics to obtain a pre-screened behavior decision tree. Human driving characteristics include two heuristic rules: one is that no two behavior changes are allowed within a single decision time domain, and the other is that no left lane change to right lane change or right lane change to left lane change change is allowed within a single decision time domain.

[0125] By applying these two heuristic rules, the processor can reduce the initial 27 behavior sequences to fewer reasonable behavior sequences. For example, the behavior sequence {LLC, LK, LLC} contains two behavior changes within a decision time domain and will be pruned; the behavior sequence {LLC, RLC, LK} contains a behavior change from left lane change to right lane change directly and will also be pruned.

[0126] Then, the processor prunes the unsafe behavior decision sequences in the pre-screened behavior decision tree through an RSS model. The RSS model is divided into longitudinal safety distance and lateral safety distance, which is used to evaluate the safety of vehicle driving.

[0127] Through the RSS model, the processor can evaluate the safety of each behavior sequence and eliminate unsafe behavior sequences.

[0128] After eliminating unsafe behavior sequences, the processor generates corresponding vehicle candidate trajectories for each remaining behavior sequence in combination with a vehicle kinematics model. The vehicle kinematics model adopts a bicycle model, which simplifies the vehicle as a two-dimensional model with a front wheel and a rear wheel. Through the vehicle kinematics model, the processor can predict the motion trajectory of the vehicle according to the behavior sequence.

[0129] Next, the processor generates a sequence of vehicle candidate trajectories to obtain a predicted trajectory. This process uses a combination of pure pursuit model, vehicle kinematics model and IDM.

[0130] The pure pursuit model is a simple and effective path following controller that calculates the required steering angle based on the lateral deviation between the current position of the vehicle and the target path. IDM is a model used to simulate the longitudinal motion of the vehicle, which takes into account the relative distance and relative speed between the vehicle and the preceding vehicle to calculate the required acceleration.

[0131] Through the combination of pure pursuit model, vehicle kinematics model and IDM, the processor can generate accurate predicted trajectories that take into account path following, vehicle dynamics and traffic flow characteristics.

[0132] Then, the processor performs safety speed check and collision check on the predicted trajectory and eliminates behavior sequences that do not meet safety requirements. Safety speed check includes calculating the upper and lower limits of safe speed and adjusting the speed at each point in the trajectory to be within the safe range. Collision check is to compare the predicted trajectory of the vehicle with the predicted trajectories of surrounding vehicles to determine whether there is a risk of collision.

[0133] Finally, the processor calculates the overall cost of each pre-selected trajectory and selects the behavior sequence with the smallest overall cost as the optimal vehicle decision result. The overall cost function considers three aspects: path cost, decision consistency cost and safety cost, which correspond to the rationality of the vehicle driving path, the stability of the decision and the safety of the driving, respectively.

[0134] In practical applications, the electronic device will re-execute the decision-making process once every decision-making period to update the decision-making result based on the latest vehicle state and environmental information. This rolling optimization approach ensures the real-time and adaptability of the decision-making.

[0135] Embodiment Four A computer readable storage medium stores a computer program, which, when executed by a processor, implements an automatic driving decision control method as follows: S101, according to the vehicle state input by positioning, an initial behavior decision tree is generated by using Markov algorithm; S102, the sequence of the initial behavior decision tree is pruned and reconstructed according to the characteristics of human driving to obtain a pre-screened behavior decision tree; S103, through the RSS model, the unsafe behavior decision sequence in the pre-screened behavior decision tree is pruned and removed, and then combined with the vehicle kinematics model and the behavior decision tree, the vehicle candidate trajectory corresponding to each behavior sequence in the vehicle decision tree is iteratively generated; S104, the sequence of the vehicle candidate trajectory is generated to obtain a predicted trajectory; S105, the predicted trajectory is checked for safety speed and collision, and the behavior sequence that does not meet the safety speed check and collision check is removed to obtain a pre-selected trajectory; S106, the overall cost of each pre-selected trajectory is calculated, the behavior sequence with the minimum overall cost is taken as the optimal vehicle decision result and output.

[0136] In the embodiment, the computer readable storage medium can be any tangible medium containing or storing a program, such as read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory, magnetic disk or optical disk, etc. The computer readable storage medium stores a computer program containing instruction sequences for implementing the automatic driving decision control method.

[0137] When the computer program in the computer readable storage medium is executed by the processor, the processor will execute each step of the automatic driving decision control method according to the order of the instruction sequences. The specific execution process is the same as described in Embodiment Three, including generating an initial behavior decision tree according to the vehicle state input by positioning, pruning and reconstructing the behavior decision tree, generating a predicted trajectory, performing safety check, and selecting the optimal vehicle decision result.

[0138] The computer readable storage medium can be integrated in a vehicle-mounted computer, an automatic driving control unit or other computing devices capable of executing the automatic driving decision control method. By storing the computer program in the computer readable storage medium, the automatic driving decision control method can be conveniently deployed on different computing devices, improving the portability and reusability of the method.

[0139] Embodiment Five A computer program product comprising computer instructions instructing a computer to execute an automatic driving decision control method as follows: S101, generating an initial behavior decision tree according to the vehicle state input by positioning, using a Markov algorithm; S102, pruning and reconstructing the sequence of the initial behavior decision tree according to human driving characteristics to obtain a pre-screened behavior decision tree; S103, pruning and removing unsafe behavior decision sequences in the pre-screened behavior decision tree through an RSS model, and then combining a vehicle kinematics model and the behavior decision tree to iteratively generate vehicle candidate trajectories corresponding to each behavior sequence in the vehicle decision tree; S104, generating a prediction trajectory by generating a sequence of vehicle candidate trajectories; S105, performing safety speed checking and collision checking on the prediction trajectory, and removing behavior sequences that do not meet the safety speed checking and collision checking to obtain a pre-selected trajectory; S106, calculating the overall cost of each pre-selected trajectory, and selecting the behavior sequence with the minimum overall cost as the optimal vehicle decision result and outputting it.

[0140] In this embodiment, the computer program product comprises computer instructions which can be stored in a computer readable storage medium or transmitted through a network. The computer instructions instruct the computer to execute each step of the automatic driving decision control method.

[0141] When the computer executes these instructions, it will execute each step of the automatic driving decision control method in the order of the instructions. The specific execution process is the same as described in Embodiment Three and Embodiment Four, including generating an initial behavior decision tree according to the vehicle state input by positioning, pruning and removing the behavior decision tree, generating a prediction trajectory, performing safety checking, and selecting the optimal vehicle decision result.

[0142] The computer program product can be provided in the form of a software package, an application, a plug-in or other forms, and can be distributed through various channels such as an application store, an intranet or direct installation. By encapsulating the automatic driving decision control method as a computer program product, it can be easily deployed on different computing devices, improving the accessibility and ease of use of the method.

[0143] It should be noted that Embodiment One, Embodiment Two, Embodiment Three, Embodiment Four and Embodiment Five are all automatic driving decision control methods.

[0144] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit the same; and although the present application has been described in detail with reference to the foregoing embodiments, it should be appreciated by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features thereof can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An automatic driving decision control method, characterized in that: include: According to the vehicle status input by positioning, the Markov algorithm is used to generate the initial behavior decision tree; The sequence of the initial behavior decision tree is pruned and reconstructed according to human driving characteristics to obtain a pre-screened behavior decision tree; The RSS model is used to prune unsafe behavior decision sequences in the pre-screened behavior decision tree. Then, the vehicle kinematic model and the behavior decision tree are combined to iteratively generate candidate vehicle trajectories corresponding to each behavior sequence in the vehicle decision tree. Generate trajectories for the sequence of candidate vehicle trajectories to obtain predicted trajectories; Perform safety speed check and collision check on the predicted trajectory, and eliminate the behavior sequences that do not meet the safety speed check and collision check to obtain the pre-selected trajectory; Calculate the overall cost of each pre-selected trajectory, and output the behavior sequence with the minimum overall cost as the optimal vehicle decision result.

2. The automatic driving decision control method according to claim 1, characterized in that: The initial behavior decision tree is generated using the Markov algorithm based on the vehicle state input by the positioning, including: Get the vehicle status of the vehicle; Based on the vehicle state, an initial behavior decision tree is generated using the Markov method, and the action set of the vehicle decision process is constructed as A={LLC,LK,RLC}, where LLC represents a left lane change, LK represents staying in the original lane, and RLC represents a right lane change.

3. The automatic driving decision control method according to claim 1, characterized in that: The sequence of the initial behavior decision trees is pruned and reconstructed according to human driving characteristics to obtain a pre-screened behavior decision tree, including: The behavior sequences of vehicle changes are pruned based on human driving characteristics, and all unreasonable behavior sequences in the total behavior decision tree are pruned; unreasonable behaviors include sequences with simultaneous left and right lane changes and multiple vehicle behavior changes; After the optimal decision sequence has gone through a decision cycle, the behavior decision tree is reconstructed as the pre-screened behavior decision tree; Among them, human driving characteristics include: Heuristic rule 1: Two behavior changes are not allowed in a single decision domain; Heuristic rule 2: Lane changes from left to right or from right to left are not allowed in a single decision domain.

4. The automatic driving decision control method according to claim 1, characterized in that: The RSS model is divided into longitudinal safety distance and lateral safety distance; Longitudinal safety distance d min It is the distance that can avoid collision in the worst case scenario, specifically: in, v r is the speed of the following vehicle, ρ is the reaction time, a max is the maximum acceleration, β min is the comfortable deceleration, v f is the speed of the preceding vehicle, β max Maximum deceleration; Lateral safety distance: Between cars c1 and c2 traveling at lateral velocities v1 and v2, if both cars apply maximum lateral acceleration and accelerate toward each other within the time interval [0, ρ], and then both cars apply lateral braking with minimum lateral deceleration until the relative lateral velocity reaches zero, both cars can avoid collision in this case. The specific lateral safety distance is: Among them, v1 is the speed of the car on the right, v2 is the speed of the car on the left, ρ is the time interval, β 1,lat,min is the minimum lateral deceleration of the right vehicle, β 2,lat,min is the minimum lateral deceleration of the left vehicle, a 1,max is the maximum lateral acceleration of the right vehicle, a 2,max is the maximum lateral acceleration of the left vehicle, µ is the lateral distance between the last two cars.

5. The automatic driving decision control method according to claim 1, characterized in that: The step of generating a trajectory for a sequence of candidate vehicle trajectories to obtain a predicted trajectory includes: Based on the sequence of candidate vehicle trajectories, a pure tracking model, vehicle kinematic model and intelligent driver IDM model are used to generate rolling prediction trajectories.

6. The automatic driving decision control method according to claim 1, characterized in that: The safety speed check on the predicted trajectory includes: The upper limit of the vehicle's speed is calculated based on the safe distance between the vehicle and the vehicle in front; the lower limit of the vehicle's speed is calculated based on the safe distance between the vehicle and the vehicle behind; If the vehicle speed exceeds the upper limit of the safe speed, adjust the vehicle speed to the safe speed range and reduce the speed of each trajectory point in the candidate trajectory corresponding to the behavior sequence; If the vehicle speed is lower than the safe speed lower limit, adjust the vehicle speed to the safe speed range and increase the speed of each trajectory point in the candidate trajectory corresponding to the behavior sequence.

7. The automatic driving decision control method according to claim 1, characterized in that: The collision check includes: After obtaining the trajectory predictions of important surrounding vehicles and the candidate trajectories of the vehicle based on the input of the prediction module, a collision safety check is performed by judging whether there is a collision risk between the vehicle and the important surrounding vehicles at each moment in the decision time domain. Behavior sequences that fail the collision safety check are marked as infeasible behavior sequences and are eliminated from the behavior decision tree.

8. The automatic driving decision control method according to claim 1, characterized in that: The function of the overall cost is expressed as: Je (τ) indicates that the vehicle must follow the global optimal path as much as possible. When the vehicle needs to change lanes under special circumstances, there is also a lane change cost. Js (τ) indicates high decision consistency, which encourages the current behavior sequence to be similar to the behavior sequence in the previous decision domain. Jn (τ) indicates that vehicle driving safety is the most basic and important requirement. λ1, λ2, and λ3 are weight coefficients. Preferably, the process of taking the behavior sequence with the minimum overall cost as the optimal vehicle decision result and outputting it includes: By calculating the overall cost of each behavior sequence, the behavior sequence with the minimum overall cost is taken as the optimal vehicle behavior sequence: Where, J total (τ) is the total cost.

9. An automatic driving decision control system, characterized in that: include: The decision tree generation module is used to generate an initial behavior decision tree using the Markov algorithm according to the vehicle status input by the positioning; The sequence pruning module is used to prune and reconstruct the sequence of the initial behavior decision tree according to human driving characteristics to obtain a pre-screened behavior decision tree; The sequence elimination module is used to prune unsafe behavior decision sequences in the pre-screened behavior decision tree using the RSS model. It then combines the vehicle kinematic model and the behavior decision tree to iteratively generate candidate vehicle trajectories corresponding to each behavior sequence in the vehicle decision tree. The trajectory generation module is used to generate trajectories for the sequence of candidate vehicle trajectories to obtain predicted trajectories; The sequence elimination module is used to perform safety speed checks and collision checks on the predicted trajectory, and eliminate the behavior sequences that do not meet the safety speed checks and collision checks to obtain the pre-selected trajectory; The decision output module is used to calculate the overall cost of each pre-selected trajectory and output the behavior sequence with the minimum overall cost as the optimal vehicle decision result.

10. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the automatic driving decision control method according to any one of claims 1 to 6 when executing the computer program.

11. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the automatic driving decision control method according to any one of claims 1 to 8.

12. A computer program product comprising computer instructions, characterized in that: The computer instructions instruct the computer to execute the automatic driving decision control method described in any one of claims 1 to 8.

Images