An automatic driving decision planning system and method for a tubeless intersection environment

Abstract

The application discloses an automatic driving decision scheme for a control signal-free intersection, which divides automatic driving decision into two main bodies, namely, a prediction-based auxiliary decision and real-time planning, and each main body is independently operated; the auxiliary decision part is responsible for predicting a target vehicle and making an optimal action decision, and outputs a decision result to the real-time planning part; the real-time planning part performs real-time trajectory planning and conflict detection based on observation information and auxiliary decision information, and adjusts the action when necessary. The framework structure is clear, the transparency of the model is ensured, and the safety and controllability in the road driving process are ensured, the interaction and uncertainty between vehicles are fully considered, the prediction result-oriented decision tree search method is adopted for optimal action decision, the operation consumption is effectively reduced, and the real-time decision efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic control, and particularly relates to an automatic driving vehicle decision planning system in a traffic environment of a control signal-free intersection and a strategy tree search decision method based on long-term prediction result guidance. BACKGROUND

[0002] Intersection scenarios without traffic signal guidance have attracted extensive attention in the field of automatic driving, and more intelligent decisions are needed in such an environment. According to the investigation report of the American Highway Traffic Safety Agency, more than 1 / 4 of traffic accidents in the United States are related to intersections, and more than 50% of the accidents occur at uncontrolled intersections.

[0003] In the decision-making process, two problems are very prominent: the first is the improper understanding of the behaviors of other vehicles. For example, without any prediction, when facing vehicles with potential conflicts, the unmanned vehicle may waste a lot of unnecessary time waiting outside the intersection (the behavior is too conservative) or take the risk to enter the intersection first (aggressive). Therefore, it is crucial to predict the situation in the future for a period of time when entering the intersection. At present, vehicle behavior prediction based on graph neural networks is widely used, and the model has the function of describing the interaction characteristics between vehicles and the road environment. Based on this, some scholars have proposed the concept of "anchor trajectory" to list possible future trajectories, and the trajectory prediction problem is converted into an anchor selection + offset regression problem. Compared with this, the target-drive trajectory model, such as TNT, is proposed, which has the following characteristics: 1) making full use of expert knowledge such as road geometry to generate reliable candidate trajectories; 2) outputting multi-modal probability trajectories.

[0004] The second problem is caused by the ignorance of uncertainty. Due to observation errors, complex driving features, and prediction biases, uncertainty is inevitable. Partially Observable Markov Decision Process (POMDP) provides an explicit way to model the uncertainty in the environment, i.e., belief. As the situation advances (the search depth increases), the dimension of the model increases exponentially in the observation space and the action space, making it difficult to run online in real time. In addition, because the actions in the model are discretized, it will also cause the trajectory to be not smooth. At the same time, because the prediction length in the model is insufficient (usually 1-3 steps, less than 2s), it is not conducive to long-term planning. Therefore, Multipolicy Decision-Making (MPDM) improves POMDP by considering constraints from behavior and environment and replacing basic actions with semantic policies, thereby reducing the size of the action space. On this basis, some works also reduce the complexity of the action space and the observation space by formulating a decision tree branching guide policy.

[0005] Although some works have designed decision planning models for uncontrolled intersections, such as a "leader car-following car" game model to model the interaction between vehicles, they only consider a simple scenario of three vehicles and are difficult to generalize to a traffic environment with multiple vehicles and complex interactions. Another work uses a reinforcement learning neural network to mimic the way human drivers adjust the throttle and steering wheel in different situations to control the vehicle, but the training of the model requires a large amount of data, the model itself is a "black box" and difficult to explain, and the safety of the vehicle after executing the model output cannot be guaranteed.

[0006] Therefore, there is an urgent need to develop new decision-making methods for uncontrolled intersections. SUMMARY

[0007] Therefore, the present application provides an automatic driving decision planning system applicable to uncontrolled intersections and a decision tree search decision-making method based on long-term lookahead prediction results, which can make reasonable predictions about the behavior of target vehicles in advance and explicitly quantify the uncertainty of the prediction results into the decision-making method, achieving intelligent and reliable decision planning in complex and frequently interacting uncontrolled intersection environments.

[0008] The automatic driving decision planning system applicable to uncontrolled intersections provided by the present disclosure comprises a co-pilot module and a main driver module, wherein:

[0009] The co-driver module is used for assisting decision-making based on situation prediction, and includes: trajectory prediction on a social vehicle that may have a conflict, and judgment on whether there is a conflict with the current trajectory planning of the controlled vehicle; when there is a conflict, the optimal action sequence is searched and selected with safety as the primary principle, and output to the main driver module;

[0010] The main driver module is used for real-time trajectory planning, and includes: trajectory planning and conflict detection on the planned trajectory according to the current observation and the auxiliary decision-making result of the co-driver module, and action adjustment when it is determined that there is a conflict, that is, updating the trajectory.

[0011] Further, the co-driver module is run once every 1s, and the optimal action sequence within 3s is obtained each time; and the main driver module is run once every 0.1s.

[0012] Further, the system further includes a social vehicle selection module for judging the intention and / or driving style of the social vehicle; wherein the intention judgment is used for further screening the vehicles for trajectory prediction to determine the sampling path; and the driving style judgment is used for controlling the sampling density of the target point.

[0013] The present disclosure also provides an action sequence search method oriented to prediction results, which can be applied to the above-mentioned system, and includes the following steps:

[0014] Based on the current possible action of the controlled vehicle, each step is simulated forward according to the vehicle kinematics;

[0015] After each step of simulation is completed, the safety of each candidate action is evaluated based on the collision possibility of the controlled vehicle and the target vehicle;

[0016] Based on the evaluated action, the next step of forward simulation search is continued until a series of action sequences that meet the search depth requirement are obtained.

[0017] Further, the search method specifically includes the following steps:

[0018] For the current state of the controlled vehicle Generate a candidate action pair set A raw , wherein, A is a longitudinal acceleration action set limited by the maximum jerk, and I is a selectable lane path set;

[0019] Each candidate action a jd =(a jd ,id jd )∈A raw is taken as the starting state, forward simulation is performed according to the vehicle kinematics model to generate the trajectory of the controlled vehicle within a future T period ajd is longitudinal acceleration, id jd is lane path number; j is action number, d is search depth;

[0020] After each forward simulation is completed, based on the trajectory of the controlled vehicle and the target vehicle in the process, the candidate action a jd is checked for safety evaluation;

[0021] Based on the candidate action with high safety, the next step search is continued until the search depth requirement is met.

[0022] Further, the safety evaluation method comprises:

[0023] Each candidate action a jd is accompanied by a score list R, and the maximum value of the list is the final risk score r(a jd ), as shown in equation (4a):

[0024]

[0025] In the score list, the risk score of the kth predicted trajectory of the controlled vehicle and each target vehicle n is stored, which is derived from the mean value of the time risk value sequence r n,k , r n,k is the time list of the risk value of the kth predicted trajectory of the controlled vehicle and the target vehicle n at all t∈T time points;

[0026]

[0027] wherein, is the risk value of the kth predicted trajectory of the controlled vehicle and the target vehicle n at t time point, and the calculation method is shown in equation (4c):

[0028]

[0029] wherein, represents the safety circumscribed rectangle of the controlled vehicle at t time point, which is the rectangle obtained by expanding the shape size of the vehicle itself by a safety margin, represents the safety circumscribed rectangle of the target vehicle n; represents the circumscribed rectangle of the vehicle body of the controlled vehicle, which is determined by the shape size of the vehicle itself, represents the circumscribed rectangle of the predicted vehicle body of the target vehicle; represents the probability of the occurrence of this predicted trajectory, represents the predicted collision time;

[0030] The calculation method of the predicted collision time is shown in equation (4d),

[0031]

[0032] the predicted state of the target vehicle n at time t the coordinate value converted to the coordinate value in the local Frenet coordinate system of the controlled vehicle and the position in the Frenet coordinate system of the controlled vehicle difference, divided by the speed v of the controlled vehicle in the s direction s wherein s represents the longitudinal axis of the local Frenet coordinate system.

[0033] Further, the time length of the trajectory prediction of the target vehicle is 3 seconds, and the search depth length of the action sequence is 3.

[0034] In addition, the present disclosure also provides an automatic driving decision-making method for an uncontrolled intersection applying the above system, comprising the following steps:

[0035] Step S1, performing virtual lane division on the intersection without lane division;

[0036] Step S2, selecting a target vehicle;

[0037] Step S3, performing trajectory prediction on the target vehicle;

[0038] Step S4, judging whether the predicted trajectory of the target vehicle conflicts with the current trajectory planning of the controlled vehicle;

[0039] Step S5, when there is a conflict, performing action sequence search with safety as the primary principle;

[0040] Step S6, selecting the optimal action sequence from the obtained series of action sequences.

[0041] Further, based on the TNT model, trajectory prediction is performed on the target vehicle.

[0042] Further, the decision-making method further comprises step S7 and / or step S8, wherein:

[0043] Step S7, using a random forest model to judge the intention of the target vehicle; based on the intention, further filtering the target vehicle for trajectory prediction to determine the sampling path;

[0044] Step S8, based on an IMM interactive multi-model, judging the driving style of the target vehicle; based on the driving style, determining the sampling density of the target sample.

[0045] Further, the step S8 specifically comprises:

[0046] Suppose that any target vehicle n has three possible driving styles, which are aggressive, normal and conservative, denoted as {ξ a ,ξb ,ξ c Each driving style ξ corresponds to a certain acceleration range;

[0047] When the target vehicle is first observed, the probabilities of belonging to the three driving styles are initialized to the same value, and thereafter, the estimated state of the vehicle based on the observation of the vehicle is compared with the actual state reached by the vehicle , and the likelihood Λ is calculated ξ ;

[0048] Then the prediction part of the interactive multi-model is used to update the probability of belonging to the style, where P is the covariance, ∈ is a small constant to avoid zero division, and ξ' is the normalized ξ;

[0049]

[0050] The target sample will be sampled at a resolution r = r0 / Λ ξ , where r0 is a fixed sampling resolution.

[0051] Further, the above decision-making method further comprises the following steps:

[0052] According to the optimal state sequence corresponding to the optimal action sequence, and the real-time observation result of the social vehicle, real-time trajectory planning and conflict detection are performed.

[0053] When it is determined that there is a conflict, the action is adjusted, that is, the trajectory is updated.

[0054] Further, when it is determined that there is a conflict, a game model is used to adjust the action.

[0055] The automatic driving decision planning scheme provided by the present disclosure divides the automatic driving decision into two main parts: auxiliary decision based on situation prediction and real-time planning, and each part is relatively independent. The former is used for target prediction and optimal action decision, and outputs the result to the real-time planning part. The real-time planning part performs real-time trajectory planning and collision detection based on observation information and auxiliary decision information, and adjusts the action when necessary. Among them, the search of the action sequence adopts a "prediction-guided strategy tree search method (Prediction-Guided Strategy Tree, PGST)", to ensure the high efficiency and feasibility of the search.

[0056] Compared with the prior art, the present disclosure has the following advantages:

[0057] (1) The task of autonomous driving decision-making is clearly divided. The two main entities operate relatively independently, which greatly reduces the complexity of the autonomous driving decision-making process and the branch interference caused by handling various real-time situations, improves the transparency of the model, and ensures the safety of driving under various road conditions.

[0058] (2) The decision-making method takes into account the interactive behavior and uncertainty of vehicles during road driving, and integrates the learning-based target vehicle prediction model and the tree search-based action decision model, which can understand and predict the development trend of the dynamic environment.

[0059] (3) Based on the prediction judgment of intent and driving style, the input data of the target vehicle trajectory prediction model was optimized, which reduced the amount of data computation and improved the accuracy of model prediction.

[0060] (4) Clearly consider the uncertainty of the prediction results, generate a “chaotic prediction forward simulation scenario” based on the predicted possible trajectory distribution of the target vehicle, consider all possible future dangerous scenarios, and evaluate the safety level of the action sample.

[0061] (5) The search for action sequences adopts a decision tree search guided by the prediction results. After each step of simulation prediction, the prediction results are evaluated based on safety criteria. Only those that pass the evaluation are used for the next step of simulation search. This effectively reduces the observation space and action space, reduces the computational burden, and improves the decision-making speed.

[0062] (6) The auxiliary decision frequency is 1Hz, which can predict the trajectory of the target vehicle within 3 seconds each time and output the optimal state sequence within 3 seconds, thus realizing long-cycle prediction; combined with 10Hz real-time planning, it can have sufficient foresight and adapt to real-time changes in road conditions. Attached Figure Description

[0063] Figure 1 This is a schematic diagram illustrating the process of using the "prediction-oriented decision tree search" method of this invention at an uncontrolled intersection.

[0064] Figure 2 An exemplary framework diagram and flowchart for passenger-side decision support and driver-side trajectory planning at uncontrolled intersections;

[0065] Figure 3 A schematic diagram of the conversion between the global coordinate system and the local Frenet coordinate system is shown. The global coordinate system is the set geodetic coordinate system, and the Frenet coordinate system takes the vehicle's position center as the longitudinal zero point, the tangent direction of the road's forward movement as the s direction, and the direction perpendicular to the tangent direction as the d direction.

[0066] Figure 4 A schematic diagram of collision detection is shown, where a) is the circumscribed rectangle G representing the shape of the vehicle.VEH And the circumscribed rectangle G with a safety margin SAFE b) The result of unfolding the vehicle trajectory described by the circumscribed rectangle in the spatiotemporal domain; c) A collision-free scenario; d) A collision scenario;

[0067] Figure 5 The following are several keyframe scenarios illustrating the use of the method disclosed herein to determine the direction of oncoming traffic at an intersection without traffic signals.

[0068] Figure 6 The following keyframe scenarios are shown, illustrating how the method of this disclosure is used to determine right-turning vehicles at an intersection without traffic signals.

[0069] Figure 7 The following are several keyframe scenarios illustrating the use of the method disclosed herein to determine left-turning vehicles at an intersection without traffic signals.

[0070] Figure 8 The diagram shows the position curve, speed curve, and acceleration curve for straight, right, and left turns. The triangle symbol represents the time of each decision, and the shaded area represents the driving process in which the main vehicle interacts with other vehicles.

[0071] Figure 9 Several keyframe scenarios are shown, illustrating the decision-making process using the method of this disclosure in complex left-turn situations at intersections without traffic signals. Detailed Implementation

[0072] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0073] This disclosure primarily addresses the decision-making and planning problem of autonomous driving at intersections without control signals where there are interactions and uncertainties. The basic solution is twofold: firstly, autonomous driving requires real-time motion planning; secondly, predicting and considering the future state of the dynamic environment to make forward-looking decisions is also crucial. These two driving functions should be closely integrated, similar to a "Primary Driver" (PD) and a "Subordinate Driver" (SD) in real-world traffic scenarios.

[0074] Following this approach, an exemplary decision-making and planning system for uncontrolled intersections is shown in the attached figure. Figure 2 As shown, it includes:

[0075] The system consists of a co-pilot module responsible for assisting decision-making and a pilot module responsible for real-time planning, wherein:

[0076] The co-pilot module is configured to perform virtual lane division on an intersection without lane division, perform trajectory prediction on a social vehicle that has a possibility of collision, and determine whether the social vehicle will collide with a current trajectory plan of the controlled vehicle. When a collision exists, the co-pilot module searches and selects an optimal action sequence based on a safety principle, and outputs the optimal action sequence to the pilot module. The optimal action sequence includes a selected lane and a longitudinal acceleration.

[0077] The pilot module continuously performs trajectory planning and collision detection based on real-time observation results and decision results provided by the co-pilot module, and adjusts actions (i.e., updates a trajectory) when a collision is determined to exist. The trajectory refers to a sequence of vehicle states according to time, including a position, a speed, an acceleration, and the like of the vehicle at each time.

[0078] Preferably, the SD model is run once every 1 second, and the running frequency is 1 Hz. The SD model outputs an optimal action sequence within 3 seconds each time. The PD model has a higher running frequency, which is 10 Hz. If a decision needs to be made at present, the SD model is started, and the trajectory prediction model is first called to predict a trajectory of a target vehicle.

[0079] Preferably, the diagram further includes a social vehicle module. The social vehicle module is configured to determine an intention and a driving style of a social vehicle, filter a vehicle that needs to be trajectory predicted based on the intention, and determine a sampling path based on the driving style, and control a sampling density of a target point.

[0080] Figure 2 An exemplary automatic driving decision-making process for an uncontrolled intersection is also shown in the diagram.

[0081] When a host vehicle (i.e., a current controlled vehicle) does not enter an intersection and follows a vehicle, the host vehicle can be controlled by an intelligent driver model (IDM). When it is determined that a decision needs to be made, the decision-making process is as follows:

[0082] The co-pilot first predicts a trajectory of a target vehicle (preferably by using a neural network TNT model). If there is no collision with a current trajectory plan of the host vehicle, the target vehicle can continue to travel according to the current state, or the current state and a vehicle model can be considered to perform more optimized planning. Otherwise, the co-pilot enters a prediction-oriented decision tree search process, searches and selects an optimal action sequence and a corresponding optimal state sequence, and delivers the optimal action sequence and the corresponding optimal state sequence to the pilot.

[0083] The pilot performs trajectory smoothing planning based on the received optimal state sequence. The trajectory smoothing is solved by constructing an optimization problem. Meanwhile, the pilot continuously performs collision detection between a current planning trajectory and a social vehicle based on real-time observation of the social vehicle. If a collision exists, the pilot adjusts actions (preferably by using a game model), i.e., re-plans the trajectory.

[0084] For social vehicles, further judge its intention and driving style, which are used to determine possible vehicles and possible target positions as part of the input of the prediction model, and the driving style is updated by comparing its real action with the likelihood of the estimated action through the game process, using the prediction part of the Interacting Multiple Model (IMM).

[0085] The following gives further examples and explanations of the action sequence search method guided by prediction results and the more complete autonomous driving decision-making method at uncontrolled intersections.

[0086] According to the exemplary autonomous driving decision-making method at uncontrolled intersections of the present disclosure, the following steps are included:

[0087] Step 1: Variable setting

[0088] Define the static environment of the uncontrolled intersection as E, and the lane centerline as l E , the label of the host vehicle is set to 0, and the labels of other target vehicles are set to n = 1, 2, …, N (variable upper index). From the start time t0 of the decision-making period T F , it is assumed that the state of the host vehicle and the state of other target vehicles are obtained, where the state variable X = (x, y, v, a, s, d, φ, id) includes global coordinates (x, y), local Frenet coordinates (s, d), and velocity v, acceleration a, heading angle φ, and the current path number id. The state sequence in the historical time period T H is denoted as (no vehicle number represents the host vehicle and all target vehicles). For the host vehicle, all reference paths are denoted as where the reference path currently traveled by the host vehicle is denoted as

[0089] Step 2: Divide reference paths

[0090] The original planning path will guide the host vehicle to travel through the intersection, which can be generated by the upstream path planner. For the uncontrolled intersection environment, there is usually no lane division, so it is necessary to divide the reference path, i.e., the virtual lane, and each virtual lane corresponds to a different lateral distance deviation.

[0091] Given the environment map or road boundary, first find the feasible lane for the host vehicle to drive off, and then according to the different lateral offset d iThe end point of the exit is determined. After the end point is determined, the reference path is represented by a spline curve, and the mutual conversion between the global coordinates and the Frenet local coordinates is realized.

[0092] Step 3: Selection of target vehicles

[0093] For vehicles that need to pass through the intersection, it is relatively easy to determine the target vehicles that need to be paid special attention to once the intention is determined. Since the shape and size of the intersection are different, a fixed distance threshold is not used in the exemplary embodiment to screen the targets, but vehicles that have a potential interaction with the host vehicle are selected. Specifically, for vehicles within the observation distance range of the host vehicle, first, according to their driving stages, including about to enter, has entered, and has exited, are classified. Vehicles that have exited the intersection are no longer considered; vehicles inside the intersection are all considered as target vehicles, because these vehicles usually have right-of-way or there may be potential conflicts. According to the priority, the vehicles closest to the stop line outside the intersection are considered as target vehicles. Finally, a total of vehicles are selected as target vehicles at t0.

[0094] Step 4: Trajectory prediction of social vehicles

[0095] The prediction of vehicles at the intersection is different from that in the high-speed environment, and it needs to better reflect the uncertainty and diversified maneuverability caused by interaction. First, it is explained that for the trajectory prediction of vehicles at the intersection, an excellent prediction model needs to consider the following factors:

[0096] Road constraints: the behavior of vehicles is constrained by road geometry;

[0097] Interaction: the interaction between vehicles is more frequent and obvious, including cooperation and conflict;

[0098] Uncertainty of behavior: due to the driver's own driving style and external influences, the behavior is changing, so there will be multiple modal probability trajectories in the future.

[0099] According to the above characteristics, the TNT model is preferred in this embodiment, which includes three simple and observable steps, namely local target prediction, motion estimation based on local target points, and trajectory scoring. First, the model will uniformly sample possible target points along the lane center line (the reference path used by social vehicles, which can be considered as a map, generated according to the shape of the intersection); then the network encodes the interaction between vehicles and the interaction between vehicles and the environment, to generate prediction trajectories based on target points. Finally, the scoring step will estimate the probability of the prediction trajectory, and select the top k TNT trajectories as the final result.

[0100] To further reduce the amount of calculation and improve the accuracy of the model, the present disclosure can further include the following steps:

[0101] First, the intention of the vehicle is preferably predicted by a model of random forest (RF), which not only provides a clue for conflict estimation, but also facilitates the determination of the possible reference trajectory of the target vehicle. According to the prediction result of the vehicle intention, the target vehicle is further filtered from the target vehicle in step 3 to enter the TNT model prediction to determine the sampling path of the target point.

[0102] Second, the sampling target point is determined in a more targeted manner. The original TNT model samples the target point for all target vehicles according to a fixed resolution r0, ignoring the current state X0 of the vehicle and the different driving styles ξ of different vehicles. The present disclosure preferably adopts an IMM interactive multi-model to predict the driving style and determine the sampling density of the target point based on the driving style to improve the accuracy of trajectory prediction.

[0103] Specifically, the following method can be used:

[0104] Suppose that the possible driving style of any target vehicle n has three types, which are aggressive, normal and conservative, denoted as {ξ a , ξ b , ξ c}. Each driving style ξ corresponds to a corresponding acceleration range

[0105] When the target vehicle is first observed, the probability of belonging to the three driving styles is initialized to the same value. After that, the estimated state of the target vehicle ( Figure 2 completed by the game model in the illustrated embodiment) is compared with the actual state of the vehicle reached, the likelihood Λ ξ is calculated, and the driving style probability is updated using the IMM interactive multi-model prediction. The calculation method can refer to formula (1), where P is the covariance, and ∈ is a small constant to avoid zero division:

[0106]

[0107] Each time the target sample is composed of samples collected by each driving style ξ simulation model:

[0108]

[0109] In formula (2), the distance range After that, the target samples will be sampled according to the sampling function g with resolution r = r0 / Λ ξ Sampling.

[0110] In summary, the samples are more densely distributed in the driving area that matches the driving style, and the remaining area is more dispersed. This adaptive sampling method associates the learning-based prediction model with the general vehicle model, generates more targeted samples, and thus improves the accuracy of trajectory prediction.

[0111] Step 5: Prediction-oriented decision tree search method

[0112] If the trajectory of the other vehicle based on TNT prediction does not have a potential conflict with the current planned trajectory of the host vehicle, it can be output according to the current state, or it can be considered in combination with the current state and the vehicle model to make a more optimized plan.

[0113] Otherwise, in the known k TNT Under the prediction trajectory and its score, a prediction-based decision search tree is initialized, and the best action sequence with a length of D is finally found The action sequence includes: selecting which lane, longitudinal acceleration, etc., where the longitudinal direction is in the Frenet local coordinate system of the vehicle body.

[0114] In this disclosure, the search for the action sequence provides a prediction result-oriented decision tree search method, that is, based on the current set of all possible actions, the trajectory within a certain time is obtained through forward simulation, the safety is evaluated based on the trajectory, and the forward simulation search is continued based on the evaluated state, thereby obtaining a series of action sequences with a search depth of D

[0115] The method is further described as follows.

[0116] Given the environmental observation and the self-positioning attitude information, the TNT model will use the historical state The lane center line (generated by the map module and used for social vehicle trajectory prediction) corresponds to the vector l E And the local target point set sampled As input, the target vehicle The prediction result in the future T F Time Includes multi-modal trajectory state And its score And the circumscribed rectangle corresponding to the trajectory (Here, it includes the circumscribed rectangle of the vehicle body And the circumscribed rectangle with a safety margin ), as shown in Figure 4 a).

[0117] In POMDP (Partially Observable Markov Decision Process), forward simulation is performed in a series of sampled scenarios, which are permutations and combinations of observations sampled based on the confidence of the environment, which is prone to cause dimension explosion of space. In the algorithm of the present disclosure, the prediction results are directly fused into a mixed observation scenario, and all prediction results and their scores are projected into a three-dimensional space-time. In addition to the lightweight calculation, it has two advantages: first, as long as the prediction is accurate enough, the mixed scenario covers all the most likely dangerous situations; second, all original action samples will be explicitly evaluated for safety risks under the same prediction results.

[0118] Unlike many works that fix the sampled action, the exemplary algorithm of the present disclosure allows the sampled action to be adaptive to the current state of the host vehicle Generate a set of adaptive action pairs A raw In order to reduce the dimension of the action space, it is preferred to use action pairs composed of acceleration actions and semantic lane-changing actions:

[0119] Given the current state of the host vehicle Consider the maximum jerk to get a longitudinal acceleration action set A, which is composed of the current acceleration, the current acceleration plus the maximum jerk, the current acceleration minus the maximum jerk, etc. The value of each element is different from each other, and is limited by the maximum and minimum speed, acceleration range. In the transverse direction of the Frenet coordinate system, the adjacent lanes in the candidate path form a path set I. The original action set is the combination of acceleration and candidate path, that is, For safety, only one path change is allowed during the search depth of one decision.

[0120] Then, forward simulation is performed with each a jd =(a jd , id jd )∈A raw as the starting state, that is, according to the vehicle kinematics model, the trajectory of the host vehicle in the future T period is generated where a jd is the longitudinal acceleration, id jd is the lane path number; j is the action number, d is the search depth; the trajectory obeys the vehicle model as shown in equation (4), where Δt is the discrete sampling time, l r and l f are the lengths backward and forward from the center of mass, which can be approximated as half the length of the vehicle, Turn the corners to the front wheels.

[0121]

[0122] Subsequently, regarding motion pruning, this disclosure provides an exemplary hazard rating and reselection mechanism, namely, based on the obtained trajectory, checking the collision situation between the trajectory and the target social vehicle, and helping to select motion sequences while ensuring safety, such as... Figure 4 As shown in d).

[0123] The decision tree contains two types of containers, one of which is an action container A that holds the parent node at each level. buff and its associated state container M buff Collision indicator container C buff and rating container S buff Another type is the temporary storage container for each layer, including A. temp M temp C temp ,S temp This is used to collect candidate actions that are temporarily at risk.

[0124] After each forward simulation step is completed, a safety assessment is performed on each candidate action, i.e., the safety of each candidate action 'a' is calculated. jd Risk score r(a) jd ), of which complete safety (r(a) jd Actions with a value of 0, along with their corresponding states and scores, will be directly placed into the buffer container at index buff, becoming the parent node; the remaining candidate actions will be temporarily placed into a temporary storage container (index temp). If A buff The number of candidate actions is less than N th Then it will start from sorted A temp Actions with relatively low risk scores are selected again to participate in the next round of expansion. It should be noted that actions that have already collided (i.e., collision = True) are very unlikely to be selected. Even if they are selected, their child nodes will be marked as collisions with a risk score of 1.0, which greatly reduces their chances of being selected.

[0125] The preferred method for risk scoring for each candidate action is as follows:

[0126] Specifically, each action is accompanied by a score list R, where the maximum value of the list is the final danger score, as shown in (4a):

[0127]

[0128] In the score list, the dangerous score of the kth predicted trajectory of the host vehicle and each target vehicle n is stored, which is the average of the dangerous value (4b) at all times t∈T

[0129]

[0130] r n,k is the sequence of the dangerous value at each time t∈T

[0131] At each time t, the dangerous value is calculated by detecting the geometric relationship between the circumscribed rectangle of the host vehicle and the circumscribed rectangle of the target vehicle Specifically, as shown in equation (4c):

[0132]

[0133] wherein, represents the safe circumscribed rectangle of the host vehicle at time t, which is a rectangle obtained by expanding the vehicle's own shape size + safety margin, represents the safe circumscribed rectangle of the target vehicle n, the center position of which can be predicted by a neural network, represents the circumscribed rectangle of the host vehicle, the shape of which is determined by the length and width of the vehicle itself, represents the circumscribed rectangle of the target vehicle; represents the probability of the occurrence of this trajectory predicted by the neural network, represents the predicted collision time.

[0134] At time t, if the safe circumscribed rectangle of the host vehicle and the safe circumscribed rectangle of the target vehicle n have no intersection, it is considered that no collision will occur, and the dangerous coefficient is recorded as 0.0; otherwise, the dangerous value will be calculated according to equation (4c), which includes the prediction probability of this trajectory collision time and warning time. In particular, if the circumscribed rectangle of the host vehicle has an intersection with the target vehicle , it is considered that the possibility of collision is very large, and the dangerous coefficient at this time is directly recorded as 1.0. As Figure 4 a), the trajectory is projected into the time-space domain as Figure 4 b), at each time t, the geometric relationship between the rectangles is detected, c) is no collision, and d) is collision.

[0135] wherein the calculation of the collision time is shown in equation (4d), i.e., the predicted state of the target vehicle n at time t ​The coordinate value converted into the local Frenet coordinate system of the host vehicle (upper subscript 0) is subtracted from the longitudinal position in the Frenet coordinate system of the host vehicle, and then divided by the speed of the host vehicle in the s direction (longitudinal direction).

[0136]

[0137] Attached Figure 1 The process of using the "predictive guidance type decision tree search" method in the intersection without control signals is shown in the following figure. Among them:

[0138] a) The bold curve path in the figure represents the reference path generated in advance according to the road shape, which can be generated by the path planner; (The reference path (solid curve) is generated for the host vehicle; The lane center line (dashed curve) can be considered as generated by the map module, which is used for predicting the trajectory of the social vehicle)

[0139] b) The figure shows the target vehicle determined by predicting the intention of the social vehicle at the intersection, which is marked as a circle;

[0140] c) The figure shows the result of predicting the target vehicle trajectory based on the neural network, and the coincidence of the candidate trajectory of the host vehicle and the predicted trajectory in space-time means the occurrence of potential conflict;

[0141] d) The decision tree search method, the action is composed of lateral path selection (different gray colors represent) and longitudinal acceleration selection (graphical representation), in the screening process of each layer, the safety of the state corresponding to the action is evaluated, and the score is sorted. The action with a danger index of 0.0 will be directly selected as the parent node to enter the next round of expansion, and the rest will be re-screened according to the sorting, and the action with a relatively small danger index will be selected as the parent node. The action with a danger index of 1.0 is considered as an inexecutable action, and the danger index of its child node is always 1.0 to ensure that it is not selected with a high probability.

[0142] Step 6: Select the optimal action sequence

[0143] After step 5, a series of actions are obtained, and the target function is designed to evaluate the action sequence from multiple angles. For each action sequence with a search depth D It is accompanied by a state sequence in the state container And the danger score in the score container. The designed target function will consider the safety comfortable state lane change deviation from the ideal speed and the distance from the target position These costs are weighted linearly combined as shown in equation (5), where λ is a coefficient multiplied by the actual cost of the term. Ultimately, the action sequence with the lowest cost is chosen as the optimal decision result.

[0144]

[0145] Safety is the most important item in the evaluation index, which represents the possibility of collision between the host vehicle and social vehicles, as shown in equation (6), where F safe is the basic cost value, D is the search depth, multiplied by the danger score of the corresponding state (action), and c safe is the normalization factor.

[0146]

[0147] Regarding comfort, acceleration and jerk will be considered, as shown in equation (7).

[0148]

[0149] The lane change cost is represented by the deviation cost of the reference lane. Because it is relatively dangerous to change direction recklessly within the intersection, the lane change cost needs to be considered once the lane change occurs.

[0150]

[0151] The difference between the actual speed and the ideal speed results in a cost as shown in equation (9), and the cost of the target distance is shown in equation (10).

[0152]

[0153]

[0154]

[0155] where γ is a discount factor that takes into account the increasing uncertainty over time.

[0156] Considering that the optimal action sequence + vehicle model = optimal state sequence, since the vehicle model is fixed, given the initial state, the best action sequence and the optimal state sequence are one-to-one correspondence. After determining the optimal action sequence, the corresponding optimal state sequence is taken as the auxiliary decision result, which is provided to the main driving module by the co-pilot module.

[0157] The main driving module can perform trajectory smoothing planning (the trajectory refers to a sequence of vehicle states according to time, including the position, speed, acceleration and the like of the vehicle at each time point) and collision-based conflict detection according to the received optimal state sequence and the real-time observation result of the target social vehicle, and when it is determined that there is a conflict, the trajectory planning is updated, and a game model is preferably used.

[0158] It is considered in the present disclosure that the auxiliary decision and trajectory planning can be run in parallel. Generally speaking, the decision belongs to the macroscopic behavior, needs to be carefully judged and is preferably consistent; and the task of the planning is to adjust the trajectory in time according to the dynamic change of the environment to ensure safety and comfort. Therefore, in the present embodiment, the co-pilot module adopts a running frequency of 1 Hz, and predicts the trajectory of the target vehicle for 3 seconds each time, and correspondingly, the search depth of the optimal action sequence is 3 seconds, that is, the optimal state decision for 3 seconds in the future is given each time; the main driver adopts a frequency of 10 Hz, and each time the action is adjusted, the trajectory for 1 second in the future is regenerated.

[0159] Application example:

[0160] In order to maximize the authenticity of the traffic scene, the data INTERACTION collected in the real scene is selected for experimental test, and the data set collects driving data of various traffic interaction scenes in different countries, which is challenging. At the same time, in order to further represent the reaction of the social vehicle to the behavior of the main vehicle, the following simulation assumptions are made: when there is no dangerous conflict between the social vehicle and the main vehicle, the social vehicle drives according to the original data set trajectory (reflecting the real driving characteristics of human drivers), and when there is a dangerous collision between the two, an action pool is generated based on the real driving style of the vehicle, and an action is selected from the action pool by using a random sampling method to control the vehicle, so as to reflect the characteristics of the interaction and increase the random diversity of the scene.

[0161] Figures 5-7 The representative cases of each intention are shown, including the 132nd vehicle straight through the intersection, the 88th vehicle turning right and the 144th vehicle turning left. In the figure, in order to compare the decision planning result of the method of the present disclosure with the actual driving result of the human driver, the main vehicle driven by the human driver is represented as a rectangle without any mark, the unmanned vehicle controlled by the method of the present disclosure is represented as a rectangle with a label 'ego', and the trajectory generated in each planning period is a thick curve. The social vehicle is represented by a rectangle with a vehicle label, and the multiple dashed lines extended therefrom represent the predicted future trajectory.

[0162] Figure 5The 132nd vehicle is shown to be driving straight through the intersection. Several key frames are captured during the driving process, which are 2.9s, 3.9s, 4.9s, 5.6s, 6.6s and 7.1s frames. When the 132nd vehicle is about to enter the intersection, the first vehicle to interact with it is the 128th vehicle (circled). Because of the prediction as an advanced judgment, the host vehicle successfully predicts the possible danger and generates a deceleration courtesy behavior, generating a low-speed forward trajectory. At 3.9s, the 128th vehicle has passed the front of the host vehicle, and the host vehicle slightly accelerates to improve the traffic efficiency under the premise of safety. After passing through the center of the intersection, there is no vehicle in the scene that poses a danger to the host vehicle, and the host vehicle is in a free driving state, maintaining a smooth acceleration to pass through the intersection. In contrast, the driving strategy of human drivers lacks reasonable and accurate prediction of other vehicles waiting to pass through the intersection, and human drivers can only adopt more conservative behavior, choosing to observe outside the intersection, wasting waiting time and reducing the traffic efficiency of vehicles at the intersection.

[0163] Figure 6 The 88th vehicle is shown to be driving straight through the intersection. Several key frames are captured during the driving process, which are 2.5s, 3.0s, 3.5s, 4.0s, 4.5s and 5.0s. Under the premise of only considering motor vehicles, from the perspective of road right, the vehicle turning right needs to yield to the straight-through vehicle and the vehicle turning left. As can be seen from the 73rd vehicle (circled), when the host vehicle begins to turn right, the 73rd vehicle is about to drive straight through. Since the 73rd vehicle is closer to the exit of the intersection than the host vehicle and has the right of way, the decision of the host vehicle is to decelerate and yield to the 73rd vehicle. As can be seen from the acceleration curve, before 3.5s, the host vehicle is basically in a continuous deceleration phase. After determining that the 73rd vehicle has passed, the host vehicle adjusts the speed according to the real-time traffic situation and smoothly and safely completes the right turn, which has a high similarity with the behavior of human drivers and well restores the decision-making strategy of human drivers. Figure 6

[0164] Figure 7 The 144th vehicle is shown to be driving straight through the intersection. Several key frames are captured during the driving process, which are 2.0s, 3.0s, 3.5s, 4.0s, 5.0s and 6.0s. Left turn is relatively complex. First, the vehicle turning left is often not fixed in a lane, and may turn a large or small bend. Another is that the vehicle passes through a relatively long distance, increasing the probability of interacting with other social vehicles. In the stage when the host vehicle begins to turn left, it finds that the 131st vehicle (circled) is driving straight through from its left, so the host vehicle chooses to decelerate and yield. The continuous deceleration before 4s successfully avoids the 131st vehicle. After avoiding the conflicting vehicle, the host vehicle accelerates at a small amplitude to smoothly pass through the intersection. Figure 8 Figures 5-7 ​​The speed and acceleration curve of the decision planning can be seen, and the host vehicle will adjust its behavior according to the real-time dynamic changes of the scene to ensure safety and relative comfort.

[0165] In order to quantitatively compare the difference between the method (PGST) and the driving effect of human drivers (Ground Truth, GT), the acceleration variance of the decision planning process, the passing time and the like are compared. The speed variance can reflect the overall forward stability of the vehicle. When frequent "go-stop" behavior occurs in the intersection, the speed variance is larger. From the experimental data, it can be seen that the driving speed planned by the framework is relatively stable and the comfort is higher. The passing time index is mainly used to measure the passing efficiency of the vehicle. Short passing time means that the vehicle has a higher understanding of the environment and can integrate into the intersection environment and implement its own strategy in a shorter time. Human drivers often hesitate outside the intersection due to lack of long-term prediction of the environment, and waste a lot of time. The re-planning frequency refers to the behavior of fine-tuning the trajectory when the distance between the host vehicle and the social vehicle is too close. From the table, it can be seen that the re-planning frequency of the method is within the acceptable range, and the consideration of the safety index can be changed by adjusting the safety distance parameter and the cost function and the like.

[0166] Table 1 Comparison of decision planning results of the method and human driver driving

[0167]

[0168]

[0169] Figures 5-7 , 9 using dashed lines to show are the prediction results of the target vehicle. It is worth mentioning that the prediction performance of the neural network in the whole process is also relatively stable. For example, in the straight-through scene, the 129th social vehicle is correctly predicted to turn left, the 128th vehicle is straight through, and the 127th vehicle is right turn. At the same time, in the right turn scene, in the face of the gradually slowing down of the 80th and 84th vehicles, the future forward speed is predicted to be slow and almost stationary. In contrast, for the 71st and 73rd vehicles that will drive out of the intersection, their speed is predicted to accelerate out of the intersection. It can be seen that because the input of the neural network utilizes the road information in the environment, the predicted trajectory is almost well constrained within the drivable area of the lane, and at the same time, due to the good representation of the graph neural network of the mutual influence between vehicles, the predicted trajectories are rarely contradictory to each other.

[0170] In most cases, the proposed PGST algorithm can deal with the obstacle avoidance problem at the intersection, which indicates the stability of the prediction model and the inclusiveness of the decision module for prediction errors. However, the following situations will pose certain difficulties to the decision planning system: (1) the deviation of the prediction is too large, for example, misjudging the intention of the target vehicle or its dynamic intention; (2) the randomness of the social vehicle action, such as simulated fatigue driving, drunk driving, etc., increases the uncertainty of the system; (3) the safety distance reserved by the vehicle is insufficient due to too fast speed. The above situations are very complex, Figure 9 The No. 57 vehicle shown in the simulation is a typical example. In the experiment, a level-k game model is used for modeling. Through the game process, the host vehicle can determine the action that conforms to its maximum benefit, and at the same time, predict the possible action of the target vehicle, thereby helping to update the estimation of the driving style of the target vehicle.

[0171] From Figure 9 As can be seen from the simulation, at the beginning, the host vehicle will be in a dilemma and will face a large number of vehicles from the left branch and the No. 52 social vehicle from the right branch. At the beginning, the host vehicle has the right of way and therefore passes through the conflict area first. Then the host vehicle starts to interact with the No. 52 vehicle (circled). Compared with the No. 52 vehicle, the host vehicle is closer to the conflict area, and at the same time, according to the principle that the vehicle entering from the right branch has priority, the host vehicle makes the decision to slow down and give way. At 6.8s, the host vehicle and the No. 54 vehicle (rhombus marked) are very close to each other, forming a confrontation situation. Because at this time, there is no obvious priority between the two, and the No. 54 vehicle is driving very slowly, it is difficult to identify its true driving intention. For safety considerations, the host vehicle actively reduces the speed to ensure safety. At the same time, the No. 54 vehicle controlled by the random sampling model makes an acceleration response to break the deadlock between the two. Finally, the No. 54 vehicle slightly accelerates through the conflict area, while the host vehicle stops and gives way until 9.8s. After that, the No. 54 vehicle enters the front of the host vehicle's driving position, and the host vehicle starts to follow the driving.

[0172] It can be seen that the automatic driving decision scheme proposed in the present disclosure fully considers the interaction behavior and uncertainty between vehicles in the road driving process, can understand and predict the development trend of the dynamic environment, and the driving decision model guided by the prediction result performs better than human drivers in the test scene.

[0173] To sum up, the above is only a preferred embodiment of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method of a prediction result-oriented action sequence search, characterized by, The method is applied to an automatic driving decision planning system for an uncontrolled intersection, and the system comprises a co-pilot module and a main pilot module, wherein: The co-pilot module is configured to make an auxiliary decision based on situation prediction, including: performing trajectory prediction on a social vehicle that may have a conflict, and determining whether the social vehicle will have a conflict with a current trajectory planning of a controlled vehicle; when a conflict exists, searching for and selecting an optimal action sequence with safety as a primary principle, and outputting the optimal action sequence to the main pilot module; The main pilot module is configured to perform real-time trajectory planning, including: performing trajectory planning and conflict detection on the planned trajectory according to a current observation and an auxiliary decision result of the co-pilot module, and performing action adjustment, i.e., updating the trajectory, when it is determined that a conflict exists; The co-pilot module is operated once every 1 second, and an optimal action sequence within 3 seconds is obtained each time; the main pilot module is operated once every 0.1 second; The system further comprises a social vehicle selection module configured to determine an intention and / or a driving style of a social vehicle; wherein the intention determination is used to further screen vehicles for trajectory prediction, and to determine a sampling path; and the driving style determination is used to control a sampling density of target points; The search method comprises the following steps: Based on each possible action of the controlled vehicle, forward simulation is performed step by step according to vehicle kinematics; After each step of simulation is completed, the safety of each candidate action is evaluated based on a collision possibility between the controlled vehicle and a target vehicle; Based on the evaluated action, the next step of forward simulation search is continued until a series of action sequences that meet the search depth requirement are obtained; The search method specifically comprises the following steps: for the current state of the controlled vehicle generating a set of candidate action pairs A raw wherein, A is a set of longitudinal acceleration actions limited by a maximum jerk, and I is a set of optional lane paths respectively for each candidate action a jd = (a jd , id jd ) ∈ A raw As a starting state, forward simulation is performed according to the vehicle kinematic model to generate the trajectory of the controlled vehicle in the future T period a jd is the longitudinal acceleration, id jd is the lane path number; j is the action number, and d is the search depth; Upon completion of each forward simulation, the trajectory checking for collisions between the controlled vehicle and the target vehicle during the course of the candidate action a jd safety assessment is performed; Based on the candidate action with high safety, the next step of search is continued until the search depth requirement is met.

2. The search method of claim 1, wherein, The safety evaluation method comprises: Each candidate action a jd is accompanied by a list of scores R, the maximum value of which is the final danger score r(a jd ), as shown in equation (4a): In the score list, the dangerous score of the controlled vehicle with the kth predicted trajectory of each target vehicle n is stored It is derived from the mean value of the time dangerous value sequence r n,k , r n,k is the time list of the dangerous value of the controlled vehicle with the kth predicted trajectory of the target vehicle n at all t∈T time wherein, is the risk value of the kth predicted trajectory of the controlled vehicle and the target vehicle n at time t, and the calculation method is shown in equation (4c): wherein, represents the safety circumscribed rectangle of the controlled vehicle at time t, which is the rectangle obtained by expanding the vehicle's own shape size by a safety margin, represents the safety circumscribed rectangle of the target vehicle n; represents the circumscribed rectangle of the controlled vehicle's body, whose shape is determined by the vehicle's own shape size, represents the predicted circumscribed rectangle of the target vehicle's body; represents the probability of the occurrence of this piece of predicted trajectory, represents the predicted collision time; A calculation method of a collision time is shown in formula (4d), the predicted state of the target vehicle n at time t the coordinate value in the controlled vehicle local Frenet coordinate system and the position in the controlled vehicle Frenet coordinate system the difference is calculated and divided by the speed v of the controlled vehicle in the s direction s where s represents the longitudinal axis of the local Frenet coordinate system.

3. The search method of any of claims 1-2, wherein, The length of time for trajectory prediction of the target vehicle is 3 seconds, and the search depth length of the action sequence is 3.

4. A method for automatic driving decision at an intersection without tube control, characterized in that, The method is applied to an automatic driving decision planning system for an uncontrolled intersection, and the system comprises a co-pilot module and a main pilot module, wherein: The co-pilot module is configured to make an auxiliary decision based on situation prediction, including: performing trajectory prediction on a social vehicle that may have a conflict, and determining whether the social vehicle will have a conflict with a current trajectory planning of a controlled vehicle; when a conflict exists, searching for and selecting an optimal action sequence with safety as a primary principle, and outputting the optimal action sequence to the main pilot module; The main pilot module is configured to perform real-time trajectory planning, including: performing trajectory planning and conflict detection on the planned trajectory according to a current observation and an auxiliary decision result of the co-pilot module, and performing action adjustment, i.e., updating the trajectory, when it is determined that a conflict exists; The co-pilot module is operated once every 1 second, and an optimal action sequence within 3 seconds is obtained each time; the main pilot module is operated once every 0.1 second; The system further comprises a social vehicle selection module configured to determine an intention and / or a driving style of a social vehicle; wherein the intention determination is used to further screen vehicles for trajectory prediction, and to determine a sampling path; and the driving style determination is used to control a sampling density of target points; The decision method comprises the following steps: Step S1, virtually dividing lanes for the intersection without lane division; Step S2, selecting a target vehicle; Step S3, performing trajectory prediction of the target vehicle; Step S4, judging whether there is a conflict between the predicted trajectory of the target vehicle and the current trajectory planning of the controlled vehicle; Step S5, when there is a conflict, performing action sequence search with safety as the primary principle; Step S6, selecting an optimal action sequence from a series of action sequences obtained; Performing trajectory prediction of the target vehicle based on a TNT model; The decision method further comprises steps S7 and / or S8, wherein: Step S7, judging the intention of the target vehicle by using a random forest model; based on the intention, further screening the target vehicle for trajectory prediction to determine the sampling path; Step S8, judging the driving style of the target vehicle based on an IMM interactive multi-model; based on the driving style, determining the sampling density of the target sample; The step S8 specifically comprises: Suppose that any one target vehicle n has three possible driving styles, which are aggressive, normal and conservative, denoted as {ξ a ,ξ b ,ξ c}, each driving style ξ corresponds to a certain acceleration range; When a target vehicle is first observed, the probabilities of belonging to the three driving styles are initialized to the same value, after which they are updated based on the observed vehicle state to the actual state reached by the vehicle are compared, and a likelihood Λ is calculated ξ ; The prediction part of the interacting multiple models is then used to update the belonging style probability, where P is the covariance, ∈ is a small constant to avoid division by zero, and ξ ′ is the normalized ξ; The target sample will then be at resolution r = r0 / Λ ξ sampling, where r0is a fixed resolution of adoption.

5. The decision method of claim 4, wherein, Further comprising the following steps: Performing trajectory planning and conflict detection according to the optimal state sequence corresponding to the optimal action sequence and the real-time observation results of the social vehicle; When it is determined that there is a conflict, performing action adjustment, i.e., updating the trajectory.

6. The decision method of claim 5, wherein, When it is determined that there is a conflict, performing action adjustment by using a game model.

Images