An unmanned vehicle blind area intersection planning method considering uncertainty

Abstract

The application relates to an unmanned vehicle blind area intersection planning method considering uncertainty, which comprises the following steps: step 1, constructing a blind area potential traffic participant model considering the subjective consciousness of a driver; step 2, constructing an uncertainty estimator, jointly reasoning the blind area uncertainty and the visible vehicle intention uncertainty, and generating the probability distribution of the uncertainty factors; and step 3, inputting the uncertainty factor estimation result into a longitudinal velocity planner based on a partially observable Markov model, and obtaining the optimal action strategy of the unmanned vehicle when passing through a non-lamp-controlled blind area intersection. Compared with the prior art, the application has the advantages of high personification degree, good safety and high passing efficiency.

Description

Technical Field

[0001] This invention relates to the field of blind spot intersection planning technology for autonomous vehicles, and in particular to a blind spot intersection planning method for autonomous vehicles that takes into account uncertainties. Background Technology

[0002] In highly complex and coupled urban traffic scenarios, trajectory planning for autonomous driving is a hot research topic. At intersections without traffic lights, autonomous vehicles can only choose their optimal course of action through interaction with other road users, whose intentions are highly uncertain. Furthermore, obstacles in the surrounding environment (vehicles, buildings, etc.) may obstruct the autonomous vehicle's field of vision. The state of potential road users in obstructed areas cannot be obtained by the vehicle's perception module, threatening its driving safety. Therefore, how to plan a safe and reliable trajectory for autonomous vehicles in the presence of blind spots has become a challenge in the field of autonomous vehicle planning.

[0003] Uncontrolled intersections present various uncertainties, one of the main factors affecting autonomous vehicle trajectory planning being the uncertainty of the intentions of surrounding vehicles. Reasonable inference of these intentions is crucial for the safe operation of autonomous vehicles. Some studies, aiming to address the interactive planning problem in adversarial urban scenarios for autonomous driving, combine road environment information with deep neural networks to infer the short-term driving intentions and long-term driving styles of surrounding vehicles, and then utilize partially observable Markov decision processes to select the optimal driving strategy for the autonomous vehicle. Learning-based methods possess inherent human-like characteristics, but suffer from poor scenario transferability and insufficient interpretability. Other studies utilize inverse programming to model vehicle interaction behavior at uncontrolled intersections. Inverse programming offers strong interpretability and can better ensure the safety of intersection traffic.

[0004] The above studies only consider the uncertainty of the intentions of visible vehicles, but do not take into account the impact of potential occlusion factors at intersections. Reachability set-based methods no longer attempt to infer possible traffic conditions in blind spots, but instead establish a vehicle reachability set based on traffic rules and dynamic models to cope with highly dynamic and high-density traffic scenarios, but are prone to providing overly conservative action strategies. Deep learning methods can infer possible traffic conditions in blind spots end-to-end, but data-driven deep neural networks are black-box models and cannot guarantee safety performance.

[0005] For the planning problem of intersections without traffic lights, most existing methods only estimate one of the uncertainties of potential obstacles in blind spots or the uncertainties of the intentions of visible vehicles, and often consider the worst case. However, in real intersection scenarios, the interactions between vehicles are highly coupled and contain a wealth of usable information.

[0006] Existing technologies do not effectively utilize the interaction behavior between vehicles for intersection planning, and the planning results are not accurate enough. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for planning blind spot intersections for unmanned vehicles that takes into account uncertainties, which has a high degree of anthropomorphism, good safety, and high traffic efficiency.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] This invention provides a method for planning blind spot intersections for autonomous vehicles that considers uncertainties. The method includes the following steps:

[0010] Step 1: Construct a model of potential traffic participants in blind spots that takes into account the driver's subjective awareness;

[0011] Step 2: Construct an uncertainty estimator to perform joint reasoning on blind spot uncertainty and visible vehicle intention uncertainty, and generate the probability distribution of uncertainty factors;

[0012] Step 3: Input the uncertainty estimation results into the longitudinal velocity planner based on the partially observable Markov model to obtain the optimal action strategy of the unmanned vehicle when passing through the intersection without traffic lights.

[0013] Preferably, the model of potential traffic participants in the blind spot considering the driver's subjective awareness in step 1 includes the following sub-steps:

[0014] Step 1-1: Determine the risk area based on the current field of view boundary and the preset blind spot boundary, and deploy the phantom vehicle within the risk area;

[0015] Step 1-2: Based on the risk collision theory, construct the longitudinal velocity model of the Phantom vehicle.

[0016] Preferably, in step 1-1, the Phantom vehicle is deployed within the risk area, specifically as follows:

[0017] A limited number of blind spot phantom vehicles are evenly distributed along the lane centerline within the risk area, and an additional phantom vehicle is placed at the boundary of the field of vision to account for the worst-case scenario.

[0018] If we assume that the Phantom vehicle travels along the lane centerline, then... The collection of phantom vehicles at any given moment is as follows:

[0019]

[0020] In the formula, This indicates the location of the blind zone boundary; Indicates the first The configuration of a Phantom vehicle is expressed as follows: ;in, The location of the Phantom vehicle; The longitudinal speed of the Phantom vehicle; It is a Boolean value representing the existence state of the Phantom vehicle.

[0021] Preferably, the longitudinal velocity model of the phantom vehicle in steps 1-2 is as follows:

[0022]

[0023] In the formula, Let be the longitudinal velocity function of the Phantom vehicle; The average speed of vehicles on the road can be obtained from traffic flow data; For collision risk parameters, the distance between the Phantom vehicle and the intersection. Relevant, defined as:

[0024]

[0025] in, For a predefined safe distance; This refers to driver attention parameters.

[0026] Preferably, step 2 includes the following sub-steps:

[0027] Step 2-1: Construct a cost function to evaluate the vehicle's driving trajectory;

[0028] Step 2-2: Construct a rational driver model based on the A* algorithm;

[0029] Steps 2-3: Compare the cost function differences between the rational driving trajectory and the actual historical trajectory of the visible vehicle under different traffic condition configurations to obtain the likelihood estimate of the traffic condition configuration corresponding to the observed historical trajectory of the visible vehicle.

[0030] Steps 2-4: Based on Bayes' theorem, obtain the probability distribution of visible vehicle intentions and blind spot traffic conditions.

[0031] Preferably, steps 2-3 specifically include:

[0032] Traffic condition configuration includes visible vehicle intentions and potential traffic conditions in blind spots;

[0033] The difference in cost functions between the rational driving trajectories and the actual historical trajectories of visible vehicles under different traffic condition configurations is parameterized as a Boltzmann distribution. The likelihood estimate of the traffic condition configuration corresponding to the observed historical trajectory values ​​of visible vehicles is as follows:

[0034]

[0035] in, Configure the vehicle to reflect the intended purpose of the vehicle. Configuration for traffic conditions in blind spots; For the visible vehicle from the initial moment to Historical trajectory observations at any given time; The temperature coefficient is the Boltzmann distribution. The function is the cost function constructed in step 2-1; Configuration for visible vehicles in traffic conditions From the initial moment to A rational driving trajectory at all times.

[0036] Preferably, steps 2-4 specifically include:

[0037] Probability distribution of traffic configuration in blind spots:

[0038]

[0039] in, Normalization factor; Let be the prior probability distribution of the visible vehicle's intention; Probability distribution configured for traffic conditions in blind spots; Likelihood estimates are configured for traffic conditions corresponding to historical trajectory observations of visible vehicles;

[0040] The probability distribution of the vehicle's intention is visible:

[0041]

[0042] in, This is the normalization factor.

[0043] Preferably, step 3 includes the following sub-steps:

[0044] Step 3-1: Construct a partially observable Markov model for an uncontrolled intersection and determine the model's state space and observation space;

[0045] Step 3-2: Integrate the estimation results of uncertain factors to determine the initial confidence level and transition function of the model;

[0046] Step 3-3: Determine the motion space of the model;

[0047] Steps 3-4: Determine the model's reward function, expressed as:

[0048]

[0049] in, Let be the reward function for the distance between the autonomous vehicle and the target point. Let be the reward function for the difference between the autonomous vehicle's speed and the reference speed. For the comfort reward function of autonomous vehicles, A reward function to avoid collisions;

[0050] Steps 3-5: Solve the Markov model online using the Monte Carlo search tree method, and select the optimal action policy in real time based on the reward function.

[0051] Preferably, step 3-1 specifically includes:

[0052] state space :

[0053]

[0054] In the formula, For the vehicle's state, the expression is: ,in, Here are the vehicle's position coordinates. For the vehicle's speed, This serves as a reference route for the vehicle. For the first The status of a visible vehicle is expressed as: ,in, For the first The location coordinates of the visible vehicles For the first Speed ​​of visible vehicles; Configuration for traffic conditions in blind spots; For the first The intended configuration of the visible vehicles;

[0055] Observation space The set of state variables that can be directly observed by the sensor:

[0056]

[0057] In the formula, These are the vehicle state space that can be directly observed by sensors and the visible vehicle state space that can be directly observed by sensors.

[0058] Preferably, step 3-2 specifically comprises:

[0059] The uncertainty estimation results in step 2 are used as the initial confidence for the configuration of visible vehicle intentions and blind spot traffic conditions, respectively; the blind spot traffic condition model considering the driver's subjective awareness described in step 1 and the rational driver model in step 2-2 are used as the state transition functions.

[0060] Compared with the prior art, the present invention has the following advantages:

[0061] 1) High degree of anthropomorphism: By comprehensively analyzing the interaction behavior between vehicles and making full use of the information contained in the historical observation behavior of visible vehicles, the traffic situation at uncontrolled intersections is understood and estimated. Joint reasoning is performed to address the uncertainty of the intentions of visible vehicles and the uncertainty of potential traffic conditions in blind spots, providing anthropomorphic reasoning results.

[0062] 2) High traffic efficiency: Most existing methods only consider the worst case for uncertainty, which is prone to providing overly conservative strategy choices. The unmanned vehicle trajectory planning method in this invention integrates the results of uncertainty reasoning, which can effectively improve the traffic efficiency of unmanned vehicles at intersections while ensuring safety. Attached Figure Description

[0063] Figure 1 This is a flowchart of the method of the present invention;

[0064] Figure 2 Framework diagram for planning intersections in blind spots for autonomous vehicles, taking into account uncertainties;

[0065] Figure 3 This is a traffic scene representation diagram in an embodiment of the present invention;

[0066] The vehicle on the right is an autonomous vehicle, and the vehicle on the left is a visible vehicle. for Always keep an eye on the vehicle's field of vision. As the boundary of the risk area, The spacing between Phantom vehicles configured in blind spots. The distance between the Phantom vehicle and the intersection. This represents the three different intentions of the visible vehicle (going straight, turning right, turning left). Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0068] Example

[0069] This embodiment provides an interactive planning method for autonomous vehicles that considers intersection uncertainties. This method utilizes an uncertainty estimator to obtain the probability distributions of visible vehicle intentions and blind spot traffic conditions. The inference results are then input into a planner based on a partially observable Markov model to select the optimal action. Figure 1 As shown, it includes the following steps:

[0070] Step 1: Construct a model of potential traffic participants in blind spots that takes into account the driver's subjective awareness;

[0071] Step 1-1: Determine the risk area based on the current field of view boundary and the preset blind spot boundary, and deploy the phantom vehicle within the risk area;

[0072] A limited number of blind spot phantom vehicles are evenly distributed along the lane centerline within the risk area, with an additional phantom vehicle positioned at the visual boundary to account for worst-case scenarios. It is assumed that the phantom vehicles travel along the lane centerline. A collection of phantom vehicles at every moment:

[0073] in, This indicates the location of the blind zone boundary; Indicates the first Configuration of a Phantom vehicle:

[0074]

[0075] in, The location of the Phantom vehicle; The longitudinal speed of the Phantom vehicle; It is a Boolean value representing the existence state of the Phantom vehicle.

[0076] Step 1-2: Based on the risk collision model, construct the longitudinal velocity model of the Phantom vehicle.

[0077] The longitudinal velocity function of the Phantom vehicle:

[0078]

[0079] in, The average speed of vehicles on the road can be obtained from traffic flow data; risk collision parameters. Related to the distance from the Phantom vehicle to the intersection, defined as:

[0080]

[0081] in, For predefined safe distance, This is a parameter related to driver attention. To avoid collision risks, the Phantom vehicle travels at a lower speed as it gets closer to the intersection, which aligns with the driving behavior of a human driver with subjective awareness.

[0082] Step 2: Perform joint reasoning on blind spot uncertainty and visible vehicle intent uncertainty to generate probability distributions for uncertain factors;

[0083] Step 2-1: Construct a cost function to evaluate the vehicle's trajectory;

[0084] Step 2-2: Construct a rational driver model based on the A* algorithm;

[0085] Steps 2-3: Compare the cost function differences between the rational driving trajectory and the actual historical trajectory of visible vehicles under different traffic condition configurations to obtain the likelihood estimate of the traffic condition configuration corresponding to the observed historical trajectory of visible vehicles.

[0086] Traffic situation configuration includes two parts: visible vehicle intentions and potential traffic situations in blind spots. Visible vehicle intentions... The potential traffic situation in the blind spot is as described in step 1-1.

[0087] The difference in cost functions between the rational driving trajectories and the actual historical trajectories of visible vehicles under different traffic condition configurations is parameterized as a Boltzmann distribution. The likelihood estimate of the traffic condition configuration corresponding to the observed historical trajectory values ​​of visible vehicles is as follows:

[0088]

[0089] in, Configure the vehicle to reflect the intended purpose of the vehicle. Configuration for traffic conditions in blind spots; For the visible vehicle from the initial moment to Historical trajectory observations at any given time; The temperature coefficient is the Boltzmann distribution. The function is the cost function constructed in step 2-1; Configuration for visible vehicles in traffic conditions From the initial moment to A rational driving trajectory at all times;

[0090] Steps 2-4: Based on Bayes' theorem, obtain the probability distribution of visible vehicle intentions and blind spot traffic conditions.

[0091] Probability distribution of traffic configuration in blind spots:

[0092]

[0093] in, Normalization factor; Let be the prior probability distribution of the visible vehicle's intention; The probability distribution configured for traffic conditions in blind spots.

[0094] The probability distribution of the vehicle's intention is visible:

[0095]

[0096] in, This is the normalization factor.

[0097] Step 3: Input the uncertainty estimation results into the longitudinal speed planner based on the partially observable Markov model to obtain the optimal action strategy of the unmanned vehicle when passing through the intersection without traffic lights.

[0098] Step 3-1: Construct a partially observable Markov model for an uncontrolled intersection and determine the model's state space and observation space;

[0099] The state space of an autonomous vehicle at an uncontrolled intersection includes the states of the autonomous vehicle and visible vehicles, as well as the intentions of the visible vehicles. and configuration of traffic conditions in blind spots , can be represented as:

[0100]

[0101] Among them, the vehicle status is:

[0102]

[0103] in Here are the vehicle's position coordinates. For the vehicle's speed, This is a reference route for the vehicle.

[0104] No. Visible vehicle status:

[0105]

[0106] in For the first The location coordinates of the visible vehicles For the first The speed of the visible vehicles.

[0107] The vehicle's intention is visible. As described in steps 2-4, the blind spot traffic situation configuration is as described in step 1-1.

[0108] The model assumes that there is no sensor noise in the scene, and the model's observation space is the set of state variables that can be directly observed by the sensors:

[0109]

[0110] Step 3-2: Integrate the estimation results of uncertain factors to determine the initial confidence level and transition function of the model;

[0111] The uncertainty estimation results described in step 2 are used as the initial confidence for the configuration of visible vehicle intentions and blind spot traffic conditions, respectively; the blind spot traffic condition model considering the driver's subjective awareness described in step 1 and the rational driver model described in step 2-2 are used as the state transition functions.

[0112] Step 3-3: Determine the action space of the model;

[0113] The action space of the autonomous vehicle is defined as a one-dimensional space, and only the selection of the longitudinal speed strategy is considered.

[0114] Steps 3-4: Determine the model's reward function, specifically:

[0115]

[0116] in, The reward function is the distance between the autonomous vehicle and the target point; The reward function is the difference between the autonomous vehicle's speed and the reference speed. For the comfort reward function of the autonomous vehicle; A reward function to avoid collisions.

[0117] Steps 3-5: Solve the model online to obtain the optimal action strategy for the autonomous vehicle.

[0118] The Markov model is solved online using the Monte Carlo search tree method, and the optimal action strategy is selected in real time based on the reward function, enabling the autonomous vehicle to pass safely and efficiently at uncontrolled intersections with blind spots.

[0119] Here is a specific application example:

[0120] In actual operation, see Figure 2 First, the probability distribution of uncertain factors is generated using an uncertainty estimator, as follows:

[0121] Step 1: Establish a model of potential traffic conditions in blind spots, see... Figure 3 Based on the autonomous vehicle's perception field of view boundary and a predefined risk boundary, a finite set of blind spot phantom vehicles is determined. Based on the risk collision model, the speed of the Phantom vehicle in each blind spot is determined. .in, The prior average road speed and the risk collision parameters . For predefined safe distance, This refers to driver attention parameters.

[0122] Step two involves comparing the cost functions of rational trajectories and real historical trajectories under different traffic conditions, based on the inverse programming framework. The rational driving model that generates the rational trajectories of visible vehicles assumes that the visible vehicles travel along the centerline of their corresponding lanes according to their current intentions, with their longitudinal speed obtained by searching the st graph using the A* algorithm. The real historical trajectories of the visible vehicles are recorded by the autonomous vehicle's perception module. The cost function metrics include longitudinal velocity, longitudinal acceleration, longitudinal distance, and path curvature.

[0123] Step 3: Calculate the likelihood estimates for all possible traffic condition configurations, and obtain the probability distribution of uncertainties based on Bayes' theorem. Parameterize the cost function difference between the rational driving trajectories and actual historical trajectories of visible vehicles under different traffic condition configurations as a Boltzmann distribution, and calculate the likelihood estimates of the traffic condition configurations corresponding to the observed historical trajectories of visible vehicles. .in, Configure the vehicle to reflect the intended purpose of the vehicle. Configuration for traffic conditions in blind spots; For the visible vehicle from the initial moment to Historical trajectory observations at any given time; The temperature coefficient is the Boltzmann distribution. The function is the cost function constructed in step two. The probability distribution of blind spot traffic configurations is calculated using Bayes' theorem. Probability distribution of visible vehicle intentions .

[0124] See Figure 2 The operation steps of the autonomous vehicle trajectory planner based on a partially observable Markov model are as follows:

[0125] Step 1: Construct a partially observable Markov decision model for an intersection without traffic lights. The state space of the model is determined as follows: The observation space is The pose, speed, reachability intention, and potential traffic conditions of the vehicle and visible vehicles in the state space are determined respectively.

[0126] Step 2: Incorporate the inference results of the uncertainty factor estimator into the planning framework to determine the initial confidence level and transition function of the model. The inference results of the uncertainty factor are expressed as a probability distribution, which can be directly used as the initial confidence input in the observable Markov model for each planning cycle. Different traffic conditions correspond to different state transition functions. The blind spot phantom vehicle transitions its state according to the blind spot traffic condition model in Step 1. It can be seen that the state transition functions under different vehicle intentions are generated according to the rational driving model in Step 2.

[0127] Step 3: Determine the action space of the model. The action space is defined as a one-dimensional space, considering only the selection of the longitudinal acceleration of the autonomous vehicle. The selectable longitudinal accelerations are chosen from a series of positive and negative values ​​with a certain step size to represent different degrees of acceleration and deceleration.

[0128] Step 4: Determine the model's reward function .in, The reward function is the distance between the autonomous vehicle and the target point, which encourages the selection of actions that bring the autonomous vehicle closer to the target point; The reward function is the difference between the autonomous vehicle's speed and the reference speed, which encourages the selection of actions that make the autonomous vehicle's speed closer to the reference speed. The comfort reward function for autonomous vehicles makes the vehicle ride smoother and more comfortable. A reward function is used to avoid collisions and ensure vehicle safety.

[0129] Step 5: Solve the Markov model online using the Monte Carlo search tree method, and select the optimal action strategy in real time based on the reward function to enable the autonomous vehicle to pass safely and efficiently at uncontrolled intersections with blind spots.

[0130] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An unmanned vehicle blind area intersection planning method considering uncertainty, characterized in that, The method comprises the following steps: Step 1, constructing a blind area potential traffic participant model considering the subjective consciousness of the driver, the construction process comprising the following sub-steps: Step 1-1: determining a risk area according to the current visual field boundary and the preset blind area boundary, and configuring a phantom vehicle in the risk area; Step 1-2: constructing a longitudinal speed model of the phantom vehicle based on the risk collision theory, specifically: wherein is the longitudinal speed function of the phantom vehicle; is the average travel speed of road vehicles, obtained from traffic flow data; is the risk collision parameter, related to the distance of the phantom vehicle to the intersection is defined as: wherein, is a predefined safety distance; is a driver attention parameter; Step 2, constructing an uncertainty estimator, jointly reasoning the blind area uncertainty and the visible vehicle intention uncertainty, and generating a probability distribution of the uncertainty factors; Step 3, inputting the uncertainty estimation result into a longitudinal speed planner based on a partially observable Markov model to obtain an optimal action strategy of the unmanned vehicle when passing through a no-lamp-control blind area intersection.

2. The method of claim 1, wherein, In the step 1-1, the phantom vehicle is configured in the risk area, specifically: A limited number of blind area phantom vehicles are uniformly distributed along the lane centerline in the risk area, and an additional phantom vehicle is configured at the visual field boundary to consider the worst case; It is considered that the phantom vehicle takes the lane center line as the path and travels along the path, then The phantom vehicle set at the moment is: In the formula, is the position of the blind area boundary; represents the first vehicle configuration, the expression is ; wherein, is the position of the phantom vehicle; is the longitudinal speed of the phantom vehicle; is a Boolean value representing the existence state of the phantom vehicle.

3. The method of claim 1, wherein, The step 2 comprises the following sub-steps: Step 2-1, constructing a cost function for evaluating the vehicle driving trajectory; Step 2-2, constructing a rational driver model based on the A* algorithm; Step 2-3, comparing the cost function difference between the rational driving trajectory and the real historical trajectory of the visible vehicle under different traffic situation configurations to obtain the likelihood estimation of the traffic situation configuration corresponding to the visible vehicle historical trajectory observation value; Step 2-4, obtaining the probability distribution of the visible vehicle intention and the blind area traffic situation configuration according to the Bayes rule.

4. The method of claim 3, wherein, The step 2-3 is specifically: The traffic situation configuration includes the visible vehicle intention and the blind area potential traffic situation; The cost function difference between the rational driving trajectory and the real historical trajectory of the visible vehicle under different traffic situation configurations is parameterized as a Boltzmann distribution, and the likelihood estimation of the traffic situation configuration corresponding to the visible vehicle historical trajectory observation value is: wherein, configuring an intention of the visible vehicle; configuring a traffic situation of the blind area; configuring a historical trajectory observation of the visible vehicle from an initial time to a time ; configuring a temperature coefficient of the Boltzmann distribution; is a cost function constructed in step 2-1; configuring a rational driving trajectory of the visible vehicle from the initial time to the time under the traffic situation; .

5. The method of claim 4, wherein, The step 2-4 is specifically: The probability distribution of the blind area traffic situation configuration: wherein, is a normalization factor; is a prior probability distribution of visible vehicle intent; is a probability distribution of blind zone traffic configurations; is a likelihood estimate of traffic configurations corresponding to visible vehicle historical trajectory observations; The probability distribution of the visible vehicle intention: wherein is a normalization factor.

6. The method of claim 3, wherein, The step 3 comprises the following sub-steps: Step 3-1, constructing a partially observable Markov model at the no-lamp-control intersection and determining the state space and observation space of the model; Step 3-2, fusing the uncertainty estimation result to determine the initial confidence and transition function of the model; Step 3-3, determining the action space of the model; Step 3-4, determining the reward function of the model, the expression being: wherein, is a reward function for the distance between the unmanned vehicle and the target point, is a reward function for the difference between the speed of the unmanned vehicle and the reference speed, is a comfort reward function for the unmanned vehicle, is a reward function for avoiding collision; Step 3-5, solving the Markov model online through the Monte Carlo search tree method, and selecting the optimal action strategy in real time according to the reward function.

7. The method of claim 6, wherein, The step 3-1 is specifically: State space : In the formula, For the vehicle's state, the expression is: ,in, Here are the vehicle's position coordinates. For the vehicle's speed, This is a reference route for the vehicle. For the first The status of a visible vehicle is expressed as: ,in, For the first The location coordinates of the visible vehicles For the first Speed ​​of visible vehicles; Configuration for traffic conditions in blind spots; For the first The intended configuration of the visible vehicles; Observation space A set of state quantities directly observable by the sensor: In the formula, respectively, are the self-vehicle state space directly observable by the sensor, and the visible vehicle state space directly observable by the sensor.

8. The method of claim 6, wherein, The step 3-2 is specifically: The uncertainty estimation result in step 2 is taken as the initial confidence of the visible vehicle intention and the blind area traffic situation configuration; the blind area traffic situation model considering the subjective consciousness of the driver in step 1 and the rational driver model in step 2-2 are taken as the state transition function.

Images