Device and computer-implemented method for determining trajectory for vehicle
By combining rule-based models and neural networks, and using a hybrid A* algorithm to optimize vehicle trajectories, the problem of balancing safety and optimization in existing technologies is solved, achieving a balance between safety and optimization and improving the flexibility and reliability of trajectory planning.
Patent Information
- Application Number
- CN202510580853.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-05-07
- Filing Date
- 2025-05-07
- Publication Date
- 2025-11-07
AI Technical Summary
Existing technologies struggle to simultaneously guarantee safety and trajectory optimization when determining vehicle trajectories, and they lack consideration for implicit quality requirements.
A rule-based model combined with neural networks is used to plan and optimize vehicle trajectories based on environmental information and behavioral patterns. A hybrid A* algorithm is used for local optimization and continuation, and a machine learning model is used for cost assessment and trajectory adjustment.
This approach optimizes vehicle trajectories while ensuring safety, improves the solvability and conformity of trajectories to human driving habits, and enhances the flexibility and reliability of trajectory planning.
Smart Images
Figure CN120902745A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The invention is based on an apparatus and a computer-implemented method for determining a trajectory for a vehicle. BACKGROUND
[0002] C. Hubschneider et al. (2017) published “Integrating end-to-end learned steering into probabilistic autonomous driving” discloses a possibility to determine a trajectory for a vehicle based on video data. Here, deep learning and classical trajectory determination schemes are combined by a factor graph approach. SUMMARY
[0003] The apparatus and the computer-implemented method for determining a trajectory for a vehicle provide an alternative for determining a trajectory which does not require a factor graph approach and which is also able to provide a safety guarantee for the trajectory.
[0004] The computer-implemented method for determining a trajectory for a vehicle is provided to:
[0005] - provide an environment model, wherein the environment model contains environment information about an environment surrounding the vehicle;
[0006] - provide at least one behavior pattern;
[0007] - plan a trajectory for the vehicle based on the environment information using an artificial neural network; or
[0008] - plan a trajectory for the vehicle based on the behavior pattern using a rule-based model;
[0009] - select and / or modify the trajectory based on the environment information and the trajectory and according to the behavior pattern and according to a cost using a rule-based model.
[0010] The rule-based model is a model which comprises:
[0011] - physical laws;
[0012] - rules, e.g. traffic rules;
[0013] - heuristics, e.g. heuristics about what is considered to be a comfortable driving;
[0014] - parameters, e.g. vehicle parameters.
[0015] These parameters can also be learned or be able to be learned. For example, in a rule-based model, the second trajectory is generated by an optimization algorithm. Examples for rule-based algorithms are the hybrid A* algorithm or model predictive control (MPC). Rule-based models have the advantage that they are understandable for humans. Rule-based models serve the safety goal, but also optimize the performance of the second trajectory.
[0016] The neural network can have learned implicit quality requirements for particularly comfortable / human-like driving. This can be achieved, for example, by an Immitation Learning scheme. For this purpose, a data set is used for training which contains, on the one hand, respective current situations, for example in the form of an environmental model, and, on the other hand, trajectories driven by human drivers.
[0017] By using a rule-based model and a neural network, both rule-based modeled trajectories and learned trajectories are modeled. This means that not only explicitly describable aspects, but also only implicitly learnable aspects are taken into account together.
[0018] For example, it is provided that:
[0019] - a plurality of behavior patterns with different priorities is predefined;
[0020] - a trajectory is determined and / or modified for each behavior pattern.
[0021] For example, it is provided that:
[0022] - a list of prioritized behavior patterns is predefined;
[0023] - a trajectory is determined and / or modified for each behavior pattern;
[0024] - the trajectory assigned to the behavior with the highest priority in the list is selected as the trajectory for the vehicle.
[0025] For example, the costs are determined depending on the environmental information and / or depending on the behavior pattern.
[0026] For example, a part of the costs is modeled using a machine learning model which is trained for assigning a respective cost depending on the environmental information and / or depending on the behavior pattern.
[0027] For example, a part of the costs is modeled using a rule-based model which is configured for assigning a respective cost depending on the environmental information and / or depending on the behavior pattern.
[0028] It can be provided that the behavior pattern is selected from a plurality of predefined behavior patterns depending on the costs.
[0029] It can be provided that the vehicle performs contingency planning, wherein at least one safety goal and at least one goal characterizing a performance of a trajectory are predefined, wherein a trajectory is planned in a first time range, which trajectory realizes the goal characterizing the performance as well as possible and which trajectory fulfils the at least one safety goal in the first time range, wherein a continuation of the trajectory planned in the first time range is planned in a second time range, which is longer compared to the first time range, which continuation is allowed to realize the goal characterizing the performance less well compared to the trajectory planned in the first time range, wherein a modified trajectory is determined as a trajectory for the vehicle until the end of the second time range before the end of the first time range using a rule-based model on the basis of the environmental information and the trajectory planned in the second time range and according to a behavior and according to a cost, or wherein the trajectory planned in the second time range is determined as the modified trajectory for the vehicle until the end of the second time range when no modified trajectory has been determined as a trajectory for the vehicle until the end of the second time range until the end of the first time range. The continuation can be a multimodal continuation. For example, the continuation can be planned by means of a hybrid A* algorithm. In an example, the continuation is not implemented at first, but is arranged by replanning. However, the continuation guarantees that even at the end of the short-term planning range, i.e. the first time range, a solution with an acceptable performance still exists. This is an advantage.
[0030] By rule-based continuation of the trajectory by means of the modified trajectory, the vehicle generally follows a trajectory with a particularly high performance. There is no strict requirement for the planning range to be necessary. By means of the continuation, solvability of the planning problem is also ensured after the planning range of the network when the trajectory is determined by a neural network.
[0031] For example, it is provided that:
[0032] - the environmental information contains information about the vehicle's surroundings at a first point in time;
[0033] - a plurality of trajectories for the vehicle is determined on the basis of the information about the vehicle's surroundings at the first point in time;
[0034] - the vehicle is moved on a first trajectory;
[0035] - during the movement of the vehicle on the first trajectory, environmental information is determined, which contains information about the vehicle's surroundings at a second point in time;
[0036] - a second trajectory is determined on the basis of the information about the vehicle's surroundings at the second point in time;
[0037] - and not moving the vehicle on the first trajectory, but moving the vehicle on a second trajectory.
[0038] For the training, it can be provided that training data comprising a reference trajectory for a trajectory of the vehicle are provided, wherein the reference trajectory simulates a driving behavior of a human or represents a detected driving behavior of a human, wherein the artificial neural network is trained for determining a trajectory for the vehicle which is in as good an agreement as possible with the reference trajectory.
[0039] An apparatus for determining a trajectory for a vehicle, in particular a control device, is configured to carry out the method. BRIEF DESCRIPTION OF DRAWINGS
[0040] Further advantageous embodiments can be derived from the following description and the attached drawings. In the drawings, the following is shown:
[0041] Figure 1 Schematic illustration of an apparatus for determining a trajectory for a vehicle;
[0042] Figure 2 Schematic illustration of an architecture for determining a trajectory for a vehicle;
[0043] Figure 3 Flowchart with steps of a method for determining a trajectory for a vehicle;
[0044] Figure 4 Example for planning a trajectory for a vehicle. DETAILED DESCRIPTION
[0045] In Figure 1 An apparatus 100 for determining a trajectory for a vehicle 102 is schematically shown in
[0046] The apparatus 100 is, for example, a control device of the vehicle 102.
[0047] In this example, the apparatus 100 is configured to determine a trajectory 104 for the vehicle 102.
[0048] In this example, the apparatus 100 is configured to move the vehicle 102 on the trajectory 104.
[0049] In Figure 2 An architecture 200 for determining a trajectory for a vehicle 102 is schematically shown in
[0050] The architecture 200 comprises an environment model 202, a behavior generation module 204 for pre-specifying at least one behavior pattern, a model 206 for determining a trajectory for the vehicle 102, a cost function 208, and a behavior verification and adjustment module 210.
[0051] The environment model 202 contains environment information about the environment surrounding the vehicle.
[0052] The behavior generation module 204 is configured for predefining at least one behavior pattern 214.
[0053] The behavior generation module 204 is configured, for example, for predefining one boundary condition or a plurality of boundary conditions as a behavior pattern 214 or a prioritized list of behavior patterns. In this context, "safe expansion" means, for example, that the vehicle 102 is able to move in compliance with traffic safety standards when moving in accordance with the behavior pattern 214.
[0054] The cost function 208 is configured for predefining a cost 216 for a trajectory.
[0055] The cost function 208 comprises a machine learning model 208-1 which is configured to determine a portion of the cost 216 depending on the environment information 212 and / or depending on the at least one behavior pattern 214. In this example, the machine learning model 208-1 is trained for determining this portion of the cost 216 depending on the environment information 212 and / or depending on the at least one behavior pattern 214.
[0056] The machine learning model 208-1 for determining this portion of the cost 216 is configured for determining this portion of the cost based on a trajectory predefinable by the model 206 for determining a trajectory.
[0057] The cost function 208 comprises a rule-based model 208-2 which is configured to determine another portion of the cost 216 depending on the environment information 212 and / or depending on the at least one behavior pattern 214.
[0058] In this example, the cost function aggregates the portions of the cost.
[0059] An implementation variant without the machine learning model 208-1 can be provided.
[0060] In an implementation variant of the machine learning model 208-1, for example, a grid map is learned for prioritized regions and velocities which then enters into the cost 216 by a corresponding weighting.
[0061] Another input variable in the machine learning model 208-1 can be a risk estimate for the surrounding environment. In the grid map, regions can be marked which have a potentially high risk, for example, a field in front of a kindergarten, or a field which is particularly frequently subjected to difficult / complex situations in the training data.
[0062] By taking the risk into account in the cost 216, it is then possible to prioritize avoiding these regions.
[0063] The rule-based model 208-2 for determining this part of the cost 216 is configured to determine this further part of the cost based on expert knowledge.
[0064] The rule-based model 208-2 can use a grid map which assigns a specific cost value to each modeled state of the vehicle 102.
[0065] If the environment model 202 contains uncertainties about estimated and possibly predicted states of other road users, these uncertainties can likewise be incorporated into the cost 216 by means of a corresponding modeling with the rule-based model 208-2.
[0066] For example, a probabilistic occupancy risk resulting from the uncertainties of the environment model 202 and their prediction is recorded as a cost to a grid map. The grid map, for example, comprises cells. For example, a different representation is chosen for each cell for the probabilistic occupancy risk. For example, a pure occupancy probability is set for each cell or a more detailed and more convincing representation by means of Subjective Logic Opinions.
[0067] It can be provided that the grid map is stretched over the current state of the vehicle 102 and over a time window of the last N states of the vehicle 102 and the current state. Thereby, the interaction with other vehicles can be depicted over a time range of N time steps. In particular, the cost 216 of the future states can be conditioned on the last N time steps of a trajectory planned by the vehicle 102.
[0068] With the cost function 208, so-called emergency planning can be taken into account. Emergency planning means that an initial trajectory is first generated for the near future which must satisfy special safety requirements. Subsequently, the initial trajectory is further planned for various possible cases in order to thus antizipativly achieve a solution with as high performance as possible. Since the combinatorial complexity constantly increases with each branching of possible cases, it is necessary to perform a final state evaluation from a certain search depth. The final state evaluation can be carried out by means of the cost 216 in the cost function 208.
[0069] In order to generate an initial solution for the subsequent emergency planning, a combination of trajectory planning with machine learning models and rule-based models can be used. For example, with the rule-based model, a trajectory is determined taking into account the previously selected behavior modes 214 according to the cost 216. Here, the restrictions of the selected individual behavior modes 214 are considered, for example, in a short time period in a comprehensive manner and in a subsequent time period separately.
[0070] The model 206 for determining a trajectory is configured to determine at least one trajectory 218.
[0071] The model 206 for determining a trajectory comprises an artificial neural network 206-1 which is configured to plan a trajectory for the vehicle 102 based on the environment information 212.
[0072] The artificial neural network 206-1 is trained to plan a trajectory which is similar to a trajectory that a human would plan based on the environment information 212.
[0073] The model 206 for determining a trajectory comprises a rule-based model 206-2 which is configured to plan a trajectory for the vehicle based on the environment information 212 and the first trajectory and in accordance with the behavior pattern 214 and in accordance with the costs 216 for the trajectory.
[0074] The rule-based model 206-2 for determining a trajectory is configured to determine a trajectory for the vehicle 102 such that adherence to the pre-specified behavior pattern 214 is guaranteed.
[0075] It is provided in this example that a plurality of trajectories 218 is determined using the artificial neural network 206-1 and the rule-based model 206-2.
[0076] It can be provided that the model 206 for determining a trajectory comprises either the artificial neural network 206-1 or the rule-based model 206-2.
[0077] The behavior verification and adjustment module 210 is configured to determine a trajectory 104. In this example, the behavior verification and adjustment module 210 comprises a rule-based model 210-1 for determining a trajectory 104 for the vehicle 102 in accordance with the at least one trajectory 218 and the costs 216.
[0078] The rule-based model 210-1 for determining a trajectory 104 for the vehicle 102 is configured to determine a trajectory 104 for the vehicle 102 such that a drivable behavior of the vehicle 102 is guaranteed.
[0079] The rule-based model 210-1 uses a hybrid A* algorithm for trajectory generation, for example, starting from an initial trajectory.
[0080] Accordingly, the initial trajectory is first converted into an initial solution in the search space of the hybrid A* algorithm. This is achieved, for example, by suitable sampling or by a direct conversion of the control points (Stützstellen) of the initial trajectory 104. A possible trajectory is created by locally optimizing the initial solution by means of the hybrid A* algorithm. The initial solution and the local optimization are defined by active nodes in the search space of the hybrid A* algorithm. Here, the nodes are considered active when they satisfy all boundary conditions imposed on them. These boundary conditions can, inter alia, originate from the behavior patterns 214 generated by the behavior generation module 214.
[0081] If a node is not active, it is discarded in the optimization. If the node representing the state of the vehicle 102 is also not active, the behavior pattern 214 from which the boundary condition originated is discarded.
[0082] In the local optimization, the cost function 208 is chosen, for example, such that as many nodes of the initial solution as possible are used. For example, a state cost function, i.e. a cost for the respective node, goes into the cost function 208. However, the cost function 208 can have other terms. For example, the cost function 208 can contain a term that favors the use of as many nodes of the initial solution as possible. The hybrid A* algorithm, for example, only explores within the boundary conditions predefined by the respective behavior pattern 214. Accordingly, the optimized solution and thus the driving movement of the final vehicle 102 on the trajectory thus determined inherits the safety guarantees defined with the boundary conditions.
[0083] In addition, the optimization can be set up to also explore nodes that continue the initial solution into another behavior, which is currently, for example, not selectable due to a lack of visibility, although it is foreseeable. In this way, a preferred behavior can already be steered in advance by the local optimization of the initial solution.
[0084] The rule-based model can be implemented on the basis of various optimization algorithms, such as the hybrid A* search, the fast exploration random tree or the Monte Carlo tree search. The machine learning model can be trained by deep learning methods, such as imitation learning or deep reinforcement learning or value function learning.
[0085] It can be provided that, first, proxy variables are learned, which are then integrated into the state cost function by the respective rule-based modeling. For example, a Preferred Region Grid is learned about the priority areas and a Preferred Speed Grid about the preferred and / or maximum speeds as proxy variables.
[0086] The device 100 is configured to implement the method for determining a trajectory.
[0087] In Figure 3 A flowchart showing the steps of the method is shown in
[0088] The method is premised on predefined environment information 202. How to determine the environment information 202 is not subject of the method. In connection therewith, the following description of the environment information 202 is thus exemplary. The method is premised on predefined at least one behavior pattern 214. How to determine the behavior pattern 214 is not subject of the method. In connection therewith, the following description of the behavior pattern 214 is thus exemplary. The method is premised on assigned costs 216. How to determine the costs 216 is not subject of the method. In connection therewith, the following description of the costs 216 is thus exemplary.
[0089] The method optionally comprises a step 300.
[0090] In the optional step 300, the artificial neural network 206-1 is trained on training data to determine trajectories for the vehicle 102 that are in as good agreement as possible with reference trajectories in the training data.
[0091] For example, the training data is provided comprising reference trajectories for trajectories of the vehicle 102, which reference trajectories simulate human driving behavior.
[0092] For example, the training data is provided comprising reference trajectories for trajectories of the vehicle 102, which reference trajectories represent detected driving behavior of humans. This enables to produce trajectories of the vehicle 102 that are particularly human-like in their driving style.
[0093] It can be provided that the artificial neural network 206-1 is provided already trained.
[0094] It can be provided that the machine learning model 208-1 is trained in the step 300.
[0095] For example, the training data is provided comprising environment information 212 and / or behavior patterns 214 assigned to the reference trajectories.
[0096] It can be provided that the machine learning model 208-1 is trained to assign respective costs depending on the environment information 212 and / or depending on the behavior patterns 214 in the training data.
[0097] It can be provided that the machine learning model 208-1 is provided already trained.
[0098] The method comprises a step 302.
[0099] In the step 302, the environment model 202 is provided.
[0100] The environment model 202 contains environment information 212 about the environment surrounding the vehicle.
[0101] The method comprises a step 304.
[0102] In step 304, at least one behavior 214 is provided.
[0103] For example, a plurality of behaviors 214 is predefined with different priorities.
[0104] It can be provided that a list of behaviors 214 is predefined that distinguishes the priorities.
[0105] The method optionally comprises a step 306.
[0106] In step 306, a trajectory is planned with the artificial neural network 206-1 based on the environment information 212.
[0107] It can be provided that the trajectory is determined with the artificial neural network 206-1 based on the behavior 214.
[0108] It can be provided that, when a plurality of behaviors 214 is predefined with different priorities, a trajectory is determined for each behavior 214.
[0109] The method optionally comprises a step 308.
[0110] Steps 306 and 308 are optional, as the method comprises at least one of these steps.
[0111] In step 308, a trajectory is planned for the vehicle 102 with the rule-based model 206-2.
[0112] For example, the trajectory is planned with the rule-based model 206-2 depending on the behavior 214 and depending on the cost 216.
[0113] In this example, the cost 216 is determined depending on the environment information 212 and / or depending on the behavior 214.
[0114] In this example, a part of the cost is modeled with the machine learning model 208-1.
[0115] In this example, a part of the cost is modeled with the rule-based model 208-2.
[0116] It can be provided that, when a plurality of behaviors 214 is predefined with different priorities, a trajectory is determined for each behavior 214.
[0117] It can be provided that the behavior 214 for which a trajectory is determined is selected from the plurality of behaviors 214 depending on the cost 216.
[0118] In this context, "planning" for example means that, with the rule-based model 206-2, a trajectory is selected from the trajectories determined for each behavior pattern 214 based on the environmental information 212 and the trajectory itself and in accordance with the behavior pattern 214 and in accordance with the costs 216.
[0119] It can be provided that the trajectory is planned by modifying the trajectory determined as described above with the rule-based model 210-1 based on the environmental information 212 and the trajectory itself and in accordance with the behavior pattern 214 and in accordance with the costs 216.
[0120] It can be provided that the trajectory is not modified, but is only selected when a better solution, i.e. a trajectory, is not found with the rule-based model 210-1 based on the environmental information 212 and the trajectory itself and in accordance with the behavior pattern 214 and in accordance with the costs 216. One example of a better solution is a trajectory with more favorable costs, for example lower costs, compared to another trajectory.
[0121] The method comprises a step 310.
[0122] In step 310, a trajectory 104 for the vehicle 102 is determined.
[0123] In this example, the trajectory 104 is determined with the rule-based model 210-1 for determining a trajectory 104 for the vehicle 102 such that a drivable behavior of the vehicle 102 is guaranteed.
[0124] In this example, the trajectory or trajectories previously determined with the artificial neural network 206-1 or determined with the rule-based model 206-2 are selected and / or modified with the rule-based model 210-1 for determining a trajectory 104 for the vehicle 102. The selected or modified trajectory represents the trajectory 104 determined with the rule-based model 210-1 for determining a trajectory 104 for the vehicle 102.
[0125] It can be provided that, with the rule-based model 210-1 for determining a trajectory 104, a trajectory determined with the rule-based model 206-2 based on the environmental information 212 is selected and / or modified in accordance with a trajectory planned by the artificial neural network 206-1 based on the environmental information 212.
[0126] There are various possibilities for the modification. Firstly, the trajectory can be locally optimized based on the costs 216. The local optimization is carried out for example by means of an optimal control scheme. Alternatively, a graph search can be used, for example a hybrid A* algorithm.
[0127] Another possible modification is a continuation of the trajectory. This modification is particularly relevant when the continuation is required by a boundary condition of the behavior pattern, whereas the artificial neural network 206-1 plans not far enough due to a limited planning horizon. For the continuation, the continuation of the trajectory is determined, for example, using a hybrid A* algorithm.
[0128] For example, when a list of prioritized behavior patterns 214 is predefined, the modified trajectory determined for the respective behavior pattern 214 is selected, which is attributed to the behavior having the highest priority in the list, as the trajectory 104 for the vehicle 102.
[0129] It can be provided that a modified trajectory is determined in parallel for each of the predefined behavior patterns. It can be provided that the rule-based model 210-1 for determining the trajectory 104 for the vehicle 102 selects from the modified trajectories determined in parallel the modified trajectory that best meets the predefined safety target as the trajectory 104.
[0130] The selection of the trajectory is made by the fact that the behavior patterns 214 have been prioritized according to the achievement of the safety target, and the trajectory belonging to the behavior pattern 214 having the highest priority is selected.
[0131] The safety target is, for example:
[0132] - compliance with traffic rules;
[0133] - avoidance of hindering other road users;
[0134] - avoidance of accidents.
[0135] It can be provided that the selected or modified trajectory is subsequently optimized again. "Optimization" means, for example, that a cost function depending on the trajectory, for example the cost 206, is defined. The cost function is minimized in the optimization. This means that the selected or modified trajectory is used as an initial value in the optimization in order to find a trajectory that is optimal in terms of the cost function compared to the selected or modified trajectory, which then replaces the selected or modified trajectory.
[0136] The method comprises a step 312.
[0137] In step 312, the vehicle 102 is moved on the trajectory 104 for the vehicle 102.
[0138] The method can be provided that the trajectory 104 on which the vehicle 102 is moved is modified in accordance with the environmental information 212.
[0139] To this end, at a first time instant, environment information 212 is detected. The environment information 212 detected at the first time instant contains information about the vehicle's surroundings at the first time instant.
[0140] Subsequently, a trajectory is determined based on the information about the vehicle's surroundings at the first time instant using a method for determining a trajectory.
[0141] Subsequently, during the movement of the vehicle 102 on the trajectory, at a second time instant, new environment information 212 is determined. The new environment information 212 contains information about the vehicle's surroundings at the second time instant.
[0142] Subsequently, the trajectory is modified based on the information about the vehicle's surroundings at the second time instant using the method for determining a trajectory.
[0143] Then, the vehicle 102 is moved on the modified trajectory at the second time instant.
[0144] In Figure 4 One example for planning a trajectory for the vehicle 102 is shown in Fig. 4. In this example, the trajectory is described for driving through a roundabout 400.
[0145] First, an initial trajectory 402 is determined.
[0146] In this example, the initial trajectory 402 is determined using an artificial neural network 206-1. In the example, first, the initial trajectory enables a particularly human-like driving style of the vehicle 102.
[0147] In this example, the learned human-like driving style results in that the artificial neural network 206-1 has learned that the view on the roundabout 400 gets better the closer the vehicle 102 gets to the roundabout 400. The artificial neural network 206-1, for example, anticipates that the vehicle is expected to possibly drive into the roundabout 400. Therefore, the initial trajectory 402 extends beyond the stop line 404. The artificial neural network 206-1 plans to drive into the roundabout with the initial trajectory 402 in the situation of having an insufficient view on the entire roundabout. In contrast, a rule-based model 206-2 would plan to stop at the stop line 404 in front of the roundabout in the case of insufficient view on the entire roundabout.
[0148] According to the initial trajectory 402, the vehicle 102 is moved into the roundabout 400. In Figure 4 In Fig. 4, the course of the initial trajectory 402 is represented by nodes connected by lines, which lead into the roundabout 400.
[0149] In this example, the initial trajectory 402 is checked.
[0150] When the field of view of the vehicle 102 is still insufficient, the respective node 406 is marked as invalid due to the boundary condition.
[0151] The hybrid A* algorithm thus explores other nodes 402' in the local surroundings that enable parking.
[0152] To ensure that a new trajectory is found that leads to a better driving behavior, it can additionally be provided that the hybrid A* algorithm continues the initial trajectory 402 with nodes 402" that, in the case of an assumed sufficient field of view, would lead to a valid travel trajectory. A valid travel trajectory ends, for example, when the target region 408 is reached.
[0153] Overall, a new trajectory is thus generated in this example, which starts from the initial solution that leads to the desired driving behavior and is continued with braking trajectories. The starting points of the braking trajectories found by the hybrid A* algorithm are optimized here in such a way that the braking trajectories already pre-control the steering foresighted and comparatively preferred travel behavior.
[0154] In addition to the possible new trajectory, a modified trajectory 410 is also determined from the new trajectory.
[0155] According to the modified trajectory 410, the vehicle 102 is to park at the stop line 404 in front of the roundabout 400. In Figure 4 The course of the modified trajectory 410 is represented in the middle by nodes that are connected by lines and end in front of the roundabout 400.
[0156] It can be provided that the trajectories for the respective behavior 214 are calculated in parallel. However, if this is not possible, for example, because a single-core embedded computing unit (English: Embedded Computing Unit, ECU) is used, the trajectories can be determined, for example, on the basis of an analytical solution for the optimal trajectory. Such an analytical solution makes specific assumptions about the state cost function, for example, as a heuristic method. One example of an assumption is that the trajectory leads to as little longitudinal jerk of the vehicle 102 as possible as a comfort criterion. This limits the solution quality in terms of comfort. The solution can be determined in extremely short time for this purpose, i.e. the safety can be significantly strengthened with limited computing resources.
Claims
1. A computer-implemented method for determining a trajectory for a vehicle (102), characterized by - providing an environment model (202) (302), wherein the environment model (202) contains environment information (212) about the environment surrounding the vehicle; - providing at least one behavior (214) (304); - planning a trajectory (306) for the vehicle (102) based on the environment information (212) using an artificial neural network (206-1); or - planning a trajectory (308) for the vehicle (102) based on the behavior (214) using a rule-based model (206-2); - selecting and / or modifying the trajectory (308) based on the environment information (212) and the trajectory and according to the behavior (214) and according to a cost (216) using a rule-based model (210-1).
2. The method according to claim 1, characterized by - predefining a plurality of behaviors (214) (304) having different priorities; - determining and / or modifying one trajectory (306, 308) for each behavior (214).
3. The method according to claim 1 or 2, characterized by - predefining a list (304) of a plurality of behaviors (214) distinguished by a priority order; - determining and / or modifying one trajectory (306, 308) for each behavior (214); - selecting the trajectory assigned to the behavior having the highest priority in the list as the trajectory (104) (310) for the vehicle (102).
4. The method according to claim 2 or 3, characterized in that, - determining the cost (216) (308) from the environment information (212) and / or from the behavior (214).
5. The method according to any of the preceding claims, characterized in that, - modeling a part of the cost (308) using a machine learning model (208-1) trained to assign a respective cost from the environment information (212) and / or from the behavior (214).
6. The method according to any of the preceding claims, characterized in that, - modeling a part of the cost (308) using a rule-based model (208-2) configured to assign a respective cost from the environment information (212) and / or from the behavior (214).
7. The method according to any of the preceding claims, characterized in that, - selecting a behavior (308) from a plurality of pre-defined behaviors according to the cost.
4. A computer program having a program code for performing the method according to one of the preceding claims, when the computer program runs on a computer.
8. The method according to any of the preceding claims, characterized in that, predefining at least one safety goal and at least one goal characterizing a performance of the trajectory, wherein a trajectory is planned for a first time horizon, which trajectory realizes the goal characterizing the performance as well as possible and which trajectory fulfils the at least one safety goal for the first time horizon, wherein a continuation of the trajectory planned for the first time horizon is planned for a second time horizon which is longer compared to the first time horizon, which continuation is allowed to realize the goal characterizing the performance less well compared to the trajectory planned for the first time horizon, wherein a modified trajectory is determined as a trajectory for the vehicle (102) until the end of the second time horizon before the end of the first time horizon using a rule-based model on the basis of the environmental information and the trajectory planned for the second time horizon and in accordance with the behavior and in accordance with a cost, or wherein the trajectory planned for the second time horizon is determined as a modified trajectory for the vehicle (102) until the end of the second time horizon when no modified trajectory has been determined as a trajectory for the vehicle (102) until the end of the second time horizon until the end of the first time horizon.
9. The method according to any one of the preceding claims, characterized in that - the environmental information (212) comprises information (302) about the vehicle's surroundings at a first point in time; - a plurality of trajectories (306, 308) for the vehicle (102) is determined on the basis of the information about the vehicle's surroundings at the first point in time; - the vehicle (102) is moved (312) on the trajectory; - during the movement of the vehicle (102) on the trajectory, environmental information (212) (302) is determined, which comprises information about the vehicle's surroundings at a second point in time; - the trajectory (308) is modified on the basis of the information about the vehicle's surroundings at the second point in time; - and the vehicle (102) is not moved on the trajectory, but on the modified trajectory (312).
10. The method according to any of the preceding claims, characterized in that, - training data comprising a reference trajectory for a trajectory of the vehicle (102) is provided, wherein the reference trajectory simulates a driving behavior of a human or represents a detected driving behavior of a human, wherein the artificial neural network (206-1) is trained for determining a trajectory (300) for the vehicle (102) which coincides as well as possible with the reference trajectory.
11. A device (100), in particular a control device, for determining a trajectory for a vehicle, characterized in that The apparatus (100) is configured to carry out the method according to any one of claims 1 to 10.