Vehicle control method, device, storage medium and electronic device

Abstract

A vehicle control method, equipment, storage medium and electronic device. The method comprises: obtaining real-time traffic information of a target vehicle (202); when the real-time traffic information indicates that the target vehicle (202) has arrived at a target intersection (20), obtaining first vehicle state information of the target vehicle (202) in a current driving state, and obtaining second vehicle state information of the target vehicle (202) in an area where the target intersection (20) is located in the current driving state, wherein the target vehicle (202) includes surrounding vehicles (206) passing through the target intersection (20) and virtual vehicles (204-1, 204-2; 602-1 to 602-9) configured for the current observation field of the target vehicle (202); inputting the first vehicle state information and the second vehicle state information into a vehicle motion prediction model to obtain a prediction result, wherein the vehicle motion prediction model is used to predict the vehicle's driving action in the next control cycle based on the vehicle state information in the current control cycle; and controlling the target vehicle (202) according to the prediction result to perform the target driving action.

Description

Technical Field

[0001] The present invention relates to the field of vehicle control, and in particular, to a vehicle control method, device, storage medium and electronic device. Background Art

[0002] Nowadays, in order to simplify driving operations, more and more intelligent driving technologies are being applied to vehicles to help drivers drive more safely and reliably. Obviously, autonomous driving has become an important development trend.

[0003] However, there are various uncertain traffic conditions at intersections without traffic lights, and existing autonomous driving technologies cannot handle such situations. For example, a moving vehicle suddenly appears at an intersection, blocking the current driving field of the autonomous driving target vehicle, which may cause the target vehicle to easily collide with potential obstacles in the blocked area. That is, in the existing technology, when a vehicle passes through an intersection, it is often difficult to effectively ensure driving safety during vehicle control because the road conditions cannot be fully understood.

[0004] There is currently no effective solution to this problem. Summary of the invention

[0005] Embodiments of the present invention provide a vehicle control method and apparatus, a storage medium, and an electronic device to solve the problem of relatively low control safety of vehicle control methods in related arts.

[0006] According to one aspect of an embodiment of the present invention, a vehicle control method is provided, comprising: obtaining real-time traffic information of a target vehicle; when the real-time traffic information indicates that the target vehicle has arrived at a target intersection, obtaining first vehicle status information of the target vehicle in a current driving state, and obtaining second vehicle status information of the target vehicle in an area where the target intersection is located in the current driving state, wherein the target vehicle includes surrounding vehicles passing through the target intersection and a virtual vehicle configured for the current observation field of view of the target vehicle; inputting the first vehicle status information and the second vehicle status information into a vehicle motion prediction model to obtain a prediction result, wherein the vehicle motion prediction model is used to predict the vehicle's driving action in the next control cycle based on the vehicle status information of the current control cycle; and controlling the target vehicle according to the prediction result to perform the target driving action.

[0007] According to another aspect of an embodiment of the present invention, a vehicle control device is further provided, comprising: a first acquisition unit, for acquiring real-time traffic condition information of a target vehicle; a second acquisition unit, for acquiring first vehicle state information of the target vehicle in a current driving state, and acquiring second vehicle state information of the target vehicle in an area where the target intersection is located in the current driving state, when the real-time traffic condition information indicates that the target vehicle has arrived at a target intersection, wherein the target vehicle includes surrounding vehicles passing through the target intersection and a virtual vehicle configured for a current observation field of view of the target vehicle; an input prediction unit, for inputting the first vehicle state information and the second vehicle state information into a vehicle motion prediction model to acquire a prediction result, wherein the vehicle motion prediction model is used to predict the vehicle's driving action in the next control cycle based on the vehicle state information of the current control cycle; and a control unit, for controlling the target vehicle according to the prediction result to perform the target driving action.

[0008] According to another aspect of the present invention, a computer-readable storage medium is further provided, wherein the computer-readable storage medium stores a computer program, wherein the computer program executes the vehicle control method when it is executed.

[0009] In an embodiment of the present invention, when a target vehicle is about to pass through a target intersection, a corresponding virtual vehicle is configured in the obstructed observation field of the target intersection to simulate potential moving obstacles in the obstructed area, and the target driving action of the target vehicle when passing through the target intersection is predicted by a vehicle motion prediction model in combination with the vehicle state information of surrounding vehicles and the target vehicle that actually passes through the area where the target intersection is located. The present invention enables the target vehicle to fully understand the road condition information of the target intersection when automatically passing through the target intersection, thereby avoiding the risk of collision during driving, improving the safety of vehicle control, and thus overcoming the problem of relatively low vehicle control safety in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The accompanying drawings described herein provide further understanding of the present application and form a part of the present application. The illustrative embodiments of the present application and their description are only used to explain the present application and are not intended to limit the present application.

[0011] Figure 1 is a flow chart of an alternative vehicle control method according to an embodiment of the present invention;

[0012] Figure 2 is a schematic diagram of an application environment of an alternative vehicle control method according to an embodiment of the present invention;

[0013] Figure 3 is a schematic diagram of an alternative vehicle control method according to an embodiment of the present invention;

[0014] Figure 4is a schematic diagram of another alternative vehicle control method according to an embodiment of the present invention;

[0015] Figure 5 is a schematic diagram of another alternative vehicle control method according to an embodiment of the present invention;

[0016] Figure 6 is a schematic diagram of another alternative vehicle control method according to an embodiment of the present invention;

[0017] Figure 7 is a schematic diagram of another alternative vehicle control method according to an embodiment of the present invention;

[0018] Figure 8 is a schematic diagram of another alternative vehicle control method according to an embodiment of the present invention;

[0019] Fig. 9 is a schematic diagram of another alternative vehicle control method according to an embodiment of the present invention;

[0020] Fig.10 is a schematic diagram of another alternative vehicle control method according to an embodiment of the present invention;

[0021] Fig.11 is a schematic diagram of another alternative vehicle control method according to an embodiment of the present invention;

[0022] Fig.12 is a schematic diagram of another alternative vehicle control method according to an embodiment of the present invention;

[0023] Fig.13 is a structural diagram of a replaceable vehicle control device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solution in the embodiment of the present invention is clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all. Other embodiments obtained by ordinary technicians in this field based on the embodiment of the present invention without creative efforts belong to the protection scope of the present invention.

[0025] It should be noted that in the specification, claims and drawings of the present invention, the terms "first" and "second" are used to distinguish similar description objects and do not necessarily indicate an order or sequence. It should be understood that the data used in the present application can be interchanged where applicable, so that the embodiments of the present application can be implemented in an order different from that shown or described herein. In addition, the terms "including" and "having" are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device, including a series of steps or units, is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or inherent to such a process, method, product or device.

[0026] According to one aspect of an embodiment of the present invention, a vehicle control method is provided. Optionally, as an optional implementation, Figure 1 As shown, the vehicle control method includes:

[0027] S102, obtaining real-time traffic information of the target vehicle;

[0028] S104, when the real-time traffic information indicates that the target vehicle has arrived at the target intersection, first vehicle state information of the target vehicle in the current driving state is obtained, and second vehicle state information of the target vehicle in the area where the target intersection is located in the current driving state is obtained, wherein the target vehicle includes surrounding vehicles passing through the target intersection and a virtual vehicle configured for the current observation field of view of the target vehicle;

[0029] S106, inputting the first vehicle state information and the second vehicle state information into a vehicle motion prediction model to obtain a prediction result, wherein the vehicle motion prediction model is used to predict the driving action of the vehicle in the next control cycle based on the vehicle state information of the current control cycle;

[0030] S108, controlling the target vehicle according to the prediction result to execute the target driving action.

[0031] Optionally, in this embodiment, the above vehicle control method can be applied to, but not limited to, vehicles equipped with an advanced driver assistance system (ADAS) to help the vehicle perform efficient and safe driving control operations during automatic driving. The ADAS system uses various sensors installed on the vehicle to sense the surrounding environment at any time during the vehicle's driving, collect data, identify, detect and track static and dynamic objects, collect map data from the navigator, and calculate and analyze the system, so that the driver can be aware of possible dangers in advance, thereby effectively improving the driving safety of the vehicle.

[0032] And, in this embodiment, after obtaining the real-time traffic information of the target vehicle, if the real-time traffic information indicates that the target vehicle has arrived at the target intersection, the first vehicle state information of the target vehicle in the current driving state and the second vehicle state information of the target vehicle in the area where the target intersection is located in the current driving state are obtained. The vehicles here include surrounding vehicles passing through the target intersection and virtual vehicles configured for the current observation field of the target vehicle. The first vehicle state information and the second vehicle state information are input into the vehicle motion prediction model, which is used to predict the driving action of the vehicle based on the vehicle state information to obtain a prediction result, and the target vehicle is controlled according to the prediction result to perform the target driving action. That is, when the target vehicle is about to pass through the target intersection, a corresponding virtual vehicle is configured in the obstructed observation field of the target intersection to simulate potential moving obstacles in the obstruction area, and the vehicle state information of the surrounding vehicles and the target vehicle that actually passes through the area where the target intersection is located is combined, and the target driving action of the target vehicle when passing through the target intersection is predicted by the vehicle motion prediction model. The present invention enables the target vehicle to fully understand the road condition information of the target intersection when automatically driving through the target intersection, thereby avoiding the risk of collision during driving, improving the safety of vehicle control, and thus overcoming the problem of relatively low vehicle control safety in the prior art.

[0033] The following combination Figure 2 As shown in the following specific description: Assume Figure 2 What is shown in FIG. 2 is the real-time traffic information of the target intersection 20 currently reached by the target vehicle 202; and, assuming that the current driving purpose of the target vehicle 202 is to turn left, as shown in FIG. Figure 2 In the direction indicated by the dotted arrow, the target vehicle 202 will turn left into the lateral lane. The lateral lane includes two lanes, lane_1 and lane_2, wherein the passing direction of lane_1 is opposite to the driving direction of the target vehicle 202, and the passing direction of lane_2 is the same as the driving direction of the target vehicle 202. Further, in the present application, the rectangle 202 filled with diagonal lines represents the target vehicle 202, the rectangle 204-1 filled with grid lines and the rectangle 204-2 filled with grid lines represent the virtual vehicles 204-1 and the virtual vehicles 204-2 set at the boundaries of both sides of the current observation field of view, the rectangle filled with horizontal lines represents the surrounding vehicles 206, and the blank rectangle 200 represents the static obstacle 200 of the target intersection 20.

[0034] like Figure 2 As shown, based on the real-time traffic information of the target intersection 20 with a static obstacle 200 on the right, it can be seen that the current observation field of view of the target vehicle 202 is Figure 2A triangular area filled with points in the image. Then, a virtual vehicle 204-1 and a virtual vehicle 204-2 are set on the boundary of the current observation field of view, wherein the virtual vehicle 204-1 is set on the left side of lane_1, and the virtual vehicle 204-1 drives to the right to enter the target intersection 20, and the virtual vehicle 204-2 is set on the right side of lane_2, and the virtual vehicle 204-2 drives to the left to enter the target intersection 20. In the present application, the virtual vehicle 204-1 and the virtual vehicle 204-2 are each configured with a maximum driving speed allowed in the corresponding lane. By setting the virtual vehicle to avoid the target vehicle 202 from colliding with a potential moving obstacle after entering the target intersection 20, the potential moving obstacle of the target vehicle 202 in the blind spot of the field of view is simulated.

[0035] Furthermore, the vehicle motion prediction model inputs the vehicle status information of the virtual vehicle, the vehicle status information of the target vehicle itself, and the vehicle status information of the surrounding vehicles currently passing through the target intersection, and predicts the target driving action to be performed in the next control cycle for the target vehicle to ensure the safety of the target vehicle when passing through the target intersection.

[0036] In an embodiment of the present invention, when a target vehicle is about to pass through a target intersection, a corresponding virtual vehicle is configured in the obstructed observation field of the target intersection to simulate potential moving obstacles in the obstructed area, and the target driving action of the target vehicle when passing through the target intersection is predicted by a vehicle motion prediction model in combination with the vehicle state information of surrounding vehicles and the target vehicle that actually passes through the area where the target intersection is located. The present invention enables the target vehicle to fully understand the road condition information of the target intersection when automatically passing through the target intersection, thereby avoiding the risk of collision during driving, improving the safety of vehicle control, and thus overcoming the problem of relatively low vehicle control safety in the prior art.

[0037] As an optional solution, obtaining the second vehicle state information of the target vehicle in the area where the target intersection is located in the current driving state includes:

[0038] S1: Determine the arrival time of the target vehicle at the target intersection;

[0039] S1: Acquire surrounding vehicle status information of surrounding vehicles that enter the area where the target intersection is located at the arrival time, wherein the second vehicle status information includes surrounding vehicle status information;

[0040] S3: Determine the visual field boundary of the current observation visual field of the target vehicle matching the arrival time;

[0041] S4: Acquire virtual vehicle state information of a virtual vehicle configured at a boundary of the field of view, wherein the second vehicle state information includes the virtual vehicle state information.

[0042] Optionally, in this embodiment, before obtaining the virtual vehicle state information of the virtual vehicle configured at the boundary of the field of view, the method further includes:

[0043] 1. A first virtual vehicle is set on a first intersection line where a first lane of a target intersection intersects with a boundary of a field of view, wherein the first intersection line is a starting line for the first virtual vehicle to enter a current field of view, a driving direction of the first virtual vehicle on the first lane is the same as a driving direction of the target vehicle, and the first virtual vehicle is configured with a first maximum driving speed allowed by the first lane;

[0044] 2. A second virtual vehicle is set on a second intersection line where the second lane of the target intersection intersects with the boundary of the field of vision, wherein the second intersection line is a starting line for the second virtual vehicle to enter the current observation field of vision, the driving direction of the second virtual vehicle in the second lane is opposite to the driving direction of the target vehicle, and the second virtual vehicle is configured with a second maximum driving speed allowed by the second lane.

[0045] It should be noted that when an autonomous vehicle passes through an intersection without a signal light, it not only needs to consider static obstacles (such as traffic safety islands) and dynamic obstacles (such as moving vehicles) in the visible area of ​​the field of view; it also needs to consider potential obstacles in the occluded area outside the field of view (such as occluded moving vehicles), thereby allowing the target vehicle to imitate human autonomous behavior. Figure 3 At the intersection shown in (a)-3(d), by improving the field of view in one step, dynamic obstacles can be avoided and the target intersection can be passed.

[0046] For example, when faced with Figure 3 When approaching the intersection shown in (a), the driver will generally adopt a slow-moving forward strategy. The driver moves forward slowly and observes oncoming vehicles in the lane ahead if planning to turn right, and oncoming vehicles in both lanes if planning to turn left. In this way, the driver carefully expands the visual range until there is enough safe distance to turn.

[0047] Figure 3 The dynamic occluder shown in (b) and Figure 3 The more complex cases shown in (c) and 3(d) are handled in a similar way. The rule-based approach appears to be able to mimic human decision making given a specific traffic scenario. However, it has difficulty coping with increasing complexity and uncertainty and lacks the ability to generalize.

[0048] Therefore, in order to simulate potential obstacles, a virtual vehicle can be configured at the boundary of the observation field of view, and a driving direction and a driving speed can be assigned to the vehicle to simulate a real vehicle, so as to predict whether the target vehicle will collide with the virtual vehicle when passing through the target intersection, thereby preventing the target vehicle from colliding with potential moving obstacles in the occluded area. The collision scenario of the target vehicle 202 (i.e., the vehicle ego) is shown in FIG. Figure 4 As shown, vehicle 202 collides with another vehicle 402. Their positions, velocities, and angles are known or observable.

[0049] To achieve a collision-free trajectory, two sources of uncertainty need to be considered. The first is the movement of other vehicles (i.e., surrounding vehicles). For a vehicle with multiple route options, if one of the routes intersects with the route of the vehicle ego, then this uncertainty should be taken into account during planning. Another uncertainty comes from occluded areas. If the route of a potential vehicle (i.e., virtual vehicle) in the occluded area intersects with the route of the vehicle ego, the virtual vehicle needs to be processed. The vehicle state information of the vehicle ego includes the vehicle's position, speed, angle, and the length and width of the vehicle.

[0050] This information can be estimated from the perception results of the prior information of different types of vehicles. Based on these observations, a series of accelerations a=[a0; a1; a2; ..,] are generated along the preset route, which enables the vehicle ego to pass through the congested intersection safely and efficiently.

[0051] The following example is used to describe the real-time traffic information of the target vehicle. Figure 5 As shown in Figure 2, the field of view of vehicle ego202 is blocked. The field of view boundary of each lane is indicated by a vertical line.

[0052] Here, both static and dynamic obstacles are taken into account. We propose a general method to handle various occlusion scenarios regardless of the type of occlusion. Figure 5 The method is explained in the simple scenario shown in Figure 1. The vehicle ego plans to turn left and there are two horizontal lanes intersecting with the route of the vehicle ego. Figure 5 As shown, each lane has a left and right field of view boundary. Vehicles outside the boundary are invisible. In order to plan a safe trajectory, it is assumed that the virtual vehicle is in an obscured area. Here we adopt a worse method. The virtual vehicle is placed on the field of view boundary as a real vehicle for planning. For lane_1, the oncoming vehicle on the left may interfere with vehicle ego 202, so the virtual vehicle 602-1 is placed at the left boundary 1. Similarly, another virtual vehicle 602-2 is placed at the right boundary 2. As shown in FIG. Figure 6 (a) shown.

[0053] Dynamic obstacles may cause similar visual obstructions. Virtual vehicle 602-3 is placed at the boundary in the same manner as static obstacles. Figure 6 (b) As shown. For the case where both static occlusion and dynamic occlusion exist, the field of view boundary can be calculated in the same way, so the virtual vehicles (i.e., virtual vehicles 602-5 to 602-7) can be calculated as follows Figure 6 (d) and (e) are placed in a similar manner. When a large truck 604 is right in front of the vehicle ego 202. In this case, the area behind the truck 604 and within the boundary will be blocked. To handle this situation, a virtual vehicle is placed in the lane behind the truck. Figure 6 As shown in (c) and (f), the positions of the virtual vehicles are symmetrical with the vehicle ego 202 relative to the truck 604, namely, the virtual vehicle 602-4 and the virtual vehicle 602-8 and the virtual vehicle 602-9.

[0054] For lanes that are opposite to the direction of the vehicle ego, the virtual vehicle is always placed on the left field of view boundary. For lanes that are in the same direction as the vehicle ego, the virtual vehicle is placed on the right field of view boundary. In addition, if the area facing the vehicle ego is occluded, the virtual vehicle is also placed there. 7(a)-7(d) channels as shown in Figures.

[0055] Each lane has an occluded part (empty area) and an observable part (diagonally filled area). Occlusions are formed by static obstacles and / or dynamic obstacles. Depending on the direction of the lane and the direction of the vehicle ego, the virtual vehicle is placed at the boundary of the field of view and the occluded area facing the vehicle ego.

[0056] The embodiment provided in the present application simulates potential moving obstacles outside the observation field of view by configuring a virtual vehicle at the boundary of the observation field of view, so that the target vehicle can avoid potential risks during automatic driving, thereby ensuring the driving safety of the target vehicle when passing through the target intersection.

[0057] As an optional solution, inputting the first vehicle state information and the second vehicle state information into a vehicle motion prediction model to obtain a prediction result includes:

[0058] S1, inputting the first vehicle state information and the second vehicle state information into a partially observable Markov decision process model, wherein the vehicle motion prediction model includes the partially observable Markov decision process model;

[0059] S2, predicting and obtaining a target acceleration of a driving action to be performed by the target vehicle in a next control cycle after the current control cycle through a partially observable Markov decision process model, wherein the prediction result includes the target acceleration.

[0060] It should be noted that in this embodiment, in order to fully consider various occlusion scenarios without distinguishing the occlusion types, accurate prediction can be achieved by using but not limited to a model established based on a partially observable Markov POMDP framework.

[0061] Optionally, in this embodiment, during the automatic driving, the target vehicle independently determines the driving action of the next control cycle based on the current real-time road condition information and the vehicle state information of the relevant vehicles in each control cycle. This achieves the goal of controlling the target vehicle to move step by step so as to smoothly pass the target intersection and avoid collision risks.

[0062] Optionally, in this embodiment, controlling the target vehicle to perform the target driving action according to the prediction results includes: determining the target driving speed of the target vehicle according to the target acceleration; and controlling the target vehicle to travel according to the target driving speed to pass the target intersection.

[0063] Through the embodiments provided in the present application, the model established using the POMDP framework can integrate different vehicle state information, and with the help of the solution of the POMDP solver, accurately predict the driving action of the target vehicle in the next control cycle, thereby ensuring the safety of the target vehicle's automatic driving through the intersection.

[0064] As an optional solution, the target acceleration of the driving action to be performed by the target vehicle in the next control cycle after the current control cycle is predicted and obtained through the partially observable Markov decision process model, including:

[0065] S1, in a partially observable Markov decision process model, based on the second vehicle state information, calculating the driving probability of each target vehicle choosing to drive on any lane when passing through the target intersection, and obtaining a driving probability set, wherein the second vehicle state information includes: the position, driving speed, driving direction and driving route of the target vehicle;

[0066] S2, calculating the target acceleration based on the driving probability set and the first vehicle state information, wherein the first vehicle state information includes: the position, driving speed, driving direction and predicted driving route of the target vehicle.

[0067] For POMDP planning, all configured virtual vehicles are treated as real vehicles. This includes alternative routes for virtual vehicles with different route options. Each virtual vehicle is configured with a maximum speed allowed in the current lane for prediction. Unlike real vehicles, the placement of virtual vehicles is reset at each time step based on the current visibility. Since virtual vehicles do not move as predicted, the "optimal" actions generated by general POMDP solvers may not be good decisions and may even cause the system to crash.

[0068] In order to plan a safe trajectory, the time evolution of the traffic situation is required. Generally, at each time step, a reward is assigned to the state of the environment. Then, a strategy is obtained by solving an optimization problem to maximize the expected reward over a period of time. Since the movements of other road vehicles are unknown, the probability distribution of the movements needs to be maintained and updated every time a new observation is obtained. All of this can be done within the POMDP framework.

[0069] POMDP is a mathematical framework for continuous decision making under uncertainty. It is characterized by a tuple (S, A, O, T, Z, R). Here S, A, O represent the state space, action space, and observation space, respectively. Since POMDP is conventionally applied to applications with discrete spaces (such as games), S, A, O are all finite sets of values. T represents the state transition function, which has a probabilistic representation T(s, a, s') = p(s'|a, s), which follows the Markov assumption that its state s' depends only on the most recent state s and is independent of the previous state. Z represents the observation model Z(o, s', a) = p(o|s', a). R represents the reward function This function generates a reward value for executing action a in state s. Future rewards are exponentially calculated by factor γ. Therefore, the cumulative reward value starting from t=0 is expressed as follows:

[0070]

[0071] Since some state information (such as the routes of surrounding vehicles) cannot be measured directly, a probability distribution is configured on the state, called the belief state. The belief state b(s) represents the probability of being in state s, b:S→[0,1]. When an action a is performed and a new observation o is obtained, the belief b(s) can be updated by the Bayesian rule.

[0072]

[0073] POMDP aims to find an optimal strategy π * , the policy maps belief states to actions so as to maximize the expected value of the cumulative reward. As shown in Equation 3.

[0074]

[0075] The details of POMDP modeling are as follows:

[0076] 1) State and Observation: The state of the POMDP includes the information of the vehicle ego and all surrounding vehicles, except for vehicles in the occlusion area. To solve the occlusion problem, a virtual vehicle is configured as described above. The state also includes the information of the virtual vehicle. The state of the vehicle ego is defined as follows:

[0077] s ego =[lego ,v ego ,θ ego ,r ego ,] (4)

[0078] Among them l ego =(x ego ;y ego ) is the current position. θ is the direction. r ego Indicates the predefined route in this operation. ego Indicates the longitudinal speed. Surrounding vehicles veh i The state of is defined in the same way as above. However, for vehicles with different route options, is unknown. For example, Figure 5 The vehicle on the left side of lane 1 may go straight or turn right, which cannot be directly determined. However, the state of the vehicle ego is completely known. In addition to vehicles with multiple route options, the states of other vehicles can be observed. For simplicity, observation noise is not introduced. In actual settings, observation noise can be easily added according to the sensor fusion accuracy.

[0079] 2) Action space: The action of the vehicle ego is generated by the POMDP solver. In this application, the action is the longitudinal acceleration. The action space is A = {+3m / s 2 ,0m / s 2 ,-3m / s 2}, whose values ​​represent acceleration, maintaining the current speed, and deceleration respectively.

[0080] 3) Transition function: The transition function is used to predict the future state of the surrounding environment, given the action performed at the current time step. In this way, the time evolution of the traffic situation can be achieved during planning. Here we assume that the direction and speed of the surrounding vehicles remain unchanged, and the speed of the vehicle ego will be adjusted according to the action at each time step. The transition model is shown in Equation 5. For all surrounding vehicles, a is set to 0.

[0081]

[0082] 4) Belief Update: Since some vehicles have multiple route choices, the probability of each route is adjusted with the new observations. This is done according to Bayes' rule. For example, Figure 8 In, veh i There are two route options, r1 and r2, each with its own reference velocity, denoted by v1 and v2 respectively. Given the current position p and velocity v, the probability of route i;i=1;2 is updated by Eq. 6.

[0083]

[0084] Assuming l and v are conditionally independent, given r i , thus obtaining

[0085] p(l,v|r i )=p(l|r i )p(v|r i )=Ν(y;y i ,σ y,i )Ν(v;v i ,σ v,i ) (7)

[0086] Here we use the Gaussian model. i Representative route i The reference position y is given by x. The parameter σ can be obtained from the data y,i and σ v,i In this operation, σ y,i =10 and _σ v,i = 10, because small variances may drop to zero quickly due to continuous doubling. More complex predictors can be created using position p and velocity v as two attributes and trained based on continuous historical information. However, that is not the focus of this application.

[0087] 5) Reward function: In this operation, the reward function R(s,a)) consists of three parts, as shown below:

[0088] R(s,a)=R end (s)+R collision (s)+R v (s) (8)

[0089] If the vehicle ego passes through the intersection, it will receive a positive reward R end (s). If the vehicle ego collides with other road vehicles, it will receive a negative reward R collision (s) R collision (s) is usually set to a higher value to ensure driving safety. In addition, if the vehicle ego needs to travel at high speed and its speed is lower than the speed limit of the current lane, a negative reward is assigned.

[0090] In this operation, set R end (s) = 1000, R collision (s) = -10000. v The magnitude of should be much smaller than other rewards and adjusted based on the current speed.

[0091] 6) Regarding the issue of virtual vehicles: Usually, the optimal action can be obtained by maximizing Equation 3. However, it does not work well in the technical problem of this application. This is because a virtual vehicle is introduced. In the POMDP planning process, the virtual vehicle is treated in the same way as a real vehicle. However, the virtual vehicle does not really move forward. In fact, the position of the virtual vehicle is reset at each time step, which may cause the state to freeze given the latest visibility of the vehicle ego. For example, according to Equation 3, the vehicle ego may stand still and wait for the virtual vehicle to go first. However, it is assumed that the virtual vehicle is always at the same field of view boundary, so that the vehicle ego is always stationary. In order to avoid a frozen state, the vehicle ego needs a thrust to maximize the expected reward even if it is stationary. In order to unfreeze the vehicle ego while encouraging high driving speeds, different priorities can be assigned to different actions.

[0092] Through the embodiments provided by the present application, the driving probability of the target vehicle is obtained when selecting each lane, and the driving action of the target vehicle is determined in combination with the driving trend to obtain the driving trend of the target vehicle, so that the driving trajectories of the target vehicles will not intersect when passing through the target intersection, that is, collision is avoided. This improves the safety of vehicle control.

[0093] As an optional solution, calculating the target acceleration using the driving probability set and the first vehicle state information includes:

[0094] S1, according to the driving probability set and the first vehicle state information, search for the target path in a pre-constructed search tree, and use the acceleration corresponding to the target node in the target path as the target acceleration, wherein each node on the search tree is configured with a corresponding node reward value, the node reward value is associated with the predicted collision result based on the current node path, and each node is connected to the following sub-nodes: an acceleration node for controlling vehicle acceleration, a deceleration node for controlling vehicle deceleration, and a holding node for controlling the vehicle to maintain the current speed, the selection priority of the acceleration node is higher than the selection priority of the holding node, and the selection priority of the holding node is higher than the selection priority of the deceleration node.

[0095] As an optional solution, searching for a target path in a pre-constructed search tree according to the travel probability set and the first vehicle state information includes:

[0096] S1, traverse the search tree until the target path is found;

[0097] S11, obtain the current acceleration corresponding to the current node;

[0098] S12, when the target vehicle is traveling at a speed corresponding to the current acceleration, using the current acceleration, the driving probability set and the first vehicle state information to predict a collision prediction result between the target vehicle and the target vehicle;

[0099] S13, when the collision prediction result shows that there is no collision, the current node is used as a node of the target path, and a node is obtained from a child node connected to the current node as the current node;

[0100] S14, when the collision prediction results of all nodes in the found search path indicate that there is no collision, the search path is determined as the target path.

[0101] In this operation, the POMDP is solved online to cope with the dynamic environment. This is necessary because the field of view changes as the vehicle ego moves forward. The online solver plans a series of actions until a certain field of view. However, the vehicle ego only executes the first action and recalculates it at the next time point.

[0102] POMDP is usually solved by traversing the belief search tree, selecting the optimal action along the path with the maximum expected reward. However, the computational complexity of building a "complete tree" is too high to be practical. Therefore, various POMDP approximate solvers have been developed. Most of them are designed for discrete state spaces and do not take into account the characteristics of certain applications. In this application, we developed a customized POMDP solver that can solve the problem caused by virtual vehicles while reducing the computational complexity. The solver details are as follows.

[0103] 1) Belief Update. Typically, belief search trees take into account possible observations, so belief updates are performed during the construction of the tree. This is not necessary for applications with continuous state spaces, since the observation space is not a finite set of values. These virtual observations are generated by a probabilistic observation model and may differ significantly from real observations. Furthermore, planning is performed in an online manner and can adapt to changes in the environment. Therefore, the belief of each node remains unchanged during this operation. Fig. 9 As shown. Each node represents a belief state. The three edges branching out from the node represent acceleration ("+"), maintaining the current speed ("0"), and deceleration ("-"). The dotted arrow line represents the successful path from the root to the leaf. All nodes filled with slashes have been visited, and blank nodes will not be visited.

[0104] 3) Action Priority. The vehicle ego is expected to travel at high speed with less deceleration. This is usually achieved by assigning a negative reward when the speed of the vehicle ego is below a certain threshold (such as the speed limit of the current lane). It has been successfully applied to intersection scenarios. Here, different actions are assigned different priorities. "Accelerate" has the highest priority, followed by "Maintain current speed", and "Slow down" has the lowest priority. In this way, high-speed driving is encouraged, which can push the vehicle ego away from a stagnant state. In order to ensure that actions with higher priorities are always visited first, a search strategy is developed.

[0105] 4) Search strategy To handle action priorities, a depth-first search is performed, such as Fig. 9 Here, the solution of the present application is to try to find a path that starts from the root and ends at a leaf without any collision. When such a path is found, the tree traversal stops. In this way, acceleration is always visited first, and deceleration is visited only when acceleration and maintaining constant speed do not work. For example, if Fig. 9 The dashed line in is a complete path without collision, so the optimal action at the current time point is to accelerate. The nodes filled with slashes have been visited, while the blank nodes will never be visited because a successful path has been found. This strategy also increases the average computational cost. Assuming that each leaf has the same probability p of becoming the end point of a successful path, the expectation of how many leaf nodes need to be visited forms a geometric series, as shown in Equation 9.

[0106]

[0107] with the number of leaves |a| H Compared with, |a| represents the size of the action space, H represents the planning horizon, It is usually much smaller, with the minimum being |a| H .

[0108] In addition to the above search strategies, other tricks also help to reduce the computational complexity. Since the trajectories of other vehicles are fixed during the prediction period, the states of individual vehicles for each future time step can be pre-computed, eliminating repeated calculations at each node. Another reason for the high computational cost is the belief state that contains different combinations of states. Assume that each vehicle has 2 route options and there are 5 vehicles in the scene, then the total number of combinations is 25. Each combination is checked for collisions and rewards are distributed at each node. If speed changes are also considered, the total number of combinations will be too large to be calculated in real time. This can be achieved by decomposing the utility function, which plans for each vehicle and merges the results to select the optimal action.

[0109] As an optional solution, searching for a target path in a pre-constructed search tree according to the driving probability set and the first vehicle state information includes:

[0110] S1, traverse the search tree, and take each node in the search tree as the current node in turn to perform the following operations:

[0111] S11, obtaining a first cumulative value and a reference cumulative value set, wherein the reference cumulative value set is obtained by weighted summing the node reward value of the starting node and the node reward value of the current node of the search tree, wherein the reference cumulative value set includes a plurality of second cumulative values, and the second cumulative value is a cumulative value obtained by weighted summing the node reward value of the starting node and the node reward value of any other node at the same level as the current node;

[0112] S12, when the first cumulative value is greater than each second cumulative value in the reference cumulative value set, taking the current node as a node in the target path, and obtaining a node from a child node connected to the current node as the current node;

[0113] S13, when the current node is at the leaf node layer, the path from the starting node to the current node is used as the target path.

[0114] Optionally, in this embodiment, when the target path is found in a pre-constructed search tree, it also includes: configuring corresponding node reward values ​​for each node in the search tree, wherein the node reward value is a first reward value based on the target vehicle of the current node colliding with the search path when crossing the target intersection, a second reward value based on the target vehicle of the current node not colliding with the search path when crossing the target intersection, and a weighted sum of the reference reward values.

[0115] refer to Figure 10-12 The examples shown in fully describe the solutions in the embodiments of the present application.

[0116] The experiments in this operation were conducted in the SUMO simulator. SUMO is an open source package for traffic simulation. It can simulate road networks and other traffic elements. It also provides APIs and TraCI for online interaction. The actions of each vehicle can be controlled at each simulation step. However, the acceleration output by the POMDP cannot be directly controlled, only the maximum acceleration can be set. Therefore, we adjust the speed at each simulation time step given the current action according to Equation 5, and the time interval between planning and simulation is 0.1s. Although SUMO provides an intelligent driver model (IDM) to control the behavior of each vehicle, it is not used in this operation. There are two reasons. First, vehicles equipped with IDM will slow down to avoid collisions, and sometimes freeze traffic if a collision is predicted. Second, the intelligent behavior of surrounding vehicles makes it difficult to evaluate the method provided in this application. Therefore, all other vehicles drive in a pre-set mode without collision avoidance.

[0117] First, we analyze the behavior of the vehicle ego when only static obstacles exist. Fig.10 As shown in (a), the vehicle ego plans to turn left, so it is assumed that there are two virtual vehicles (such as Figure 6In this case, there are no real surrounding vehicles. Fig.10 (b) shows the velocity profile of the vehicle ego and its position y. As we can see, the vehicle ego first accelerates because the hypothetical virtual vehicle is far away from it. Then, the vehicle ego stops accelerating because continued acceleration may result in a collision, but at the same time, it does not need to slow down. As the vehicle ego continues to approach the horizontal lane ahead, it starts to slow down. When it determines that it is safe to turn left, it starts to accelerate until it reaches the maximum speed allowed.

[0118] In order to have a comprehensive view of the traffic, the vehicle ego needs to reach a position y>=-4m. Fig.10 As shown in (b), before obtaining full coverage, the vehicle ego starts to accelerate at t = 2:3s. Because at t = 2:3s, there is enough distance between the vehicle ego and the virtual vehicle, so turning left will not cause any collision. This is also the reaction of human drivers. In many cases, we cannot fully observe the traffic, but as long as we ensure enough distance, we can cross the intersection.

[0119] exist Fig.11 In Figure 1, we compared the proposed method with two other strategies. One is V2V communication, which allows the vehicle ego to access information about the occluded area. The other does not solve the occlusion problem, so no virtual vehicle is assumed. It can be seen that the proposed method has a similar curve to V2V communication. If no measures are taken to deal with the occlusion, the vehicle ego will accelerate until it observes the other vehicle. In this case, the two vehicles collide at t = 3:7s.

[0120] Here we assume a scenario where a moving truck blocks the view of the vehicle ego. To exclude the influence of other factors, static occlusions are removed. Fig.12 As shown in Figure 1, vehicle ego is turning left at the intersection. A truck is approaching and blocking the view of vehicle ego. At the same time, another vehicle is behind the truck and cannot be observed by vehicle ego. Since the truck blocks the view, according to Figure 6 (b) and 6(c) place the virtual vehicles sequentially.

[0121] The method provided in this application is compared with two other strategies, V2V communication and a control group (baseline) without virtual vehicles. Fig.12The speed curves of the above three methods are shown. It can be seen that both V2V and the method provided by this application can complete the left turn without any collision, while the control group collided at t=5:7s. We also noticed that the speed curves of V2V and the method provided by this application are very close, but the method provided by this application is more conservative. This application discusses the strategy for autonomous driving vehicles to pass through congested intersections. Static occlusions and dynamic occlusions are covered. A method for handling various occlusion scenarios is provided. The POMDP framework is used to deal with the proposed problem. Due to the introduction of virtual vehicles, their behavior is different from that of real vehicles, and the actions given by the general POMDP solver are no longer optimal, and sometimes the entire system is frozen. In order to cooperate with the proposed method, this application has developed a special POMDP solver. At the same time, the average computational complexity can also be reduced.

[0122] The performance of the proposed method is evaluated in static and dynamic occlusion scenarios. Experimental results show that the vehicle ego exhibits driving behaviors similar to those of human drivers. We also compare it with other strategies.

[0123] The method proposed in this application is effective in avoiding collisions. In addition, the method can be easily extended to other intersections with different lane layouts.

[0124] It should be noted that for simplicity, the above method embodiments are described as a series of action combinations. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, and some steps in the present invention can be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessary in the present invention.

[0125] According to another aspect of an embodiment of the present invention, a vehicle control device for implementing the above vehicle control method is also provided. Fig.13 As shown, the device includes:

[0126] 1) A first acquisition unit 1302 is used to acquire real-time road condition information of a target vehicle;

[0127] 2) A second acquisition unit 1304, when the real-time traffic information indicates that the target vehicle has arrived at the target intersection, acquires first vehicle state information of the target vehicle in the current driving state, and acquires second vehicle state information of the target vehicle in the area where the target intersection is located in the current driving state, wherein the target vehicle includes surrounding vehicles passing through the target intersection and a virtual vehicle configured for the current observation field of the target vehicle;

[0128] 3) an input prediction unit 1306, configured to input the first vehicle state information and the second vehicle state information into a vehicle motion prediction model to obtain a prediction result, wherein the vehicle motion prediction model is configured to predict the driving action of the vehicle in the next control cycle based on the vehicle state information in the current control cycle;

[0129] 4) A control unit 1308, used to control the target vehicle to perform the target driving action according to the prediction result.

[0130] Regarding the specific embodiment, please refer to the description of the above-mentioned vehicle control method, which will not be repeated in this embodiment.

[0131] According to another aspect of an embodiment of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, and the computer program executes the steps in any one of the method embodiments when running.

[0132] Optionally, in this embodiment, the computer-readable storage medium may be configured to store a computer program, and the computer program is used to perform the following steps:

[0133] S1, obtaining real-time traffic information of the target vehicle;

[0134] S2, when the real-time traffic information indicates that the target vehicle has arrived at the target intersection, obtaining first vehicle state information of the target vehicle in the current driving state, and obtaining second vehicle state information of the target vehicle in the area where the target intersection is located in the current driving state, wherein the target vehicle includes surrounding vehicles passing through the target intersection and a virtual vehicle configured for the current observation field of view of the target vehicle;

[0135] S3, inputting the first vehicle state information and the second vehicle state information into a vehicle motion prediction model to obtain a prediction result, wherein the vehicle motion prediction model is used to predict the driving action of the vehicle in the next control cycle based on the vehicle state information of the current control cycle;

[0136] S4, controlling the target vehicle according to the prediction result to execute the target driving action.

[0137] Optionally, in this embodiment, a person of ordinary skill in the art can understand that all or part of the steps of the method of the above embodiment can be implemented by a program that instructs the relevant hardware of the terminal device. The program can be stored in a computer-readable storage medium, which can include a flash disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0138] The serial numbers of the embodiments of the present invention are only used for description and do not represent preferred embodiments.

[0139] If the integrated unit in the above embodiment is implemented in the form of a software functional unit and sold or used as an independent product, the integrated unit can be stored in the above computer-readable storage medium. Based on this, the technical solution of the present invention substantially constitutes part of the prior art, or all or part of the technical solution can be implemented in the form of a software product, which is stored in a storage medium, and the software product includes several instructions for instructing one or more computer devices (which can be personal computers, servers, network devices, etc.) to execute all or part of the steps of the method described in the embodiment of the present invention.

[0140] In the above embodiments of the present invention, the description of the embodiments mainly highlights the key points, and for parts that are not described in detail in a specific embodiment, reference may be made to the relevant descriptions of other embodiments.

[0141] It should be understood that in several embodiments provided in the present application, the disclosed client can be implemented in other ways. The above embodiments of the device are merely exemplary. For example, the division of units is only a logical function division, and there may be other division methods in actual operation. For another example, multiple units or components can be combined or integrated into another system, or certain features can be ignored or not performed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be achieved through some interfaces. The indirect coupling or communication connection between units or modules can be achieved in electronic or other forms.

[0142] The units that appear in separate form may or may not be physically separated, and the unit parts shown may or may not be physical units, which may be located in a certain position, or may be distributed on multiple network elements. Some or all of the units may be selected according to actual needs to achieve the effect of the solution of the embodiment.

[0143] In addition, the functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The integrated unit may be implemented in the form of hardware or in the form of a software functional unit.

[0144] The above are only preferred embodiments of the present invention. It should be noted that those skilled in the art can make further improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also fall within the protection scope of the present invention.

Claims

1. A vehicle control method, characterized in that: include: Obtain real-time traffic information of the target vehicle; When the real-time traffic information indicates that the target vehicle has arrived at the target intersection, first vehicle state information of the target vehicle in the current driving state is obtained, and second vehicle state information of the target vehicle in the area where the target intersection is located in the current driving state is obtained, wherein the target vehicle includes surrounding vehicles passing through the target intersection and a virtual vehicle configured for the current observation field of view of the target vehicle; Inputting the first vehicle state information and the second vehicle state information into a vehicle motion prediction model to obtain a prediction result, wherein the vehicle motion prediction model is used to predict the driving action of the vehicle in the next control cycle based on the vehicle state information in the current control cycle; Controlling the target vehicle according to the prediction result to perform a target driving action; Wherein, inputting the first vehicle state information and the second vehicle state information into a vehicle motion prediction model to obtain a prediction result includes: inputting the first vehicle state information and the second vehicle state information into a partially observable Markov decision process model, wherein the vehicle motion prediction model includes the partially observable Markov decision process model; Predicting and acquiring, by means of the partially observable Markov decision process model, a target acceleration of a driving action to be performed by the target vehicle in a next control cycle after the current control cycle, wherein the prediction result includes the target acceleration; The method of predicting and obtaining a target acceleration of a driving action to be performed by the target vehicle in a next control cycle after the current control cycle through the partially observable Markov decision process model comprises: In the partially observable Markov decision process model, based on the second vehicle state information, the driving probability of each target vehicle choosing to drive on any lane when passing through the target intersection is calculated to obtain a driving probability set, wherein the second vehicle state information includes: the position, driving speed, driving direction and driving route of the target vehicle; The target acceleration is calculated based on the driving probability set and the first vehicle state information, wherein the first vehicle state information includes: the position, driving speed, driving direction and predicted driving route of the target vehicle.

2. The method according to claim 1, characterized in that The step of obtaining the second vehicle state information of the target vehicle in the area where the target intersection is located in the current driving state includes: Determining the arrival time of the target vehicle at the target intersection; Acquire surrounding vehicle status information of the surrounding vehicles that enter the area where the target intersection is located at the arrival time, wherein the second vehicle status information includes the surrounding vehicle status information; Determining a field of view boundary of the current observation field of the target vehicle that matches the arrival time; Virtual vehicle state information of the virtual vehicle configured at the boundary of the field of view is acquired, wherein the second vehicle state information includes the virtual vehicle state information.

3. The method according to claim 2, wherein: Before acquiring virtual vehicle state information of the virtual vehicle configured at the boundary of the field of view, the method further includes: A first virtual vehicle is arranged on a first intersection line where a first lane of the target intersection intersects with the boundary of the field of view, wherein the first intersection line is a starting line for the first virtual vehicle to enter the current observation field of view, a driving direction of the first virtual vehicle on the first lane is the same as a driving direction of the target vehicle, and the first virtual vehicle is configured with a first maximum driving speed allowed by the first lane; A second virtual vehicle is arranged on a second intersection line where a second lane of the target intersection intersects with the field of vision boundary, wherein the second intersection line is a starting line for the second virtual vehicle to enter the current observation field of vision, a driving direction of the second virtual vehicle in the second lane is opposite to a driving direction of the target vehicle, and the second virtual vehicle is configured with a second maximum driving speed allowed by the second lane.

4. The method according to claim 1, wherein: The controlling the target vehicle to perform the target driving action according to the prediction result comprises: Determining a target driving speed of the target vehicle according to the target acceleration; The target vehicle is controlled to travel according to the target travel speed to pass through the target intersection.

5. The method according to claim 1, wherein: The calculating the target acceleration based on the driving probability set and the first vehicle state information comprises: According to the driving probability set and the first vehicle state information, a target path is searched in a pre-constructed search tree, and the acceleration corresponding to the target node in the target path is used as the target acceleration, wherein each node on the search tree is configured with a corresponding node reward value, and the node reward value is associated with the predicted collision result based on the current node path, and each node is connected to the following sub-nodes: an acceleration node for controlling vehicle acceleration, a deceleration node for controlling vehicle deceleration, and a holding node for controlling the vehicle to maintain the current speed, the selection priority of the acceleration node is higher than the selection priority of the holding node, and the selection priority of the holding node is higher than the selection priority of the deceleration node.

6. The method according to claim 5, wherein: The step of searching for a target path in a pre-constructed search tree according to the travel probability set and the first vehicle state information comprises: Traversing the search tree until the target path is found; Get the current acceleration corresponding to the current node; Predicting a collision prediction result using the current acceleration, the driving probability set and the first vehicle state information, the collision prediction result indicating whether a target vehicle collides with a target vehicle when the target vehicle travels at a driving speed corresponding to the current acceleration; When the collision prediction result indicates that there is no collision, the current node is used as a node in the target path, and a node is obtained from a child node connected to the current node as the current node; When the collision prediction results of all nodes in the found search path indicate that there is no collision, the search path is determined as the target path.

7. The method according to claim 5, wherein: The step of searching for a target path in a pre-constructed search tree according to the travel probability set and the first vehicle state information comprises: Traverse the search tree and take each node in the search tree as the current node in turn to perform the following operations: Obtaining a first cumulative value and a reference cumulative value set, the reference cumulative value set being obtained by weighted summing a node reward value of a start node and a node reward value of a current node of a search tree, wherein the reference cumulative value set includes a plurality of second cumulative values, the second cumulative values ​​being cumulative values ​​obtained by weighted summing a node reward value of the start node and a node reward value of any other node at the same level as the current node; When the first cumulative value is greater than each second cumulative value in the reference cumulative value set, taking the current node as a node in the target path, and obtaining a node from a child node connected to the current node as the current node; When the current node is at the leaf node layer, the path from the start node to the current node is used as the target path.

8. The method according to claim 5, wherein: When searching for a target path in a pre-built search tree, the method further includes: A corresponding node reward value is configured for each node in the search tree, wherein the node reward value is a weighted sum of a first reward value based on the target vehicle at the current node colliding with the search path when crossing the target intersection, a second reward value based on the target vehicle at the current node not colliding with the search path when crossing the target intersection, and a reference reward value.

9. A vehicle control device, characterized in that: include: A first acquisition unit, used to acquire real-time road condition information of a target vehicle; A second acquisition unit, when the real-time traffic information indicates that the target vehicle has arrived at a target intersection, acquires first vehicle state information of the target vehicle in a current driving state, and acquires second vehicle state information of the target vehicle in an area where the target intersection is located in the current driving state, wherein the target vehicle includes surrounding vehicles passing through the target intersection and a virtual vehicle configured for a current observation field of view of the target vehicle; An input prediction unit, configured to input the first vehicle state information and the second vehicle state information into a vehicle motion prediction model to obtain a prediction result, wherein the vehicle motion prediction model is configured to predict a driving action of the vehicle in a next control cycle based on the vehicle state information in a current control cycle; A control unit, configured to control the target vehicle according to the prediction result to perform a target driving action; Wherein, inputting the first vehicle state information and the second vehicle state information into a vehicle motion prediction model to obtain a prediction result includes: inputting the first vehicle state information and the second vehicle state information into a partially observable Markov decision process model, wherein the vehicle motion prediction model includes the partially observable Markov decision process model; Predicting and acquiring, by means of the partially observable Markov decision process model, a target acceleration of a driving action to be performed by the target vehicle in a next control cycle after the current control cycle, wherein the prediction result includes the target acceleration; The method of predicting and obtaining a target acceleration of a driving action to be performed by the target vehicle in a next control cycle after the current control cycle through the partially observable Markov decision process model comprises: In the partially observable Markov decision process model, based on the second vehicle state information, the driving probability of each target vehicle choosing to drive on any lane when passing through the target intersection is calculated to obtain a driving probability set, wherein the second vehicle state information includes: the position, driving speed, driving direction and driving route of the target vehicle; The target acceleration is calculated based on the driving probability set and the first vehicle state information, wherein the first vehicle state information includes: the position, driving speed, driving direction and predicted driving route of the target vehicle.

10. A computer-readable storage medium storing a program, characterized in that: The method according to claim 1 is executed when the program is executed.

Images