Method for planning optimal driving behavior for at least partially autonomous driving vehicle

Abstract

A method for planning an optimal driving behavior for an at least partially autonomous driving vehicle includes: acquiring sensor data, the sensor data including ambient environment information and / or state information of the vehicle; ascertaining a first evaluation function for assigning a goodness to each possible state of the vehicle at discrete time points within a planned range of the vehicle on the basis of the sensor data; calculating a local optimal value of the first evaluation function at each discrete time point, wherein the local optimal value represents a possible state of the vehicle at the assigned discrete time point; solving a candidate trajectory, wherein the candidate trajectory is composed of a time sequence of local optimal values; evaluating the candidate trajectory by means of a second evaluation function, the second evaluation function evaluating a state transition between successive states of the candidate trajectory; selecting an optimal candidate trajectory from the candidate trajectories based on the first and second evaluation functions; and transmitting the selected optimal candidate trajectory to a control unit of the vehicle for planning the optimal driving behavior of the vehicle.

Description

Technical Field

[0001] The invention relates to a method for planning an optimal driving behavior for an at least partially autonomously driven vehicle. Background Art

[0002] One possible goal for autonomous driving of a vehicle is to control the vehicle based on sensor data in such a way that defined goals are achieved as quickly, comfortably and safely as possible without, for example, causing a collision or violating traffic regulations.

[0003] The driving task can be subdivided into the subtasks perception, prediction, planning and control. The task of perception is to extract relevant information from the sensor data, such as the position of objects (other vehicles or traffic participants), identify lane markings and recognize traffic signs or the like. Since the identified objects are usually dynamic obstacles, it is then necessary to predict their future positions so that collisions can be avoided. On this basis, the task of planning is to generate the trajectory that should be adjusted in the control subtask.

[0004] A special method in the KI-based approach for trajectory planning is the "end-to-end interpretable neural motion planner" (NMP) made public by Uber ATG. Here, the end-to-end approach covers all of the subtasks mentioned above and, based on so-called lidar point clouds and HD maps, generates predictions for other traffic participants on the one hand and so-called cost quantities on the other hand, wherein the HD maps contain information, for example, about the road layout, speed limits, real-time traffic information, etc. For each point in time in the planning horizon, the cost quantity assigns a cost value to each possible position in the field of view, which indicates how favorable (or unfavorable) it is to be at this location at this point in time. Next, the different trajectories that the vehicle can implement in principle are evaluated based on this cost quantity. The best trajectory is forwarded to the controller, which should control the vehicle according to the plan.

[0005] The known NMP approach uses a gridded input from a bird's eye view for this purpose and uses convolutional neural networks (CNNs) for information processing. In the field of prediction, vectorized inputs processed with the aid of graph neural networks are also increasingly used. Here, the road image is linked to information about the vehicle's own history and other agents. Next, for each connection of adjacent nodes in the image, the probability is determined that the connection is part of the future vehicle's own trajectory with this probability. On this basis, in PGP, different passes through the image are used as the basis for predicting different trajectories.

[0006] However, the weakness of NMP is that the plan is generated only based on sensor data and map information. Accordingly, the trajectory is selected without taking into account the desired route. Therefore, although NMP enables collision-free driving, it does not provide the possibility of taking into account the given navigation goals.

[0007] Another problem is the marginalization when generating cost quantities and predictions. This has the following reasons: In many situations, different behavior possibilities are possible for the vehicle. Thus, it is also possible to follow the current lane and change to another motorway. In addition, other traffic participants can act and react in different ways. Since the NMP only generates cost quantities or only predictions, these are marginalizations of all possible own intentions and all developments of the entire scenario.

[0008] In addition, in the NMP, the trajectory for the vehicle is selected from a set of randomly generated drivable trajectories in such a way that these trajectories are evaluated according to the cost amount and the best trajectory is then selected. This is time-consuming and computationally intensive on the one hand, and on the other hand, the goodness of the selected trajectory is therefore directly related to the candidates contained in the set. If the candidate is too simple, the trajectory thus selected may be significantly worse than the optimal value according to the given cost amount. If the candidate is very complex (and, for example, also contains an S curve), it may happen that, due to marginalization, a trajectory that mixes different modes is selected. If the modes "deceleration" and "acceleration" have the lowest cost respectively, a trajectory that first decelerates and then accelerates may be selected as the best candidate, although this behavior is not allowed in the given situation. By generating the own trajectory without evaluating the candidates and by specifically distinguishing the different modes from each other, such a driving behavior of the vehicle can be prevented. Summary of the invention

[0009] It is therefore an object of the present invention to provide a solution by means of which an improved trajectory can be generated in an efficient manner for mapping an optimal driving behavior for an at least partially autonomously driven vehicle.

[0010] This object is achieved by the method according to the invention for planning an optimal driving behavior for an at least partially autonomously driven vehicle.

[0011] According to a first aspect, the present disclosure relates to a method for planning an optimal driving behavior for an at least partially autonomously driven vehicle, the method having the following steps:

[0012] In a first step, sensor data are acquired which contain information about surroundings information and / or status information of the vehicle.

[0013] In a second step, a first evaluation function is ascertained, which is designed to assign a goodness to each possible state of the vehicle at a discrete point in time within a planning range for the vehicle based on the sensor data.

[0014] In a third step, at least one local optimum value of the first evaluation function at each discrete time point is determined, wherein the at least one local optimum value represents a possible state of the vehicle at the associated discrete time point.

[0015] In a fourth step, at least one candidate trajectory is determined, wherein the candidate trajectory consists of a time series of local optimal values.

[0016] In a fifth step, at least one candidate trajectory is evaluated with the aid of a second evaluation function which evaluates state transitions between successive states of the at least one candidate trajectory.

[0017] In the sixth step, at least one optimal candidate trajectory is selected from the at least one candidate trajectory based on the first and second evaluation functions.

[0018] In a seventh step, the selected at least one optimal candidate trajectory is transmitted to a control unit of the vehicle for planning an optimal driving behavior of the vehicle.

[0019] Below, some basic aspects of the present invention are described:

[0020] The invention for planning a (self) trajectory for a vehicle is first based on the known NMP approach and expands the NMP approach with the aid of methods from the PGP approach. The initial basis is the marginalized cost quantity.

[0021] First, for different modes or behavior possibilities, the cost quantity is reduced to potential waypoints of the vehicle. These waypoints are identified in each time step by, for example, applying non-maximum suppression (NMS) to the cost. Since these potential waypoints may belong to different modes, they cannot be arbitrarily linked without creating a mixture of different modes. In order to be able to distinguish between modes, a probability is determined for each pair of potential waypoints in successive time steps, similar to PGP, which describes how likely it is that these waypoints belong to the same mode or how likely it is that two waypoints are connected by a trajectory.

[0022] By combining the identified potential waypoints and the probability that these waypoints belong to the same mode, the planning of the trajectory can also be graphically illustrated as a traversal of a tree whose nodes are waypoints and whose edge weights are probabilities. The current position of the vehicle is used as the root node. First, the edge weights of the root node and all possible first waypoints are set to zero. Next, for each of the waypoints in the last time step, the optimal traversal through the tree is determined, for example by means of dynamic programming. The traversals correspond to different modes. Next, by studying only the optimal traversal, it is ensured that impermissible mixtures of modes are no longer considered. In addition, by this selection, it is not necessary to study all permutations of potential waypoints. The number of optimal traversals corresponds to the number of possible waypoints in the last time step.

[0023] Each traversal forms a possible rough trajectory. In the next step, the waypoints are decoded from this, it being ensured that the resulting trajectory is drivable and comfortable and that it follows the desired route. For this purpose, a neural network, for example an MLP (Multilayer Perceptron), is trained which decodes drivable trajectories for the vehicle from the coordinates and costs of the nodes. For each of these trajectories, the deviation of the final position from the route is then compared and the optimal trajectory is selected for the vehicle in this way.

[0024] With the help of Figure 3 The corresponding architecture which basically maps the above scenario is shown graphically.

[0025] A possible embodiment of the method provides that the information for the sensor data includes at least one map information. In this way, the trajectory to be generated can be set optimally and accurately according to different scenarios.

[0026] One possible embodiment of the method provides that the first evaluation function is respectively designed as a cost variable. This enables efficient planning of the driving behavior of the vehicle.

[0027] A possible embodiment of the method is that the time series of local optima consists of a first local optimum and a second local optimum, which are arranged relatively close to one another, thereby achieving a mapping of possible driving behaviors of the vehicle that is as realistic as possible.

[0028] One possible embodiment of the method provides that, before the step of transmitting, at least one optimal candidate trajectory is optimized, thereby further improving the planned driving behavior of the vehicle and increasing the user-friendliness for the driver of the vehicle.

[0029] One possible embodiment of the method provides that the first evaluation function is mapped by a neural network. This achieves the advantage that the planning of the optimal driving behavior is automated.

[0030] According to a second aspect, the present disclosure relates to a computer program product, which contains machine-readable instructions, which, when executed on one or more computers and / or computer instances, cause the one or more computers or computer instances to implement a method according to the present invention.

[0031] According to a third aspect, the present disclosure relates to a machine-readable data carrier and / or a download product having the computer program product.

[0032] According to a fourth aspect, the disclosure relates to one or more computers and / or computer instances with the computer program product and / or with the machine-readable data carrier and / or download product.

[0033] Further measures for improving the invention are shown in more detail below together with the description of preferred exemplary embodiments of the invention with reference to the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings show:

[0035] Figure 1 A schematic flow chart showing a method 100 for planning optimal driving behavior for an at least partially autonomously driven vehicle 10 ;

[0036] Figure 2 shows a schematic overview of a module-based architecture for the method according to the invention according to one embodiment of the invention; and

[0037] Figure 3 A schematic diagram shows an architecture for a method 100 according to the present invention according to one specific embodiment of the present invention. DETAILED DESCRIPTION

[0038] Figure 1 A schematic flow chart of a method 100 for planning an optimal driving behavior for an at least partially autonomously driven vehicle 10 is shown.

[0039] In a first step 102, sensor data 12 are acquired, which contain information 13 about surroundings and / or status information of vehicle 10. This information 13 may be designed, for example, as map information, which is generated by vehicle 10 itself and / or partially by external service providers and transmitted to vehicle 10.

[0040] In a second step 104, a first evaluation function 14 is determined, which is constructed to assign a goodness to each possible state of the vehicle 10 at a discrete time point 15 within a planning horizon of the vehicle 10 based on the sensor data 12. Here, for example, sensor data 12, such as HDMap map information, can be used as input. A corresponding cost volume is generated as output. Here, the planning horizon can have a spatial component and / or a temporal component. Here, the first evaluation function 14 can be constructed as a cost volume.

[0041] In a third step 106, at least one local optimum 16 of the first evaluation function at each discrete time point 15 is determined 106, wherein the at least one local optimum 16 represents a possible state of the vehicle 10 at the assigned discrete time point 15. In this case, the possible state of the vehicle 10 can be detected by different sensors, such as a lidar sensor, a radar sensor or a visual sensor, such as an RGB camera. More generally, in this context, it can also be said that one aspect of the present invention is to optimize the trajectory into states or state transitions (dynamics) with the help of a suitable evaluation function. For this purpose, a correspondingly suitable optimization method can be applied.

[0042] In a fourth step 108 , at least one candidate trajectory 17 is determined, wherein the candidate trajectory 17 is composed of a temporal sequence of local optima 16 . This may involve a corresponding combinatorial analysis.

[0043] Alternatively, the time series of local optimum values ​​16 may consist of a first local optimum value 16 - 1 at a first discrete time point 15 - 1 and a second local optimum value 16 - 2 at a second discrete time point 15 - 2 , which are arranged relatively close to each other.

[0044] In a fifth step 110, at least one candidate trajectory 17 is evaluated using a second evaluation function 20, which evaluates state transitions between successive states of at least one candidate trajectory 17. Optionally, second evaluation function 20 indicates the quality of at least one candidate trajectory 17.

[0045] In a sixth step 112 , at least one optimal candidate trajectory 22 is selected from the at least one candidate trajectory 17 based on the first and second evaluation functions 14 , 20 .

[0046] In a seventh step 114 , the selected at least one optimal candidate trajectory 22 is transmitted to the control unit 11 of the vehicle 10 for planning an optimal driving behavior of the vehicle 10 .

[0047] Optionally, before the step 114 of conveying, for example by leveling An optimization of at least one optimal candidate trajectory 22 may be achieved. In addition, optionally, based on the combined cost function, a local optimization of at least one optimal candidate trajectory 22 may be achieved.

[0048] Figure 2 A schematic and exemplary overview of a module-based architecture for the method according to the invention is shown according to one specific embodiment of the invention.

[0049] Here, the various modules of the architecture can be trained as follows:

[0050] - Cost volume prediction: Similar to the original NMP with maximum margin loss.

[0051] -Target set prediction: no trainable parameters.

[0052] -Action Scoring: Negative Log Likelihood with Teacher Forcing: The nodes closest to the ground-truth waypoint are shifted to the ground-truth coordinates. Next, the negative log-likelihood of the connections to the ground-truth nodes is optimized.

[0053] -Candidate plan generation: no trainable parameters.

[0054] - Trajectory Decoder: L2 Loss: The L2 loss of the waypoints is optimized for the traversal with the minimum distance to the ground truth.

[0055] Figure 3 Show reference Figure 2 A table of FIG. 1 is a schematic diagram of an architecture for a method 100 according to the present invention according to one specific embodiment of the present invention.

[0056] Here, the image is shown in Figure 2 The individual modules listed in the example have corresponding inputs and outputs of the architecture so as to obtain an optimal candidate trajectory for the vehicle at the output of the architecture based on sensor data at the input, such as map information.

Claims

1. A method (100) for planning an optimal driving behavior for an at least partially autonomously driven vehicle (10), the method comprising the following steps: - acquiring (102) sensor data (12), said sensor data comprising information about the surroundings of said vehicle (10) and / or status information (13) of said vehicle; - determining (104) a first evaluation function (14), which is designed to assign a goodness to each possible state of the vehicle (10) at a discrete time point (15) within a planning range of the vehicle (10) based on the sensor data (12); - obtaining (106) at least one local optimum value (16) of the first evaluation function at each discrete time point (15), wherein: The at least one local optimum (16) represents a possible state of the vehicle (10) at an associated discrete point in time (15); - determining (108) at least one candidate trajectory (17), wherein the candidate trajectory (17) consists of a time series of the local optimum (16); - evaluating (110) the at least one candidate trajectory (17) by means of a second evaluation function (20), the second evaluation function evaluating state transitions between successive states of the at least one candidate trajectory (17); - selecting (112) at least one optimal candidate trajectory (22) from the at least one candidate trajectory (17) based on the first and second evaluation functions (14, 20); and - transmitting (114) the selected at least one optimal candidate trajectory (22) to a control unit (11) of the vehicle (10) for planning an optimal driving behavior of the vehicle (10).

2. The method (100) according to claim 1, wherein: The information for the sensor data (12) includes at least one map information.

3. The method (100) according to any one of the preceding claims, wherein: The first evaluation functions (14) are each designed as a cost variable.

4. The method (100) according to any one of the preceding claims, wherein: Prior to the step of transmitting (114), the at least one optimal candidate trajectory (22) is optimized.

5. The method (100) according to any one of the preceding claims, wherein: The time series of the local optimum (16) consists of a first local optimum (16-1) and a second local optimum (16-2), the first local optimum and the second local optimum being arranged relatively close to each other.

6. The method (100) according to any one of the preceding claims, wherein: The first evaluation function (14) is mapped by a neural network.

7. A computer program product, comprising machine-readable instructions, which, when executed on one or more computers and / or computer instances, cause the one or more computers and / or computer instances to implement the method according to any one of claims 1 to 6.

8. A machine-readable data carrier and / or a download product having a computer program product according to claim 7.

9. One or more computers and / or computer instances having a computer program product according to claim 7 and / or having a machine-readable data carrier and / or a download product according to claim 8.

Images