Computer-implemented method for determining a movement plan to control an autonomous agent in a traffic situation

Abstract

The invention relates to a computer-implemented method for determining a movement plan (30) for controlling an autonomous agent (10) in a traffic situation (12), wherein the agent (10) comprises a sensor (22) for acquiring sensor data (24) in a sensor field of view (26), wherein the sensor data (24) allow the traffic situation (12) to be recognized.

Description

SPECIALIZATION

[0001] The invention relates to a computer-implemented method for determining a movement plan for controlling an autonomous agent in a traffic situation. The invention further relates to a data processing device, a computer program, a computer-readable data carrier, and an autonomous agent. BACKGROUND

[0002] Motion planning for self-driving cars and autonomous mobile robots in environments with significant obstructions is one of the key challenges in the industry. These obstructions are primarily caused by the physical limitations of the sensor elements and can reduce safety and performance levels during movement. Numerous practical examples exist where obstruction can significantly impact traffic situations. For instance, a pedestrian might be obscured by a building or a parked car near a crosswalk. Another example would be when a car or motorcycle is obscured from the perspective of the self-driving vehicle attempting to overtake a truck in front of it.

[0003] One of the most common ways to address this problem is the development of "occlusion-sensitive" motion planning algorithms. State-of-the-art algorithms in this area use various approaches, such as: • Reachability analysis - prediction of all possible worst-case scenarios, determination of the hidden space forward in time based on the predictions and steering the vehicle to avoid these regions [1]. • Game-theoretic approach – formulates concealment-sensitive planning as a game between the ego-vehicle and other road users who are concealed and potentially conflicting. The optimal solution is defined as a dynamic game problem and subsequently used for motion planning of the ego-vehicle [2]. However, this may not be universally applicable, as it is designed for only one traffic scenario. • Potential field method - represents the environment with a scalar field, e.g. a risk field [3] or an attracting / repelling field [4], and then the motion planning process is carried out based on the gradient of this field. • Information-theoretic approach – introduces an information-theoretic goal related to the degree of obstruction in the environment. During motion planning, this goal is optimized together with other goals (e.g., performance and safety goals), which allows the vehicle to actively acquire new information about obscured regions [5, 6].

[0004] Another important segment of the state of the art relates to 3D occupancy prediction techniques, which are then used as a module of an occlusion-sensitive motion planning algorithm. State-of-the-art 3D occupancy prediction solutions utilize modern AI methods that provide occupancy and semantic predictions in 3D space based on sensor inputs such as cameras [7, 8, 9].

[0005] There are several potential problems associated with the techniques mentioned in the prior art: • Overly conservative handling of coverings - this problem mainly occurs with techniques that consider worst-case scenarios, such as reachability analysis, and can lead to overly conservative vehicle behavior and reduce its performance. • There is no combination of information acquisition capability with risk-sensitive collision avoidance of potentially obscured objects – on the one hand [1-4], information acquisition capability may be disregarded, while risk sensitivity regarding potential obscurations certainly exists. On the other hand [5, 6], risk sensitivity may be lacking, but attempts are made to incorporate information-theoretic goals during motion planning. • Lack of general applicability - some of the sources mentioned are designed for a specific traffic scenario, meaning they may not be applicable to many other traffic scenarios.

[0006] The following sources are referenced herein: [1] R. Firoozi, A. Mir, GS Camps and M. Schwager, “Occlusion-Aware MPC for Guaranteed Safe Robot Navigation with Unseen Dynamic Obstacles.” arXiv, 16 November 2022. Accessed: 16 March 2024. [Online]. Available: http: / / arxiv.org / abs / 2211.09156; [2] Z. Zhang and JF Fisac, “Safe Occlusion-aware Autonomous Driving via Game-Theoretic Active Perception,” in Robotics: Science and Systems XVII, July 2021. doi: 10.15607 / RSS.2021.XVII.066; [3] C. van der Ploeg, T. Nyberg, JMG Sánchez, E. Silvas and N. van de Wouw, “Overcoming the Fear of the Dark: Occlusion-Aware Model-Predictive Planning for Automated Vehicles Using Risk Fields.” arXiv, 27 September 2023. Accessed: 16 March 2024. [Online]. Available: http: / / arxiv.org / abs / 2309.15501; [4] P. Lin, E. Javanmardi, J. Nakazato und M. Tsukada, „Occlusion-Aware Path Planning for Collision Avoidance: Leveraging Potential Field Method with Responsibility-Sensitive Safety.“ arXiv, 12. Juni 2023. Zugriff: 16. März 2024. [Online]. Verfügbar: http: / / arxiv.org / abs / 2306.06993; [5] B. Charrow et al., „Information-Theoretic Planning with Trajectory Optimization for Dense 3D Mapping,“ in Robotics: Science and Systems XI, Robotics: Science and Systems Foundation, Juli 2015. doi: 10.15607 / RSS.2015.XI.003; [6] H. Andersen et al., „Trajectory optimization for autonomous overtaking with visibility maximization“, 20. Internationale Konferenz über intelligente Beförderungssysteme (ITSC - International Conference on Intelligent Transportation Systems) der IEEE 2017, Oktober 2017, S. 1-8. doi: 10.1109 / ITSC.2017.8317853; [7] Y. Huang, W. Zheng, Y. Zhang, J. Zhou und J. Lu, „Tri-Perspective View for Vision-Based 3D Semantic Occupancy Prediction“, Konferenz über Computervision und Mustererkennung (CVPR - Conference on Computer Vision and Pattern Recognition) der IEEE / CVF 2023, Vancouver, BC, Kanada, 2023, S. 9223-9232, doi: 10.1109 / CVPR52729.2023.00890; [8] Li, Z. et al. (2022). „BEVFormer: Learning Bird's-Eye-View Representation from Multicamera Images via Spatiotemporal Transformers.“ In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T. (Herausgeber) Computer Vision - ECCV 2022. ECCV 2022. Lecture Notes in Computer Science, Band 13669. Springer, Cham. https: / / doi.orq / 10.1007 / 978-3-031-20077-9 1; und [9] Tian, ​​Xiaoyu and Jiang, Tao and Yun, Longfei and Mao, Yucheng and Yang, Huitong and Wang, Yue and Wang, Yilun and Zhao, Hang, "Occ3D: A Large-Scale 3D Occupancy Prediction Benchmark for Autonomous Driving", Advances in Neural Information Processing Systems, pp. 64318v64330, Curran Associates, Inc, Vol. 36, 2023

[0007] The document:

[10] Taş, Ö. Ş.: Brusius, PH; Stiller, C.: Decision-theoretic MPC: motion planning with weighted maneuver preferences under uncertainty. Preprint. [v1] Fri, 27 Oct 2023 08:23:23 UTC. 2023-10-27. pp. 1-15; https: / / doi.orq / 10.48550 / arXiv.2310.17963 This method reveals a procedure that incorporates multiple maneuver preferences into motion planning. The procedure optimizes the trajectory by considering weighted maneuver preferences along with uncertainties, while simultaneously ensuring the feasibility of a limited-chance evasive option.

[0008] The document:

[11] Taş, Ö. Ş.; Hauser, F.; Stiller, C.: Decision-time postponing motion planning for combinatorial uncertain maneuvering. In: 2018 21 st International Conference on Intelligent Transportation Systems (ITSC), Maui, Hawaii, USA, November 4-7, 2018. Piscataway, NJ, USA : IEEE, 2018. pp. 2419-2425. ISBN 978-1-7281-0323-5; https: / / doi.org / 10.1109 / ITSC.2018.8569580 reveals a method for motion planning that takes prediction uncertainties into account and, in the case of high uncertainty, postpones combinatorial decision-making to a later time within a planning horizon.

[0009] The document:

[12] Taş, Ö. Ş.; Stiller, C.: Tackling existence probabilities of objects with motion planning for automated urban driving. Preprint. [v2] Wed, 21 Oct 2020 17:00:50 UTC. 2020-10-21. 5 p.; https: / / doi.org / 10.48550 / arXiv.2002.01254 reveals a method for motion planning that combines alternative maneuvers and plans a motion based on the probabilities of these alternatives.

[0010] The document:

[13] Bouzidi, M.-K. et al.: Motion planning under uncertainty: integrating learning-based multi-modal predictors into branch model predictive control. Preprint. [v1] Mon, 6 May 2024 13:45:44 UTC. 2024-05-06. 7 p.; https: / / doi.org / 10.48550 / arXiv.2405.03470 This document discloses a method that uses branch MPC (model predictive control) to consider predictions. The method includes an online scenario selection process based on topology and collision risk criteria. Additionally, an adaptive decision deferral strategy is disclosed, which delays a planner's decision on a single scenario until the uncertainty is resolved.

[0011] Furthermore, DE 10 2020 208 637 A1, US 11 126 180 B1, US 2020 / 0 225 669 A1 and US 2024 / 0 124 019 A1 are state of the art. BRIEF SUMMARY OF THE INVENTION

[0012] The object of the invention is to provide an improved method for determining a movement plan for controlling an autonomous agent in a traffic situation.

[0013] To solve this problem, the invention provides a computer-implemented method according to claim 1. A data processing device, a computer program, a computer-readable data carrier, and an autonomous agent are the subject of the parallel claims. Advantageous embodiments of the invention are the subject of the dependent claims.

[0014] In one aspect, the invention provides a computer-implemented method for determining a movement plan for controlling an autonomous agent in a traffic situation, wherein the agent comprises a sensor for capturing sensor data in a sensor field of view, wherein the sensor data allows the recognition of the traffic situation, and wherein the method comprises the following: a) Evaluating the traffic situation at a current time t0 by determining, based on current sensor data, a current obscuration map, wherein the obscuration map reveals one or more obscured areas in the sensor field of view where no corresponding sensor data is currently available; b) Creating, based on the current occlusion map, a scenario tree, wherein the scenario tree comprises a plurality of scenarios for the traffic situation in a planning horizon, and determining a suitable movement plan for the agent to approach for each scenario, wherein the scenario tree is based on a decision shift that extends to a decision shift time T p all suitable movement plans are limited to a common movement plan, with the decision delay time T p calculated based on a covert map predicted for each scenario, revealing a point in time at which all but one scenario can be eliminated, and where the appropriate movement plans are determined based on a cost function incorporating the predicted covert maps for each scenario; and c) Reassessing the traffic situation at a future time t1 > t0 until only one scenario remains, and selecting the appropriate movement plan to guide the agent in the traffic situation via the decision delay time T p out.

[0015] One advantage of this method is that occlusions are taken into account during motion planning. The agent can be an autonomous vehicle, a robot, a drone, or similar. The occlusions can be accounted for by calculating a decision delay time T. p Based on predicted occlusions for each scenario, the agent follows the strategy until the decision delay time T. pThat is, until all relevant occlusions are resolved, a common movement plan is initially adopted. This can increase the degree of certainty during autonomous movement. Decision delay can be viewed as passive information gathering. Furthermore, the movement plans, including the common movement plan, can be determined based on a cost function that includes the predicted occlusion maps. Therefore, the agent can attempt to resolve the occlusions by following the common movement plan. The common movement plan can thus be determined in a way that optimally resolves relevant occlusions. This can be viewed as active information gathering.This can mimic human behavior: in a traffic situation where the agent is following another road user and attempts to overtake, the agent can, for example, first explore the area in front of the other road user by swerving towards the center of the road, but not yet initiating an overtaking maneuver. This can make motion planning more effective.

[0016] Step b) also preferably includes the following: b1) Determining the appropriate movement plans by minimizing the cost function, where the cost function includes an expression to maximize an information gain regarding one or more currently hidden areas.

[0017] Step c) also preferably includes the following: c1) Updating the predicted occlusion maps based on the sensor data at the future time t1 > t0 and adjusting the decision delay time T pbased on the updated predicted occlusion maps; and / or c2) Excluding one or more scenarios from the scenario tree based on the sensor data at the future time t1 > t0.

[0018] One advantage of the procedure can be that the decision postponement time T p The move plan can be adjusted while following the shared movement plan. Therefore, if occlusions are resolved earlier than expected, the agent can immediately decide on a movement plan. This can make movement planning more effective.

[0019] Step b) also preferably includes the following: b2) Creating the scenario tree by including one or more worst-case scenarios, where each worst-case scenario assumes one or more obstacles in at least one of the currently obscured areas; and / or b3) Creating the scenario tree by including one or more best-case scenarios, where each best-case scenario assumes no obstacles in any of the currently obscured areas.

[0020] One advantage of this method is that it considers both worst-case and best-case scenarios. Movement planning is therefore potentially risk-aware, but not overly conservative.

[0021] Step b) also preferably includes the following: b4) Estimating a possible maximum speed and / or acceleration of the assumed obstacle in order to calculate a reachable area of ​​the assumed obstacle in at least one of the currently obscured areas.

[0022] One advantage of this method is that it can consider a possible maximum speed and / or acceleration in the worst-case scenario. The motion planning is therefore potentially risk-aware, but not overly conservative.

[0023] Step b) also preferably includes the following: b5) Predicting the occlusion map for each scenario by mapping a position and / or orientation of the agent for each movement plan onto the one or more currently occluded areas.

[0024] Step a) also preferably includes the following: a1) Determine, based on the current sensor data, an up-to-date occupancy map that identifies one or more areas in the sensor's field of view that are occupied; and a2) Determine, based on the current occupancy map, the current cover map.

[0025] Step a2) further preferably includes the following: a2a) Determining the current occlusion map by extending a straight line from the sensor beyond the one or more occupied areas to a maximum sensor detection distance of the sensor.

[0026] Step a) also preferably includes the following: a3) Filtering and / or sorting the current obscuration map by classifying the one or more currently obscured areas according to a relevance score to assess the traffic situation.

[0027] One advantage of this method is that occlusions can be filtered according to relevance. For example, occlusions located far away from the agent or the road may be less relevant. This can make motion planning more effective.

[0028] Step a3) further preferably includes the following: a3a) Determining the relevance score based on a navigation map, a distance from the agent to one or more currently obscured areas, a reachable area of ​​one or more assumed obstacles and / or a characteristic of one or more obscured areas in the sensor's field of view.

[0029] One advantage of this method is that it may take into account the nature of the occupied area. For example, if an area is not occupied by a vehicle, pedestrian, or bicycle, it may be less relevant than another occupied area. This can make movement planning more effective.

[0030] Preferably the method further comprises the following: d) Generating a control signal to control the agent based on the common movement plan and / or the selected movement plan.

[0031] In another aspect, the invention provides a data processing device comprising means for carrying out the method according to one of the preceding embodiments.

[0032] Any feature, aspect and / or advantage described herein in relation to embodiments of the data processing device may optionally apply to embodiments of the method and vice versa.

[0033] In another aspect, the invention provides a computer program that includes instructions which, when the program is executed by a computer, cause the computer to perform the method according to one of the preceding embodiments.

[0034] Any feature, aspect and / or advantage described herein in relation to embodiments of the computer program may optionally apply to embodiments of the method and / or the data processing device, and vice versa.

[0035] In another aspect, the invention provides a computer-readable data carrier on which the computer program is stored.

[0036] Each feature, aspect and / or advantage described herein in relation to embodiments of the computer-readable data carrier may optionally apply to embodiments of the method, data processing device and / or computer program, and vice versa.

[0037] In another aspect, the invention provides an autonomous agent comprising a sensor for acquiring sensor data in a sensor field of view, wherein the sensor data enables the detection of a traffic situation, and wherein the agent further comprises a data processing device according to one of the preceding embodiments.

[0038] Each feature, aspect and / or advantage described herein in relation to embodiments of the autonomous agent may optionally apply to embodiments of the method, data processing device, computer program and / or computer-readable medium, and vice versa.

[0039] Preferred embodiments of the invention can be summarized as follows: Based on a retrieved map of concealed space, potential concealed objects, their relevance, and their future trajectories are assessed. Using the information about the concealed space and potentially concealed objects, a planning algorithm selects the optimal next action for the ego vehicle to minimize the risk of potentially concealed objects and maximize information gain, i.e., to minimize the relevant concealed space in the future. One embodiment of the solution is described in Fig. 3 illustrates and includes the following steps: 1. Creating an occupancy grid map: We use an existing approach, such as [7], [8], or [9], to obtain a 3D occupancy grid map (3D-OGM) or a 2D occupancy grid map (2D-OGM, so-called "bird's-eye view") from the current sensor measurements and, optionally, the measurement history. At a given resolution of the 3D or 2D space, the 3D-OGM and 2D-OGM contain information about which cell of the space is occupied. Additional information, such as from segmentation and classification, can be used to annotate the occupied cells with a class (e.g., vehicle, pedestrian, road, sidewalk, etc.). 2. Adding occlusion information: Based on the sensor configuration of the ego vehicle, we define a corresponding sensor model for each sensor (e.g., lidar, camera, radar) to obtain occlusion information as described in [3]. Based on each sensor model, we obtain a field of view (FoV) with a resolution (discretization), a sensor detection distance, and a sensor center. For each cell of the discretized FoV, a 3D beam, whose length is equal to the sensor detection distance emanating from the sensor center, is aligned to the cell (e.g., the center of the camera coordinate system, the camera focal length, and each pixel of the camera image define how the beam must be aligned for a pinhole camera model). The overlapping cells of the OGM can be retrieved together with the 3D OGM or, after projecting the beam onto the 2D plane, with the 2D OGM for each beam.

[0040] Starting from the sensor center, each cell is marked as visible until either the first occupied cell is reached or the end of the beam is reached. After evaluating each sensor as above, the result of this step is a 3D OGM or a 2D OGM including occlusion information, where any unoccupied cell not marked as visible is occluded.

[0041] 3. Assessing the situation (scenario, relevance of the hidden area): Based on the 3D OGM or the 2D OGM, including occlusion information and infrastructure information from an SD card (providing information about lanes, intersections, etc.), the obscured area is filtered by relevance. For this step, all obscured cells are filtered by relevance by projecting each obscured cell onto the SD card, retaining only those that project onto areas (such as roads, sidewalks, intersections, etc.) relevant to the planning phase. Furthermore, the previously planned optimal trajectory for the ego vehicle filters the remaining obscured cells using a distance metric; that is, obscured cells that are too far away are discarded.Additionally, information such as annotations in the 3D OGM or 2D OGM, or the classification or segmentation of sensor data, can be used to detect traffic situations and areas (scenarios) that pose a higher risk. To account for this, a relevance score is added to each obscured cell, with each obscured cell assigned to higher-risk scenarios having a higher relevance score. For example, the relevance score is increased for obscured cells in or near areas with high detected activity (e.g., where many pedestrians are detected on the sidewalk) or areas obscured by objects such as double-parked cars blocking the opposite lane.

[0042] 4. Calculating a forward-reachable set of hidden worst-case obstacles: The generated 3D map is then used to calculate the possible future positions of the potentially obscured objects. For this, we can either use the worst-case scenario by calculating the forward reachable set (simplified physics using the point mass model), where we constrain the possible maximum speed and acceleration of the obscured object by the most probable object type extracted from the situation assessment (pedestrian or vehicle). This set can be added as a crisis scenario to the scenario tree of the branch's model predictive control (MPC). The branch's MPC (see section 7) can then plan using all scenarios added to the scenario tree, i.e., for example, a crisis scenario and a scenario that disregards the potentially obscured object.

[0043] 5. Adaptive decision delay: To calculate the appropriate decision delay time, we use the predicted occlusion state (see section 7 for details) calculated by the branching MPC in the last time step. We therefore check each time step in the planning horizon where sufficient information is available (i.e., the occlusion is low enough) to determine whether an object is present or not. This means that this block outputs the time step at which the autonomous vehicle is expected to be able to make an informed decision.

[0044] 6. Information acquisition cost function: The cost function can be defined as in [5], [6]. We include the objective of minimizing the hidden space alongside the cost function expressions in order to maximize progress on the reference path, minimize deviation from the path, maximize distance to obstacles, and minimize lateral and longitudinal jerk and acceleration for comfortable driving. If we have more than one relevant hidden space, we weight the concealments based on the relevance extracted from point 3. To minimize the hidden space, we add an expression to the cost function to maximize the information gain between two time steps (the difference between the concealment of two time steps).

[0045] 7. Model predictive control of branching: The branching MPC is configured using the decision delay time from step 5 and the information acquisition cost function from step 6. The model used for the branching MPC can simply be the kinematic bicycle model, but extended to include the occlusion state. The occlusion state is computed using the occlusion data acquired in step 2, along with the relevant occlusions identified in step 3. We create a function that maps the current position and orientation of the autonomous system to the occluded areas of each selected occlusion. This essentially involves formulating a geometric equation that computes the field of view (FoV) and returns the occluded space relative to the autonomous system's input. This occlusion state is also used to formulate the information acquisition cost function.

[0046] The branching MPC plans with all scenarios added to the scenario tree in point 4, thereby minimizing the information gathering cost function while simultaneously satisfying the constraints. This is achieved by allowing different strategies (i.e., control inputs) in the branching MPC, which are restricted in such a way that they are identical in the initial time steps, i.e., until the calculated time of decision deferral is reached, i.e., until it is clear which strategy will be pursued.

[0047] Embodiments of the invention preferably have the following advantages and effects: - Reducing the conservatism of the autonomous system by incorporating risk through the following: • Information gathering by including the collision risk with hidden objects; • By configuring the decision delay based on future occlusion states, the planner recognizes when enough information is available to decide which action to take; • The permission for the planner to plan for different strategies that the autonomous vehicle can pursue depending on which scenario occurs. - Improving human-like behavior of the autonomous system, which adapts automatically (without being hard-coded), such as: • Initiating information gain (maximizing the field of view); • Waiting for further information in the future by slowing down the approach to conflict zones (e.g., concealed intersections). - This approach can be universally applied to various traffic scenarios involving obstructions. It is not designed for just one specific traffic scenario (not just for intersections or obstructed views due to a large truck, etc.), meaning it does not need to be reconfigured for different traffic scenarios, as in [2, 5]. - The architecture can be easily extended to handle other uncertain scenarios in motion planning (not just for the hidden object). For example, additional branches can be added to handle multimodal predictions regarding the behavior of the detected road users. - Embodiments of the invention can also be applied to motion planning in robotics, drones, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Embodiments of the invention are then explained in more detail with reference to the accompanying drawings. These show: Fig. 1 an autonomous agent in an exemplary traffic situation; Fig. 2. A first embodiment of a computer-implemented method for determining a movement plan for controlling the agent in the traffic situation; and Fig. 3 a second embodiment of the computer-implemented method. DETAILED DESCRIPTION OF THE DRAWINGS

[0049] Fig. Figure 1 shows an autonomous agent 10 in an exemplary traffic situation 12 at a time t0.

[0050] The exemplary traffic situation 12 comprises three road users 14: a first road user 14a in the form of a moving pedestrian, a second road user 14b in the form of a moving vehicle, and a third road user 14c in the form of another moving vehicle. The agent 10 travels in an agent travel direction 16. The third road user 14c travels in a direction 18, 18c that is opposite to the agent travel direction 16. The third road user 14c can be considered an obstacle 44.

[0051] The exemplary traffic situation 12 further includes a static object 20 as an obstacle 44, which is located in the agent's direction of travel 16 in front of the agent 10. The static object 20 is located between the agent 10 and the third road user 14c.

[0052] For autonomous movement, control, or steering, the agent 10 includes a sensor 22 for acquiring sensor data 24 within a sensor field of view 26. The sensor 22 can be, for example, a radar device, a lidar device, and / or a camera device. The sensor data 24 can include data on the positions, distances, speeds, accelerations, trajectories, and / or characteristics of objects 20 and / or road users 14 within the sensor field of view 25. To identify the characteristics of objects 20 and / or road users 14, a machine learning algorithm, such as a neural network, trained for classifying and / or segmenting traffic scenarios, can be used.

[0053] The static object 20 obscures an area 28 in the sensor's field of view 26, preventing the sensor 22 from acquiring any corresponding sensor data 24. The third road user 14c is located within the obscured area 28 and outside the sensor's field of view 26. In other words, due to the static object 20, the agent 10 cannot detect the third road user 14c using the sensor 22, and currently no sensor data 24 regarding the third road user 14c is available.

[0054] One concept of the embodiments according to the invention is the determination of a movement plan 30 for controlling the agent 10 in the traffic situation 12, wherein the movement plan 30 is based on the hidden area 28.

[0055] Fig. Figure 2 shows a first embodiment of a computer-implemented method for determining the movement plan 30 for controlling the agent 10 in traffic situation 12.

[0056] In step S11, the first embodiment comprises the following: - Evaluating the traffic situation 12 at the current time t0 by determining, based on the current sensor data 24, a current occlusion map 32, wherein the current occlusion map 32 reveals one or more occluded areas 28 in the sensor field of view 2, in which no corresponding sensor data 24 are currently available.

[0057] In step S12, the first embodiment comprises the following: - Create, based on the current occlusion map 32, a scenario tree 32, wherein the scenario tree 34 comprises a plurality of scenarios 36 for the traffic situation 12 in a planning horizon, and determine, for each scenario 36, a suitable movement plan 30 to approach the agent 10, wherein the scenario tree 34 is based on a decision shift that includes all suitable movement plans 30 up to a decision shift time T p limited to a common movement plan, whereby the decision delay time T p is calculated based on an occlusion map 32 that was predicted for each scenario 34 and reveals a point in time at which all scenarios 36 except one can be excluded.

[0058] In step S13, the first embodiment comprises the following: - Reassessing the traffic situation 12 at a future time t1 > t0, until a single scenario 36 remains, and selecting the corresponding movement plan 30 to approach the agent 10 in the traffic situation 12 via the decision delay time T p out.

[0059] It will be revisited Fig. 1. Referenced: In step S11, the method can determine all areas 28 within the sensor's field of view 26 that are obscured and for which no corresponding sensor data is available at time t0. The method can first determine a current occupancy map 38, which identifies an area 40 in the sensor's field of view 26 that is obscured. Subsequently, the current occlusion map 32 can be determined by extending a straight line from sensor 22 beyond the obscured area 40 to a maximum sensor detection distance 42 of sensor 22. The area 28 beyond the obscured area 40 can be marked as obscured. The occlusion map 32 can thus be determined based on the occupancy map 38.

[0060] In step S11, the procedure can further filter and / or sort the obscured areas 28 according to a relevance score. For example, the obscured area 28 in the agent's direction of travel 16 may be more relevant than a obscured area 28 in another direction. The procedure can also, for example, compare the obscured area 28 with a navigation map 58, which can indicate whether the obscured area 28 is located on a road or not. The relevance score can also be based on a distance from the agent 10 to the obscured area 28 and / or on the characteristics of the obscured area 30 in the sensor's field of view 26.

[0061] In step S12, the procedure creates a scenario tree 34 based on the current occlusion map 32, which includes one or more scenarios 36. In this step, the procedure can include one or more worst-case scenarios 48, each worst-case scenario assuming one or more obstacles 44 in the occluded area 28. For example, one worst-case scenario 48 might assume the third road user 14c as an obstacle 44 in the occluded area 28. Another worst-case scenario 48 might assume a non-existent obstacle 46 in the occluded area 28. Another worst-case scenario 48 might assume both the third road user 14c and the non-existent obstacle 46 in the occluded area 28.

[0062] In step S12, the procedure can further estimate a possible maximum speed and / or acceleration and / or a trajectory for each assumed obstacle 44 in order to calculate a reachable area of ​​the assumed obstacle 44 within the obscured area 28. For example, if the procedure assumes the third road user 14c to be a vehicle, it can assume a possible maximum speed of 70 km / h. If the procedure assumes the non-existent obstacle 46 to be a pedestrian, it can assume a possible maximum speed of 10 km / h. The reachable area of ​​the assumed obstacle 44 can also be taken into account in the relevance score.

[0063] In step S12, the procedure can further include one or more best-case scenarios 50, where each best-case scenario 50 assumes no obstacle 44 in the hidden area 28. In step S12, the procedure further determines a suitable movement plan 30 for each scenario 36. In this step, the scenario tree 30 is based on a decision deferral that considers all suitable movement plans 30 up to the decision deferral time T. p limited to a common movement plan. In other words, the suitable movement plans 30 correspond to each other until the decision postponement time T. p .

[0064] The decision postponement time T pThe calculation is based on a predicted obscuration map 52 for each scenario 36. In one scenario 36, the procedure might, for example, assume the third road user 14c to be an obstacle 44 in the obscured area 28, that is, a vehicle traveling at, for example, 25 km / h. The procedure can thus predict a time when the third road user 14c emerges from the obscured area 28. When the third road user 14c emerges from the obscured area 28, the best-case scenario 50, which assumes no obstacle 44 in the obscured area 28, can be eliminated.

[0065] For each movement plan 30, the predicted covert map 52 can further be based on a mapping of the agent 10's position and / or orientation to the currently covered area 28. In other words, the movement of agent 10 according to each movement plan 30 can be used in calculating the decision delay time T. p be taken into account.

[0066] In step S12, the suitable movement plans 30 can be determined by minimizing based on a cost function 54 that includes the predicted covert map 52 for each scenario 36. The cost function 54 can, for example, be a sum of single-scenario cost functions, each single-scenario cost function associated with a single scenario 36. The cost function 54 can include an expression for maximizing information gain regarding the currently covered area 28.

[0067] In step S13, the procedure includes reassessing the traffic situation 12 at the future time t1 > t0. In step S13, the procedure can update the obscuration maps 52, which were predicted at time t0, based on the sensor data 24 at the future time t1 > t0. The updated predicted obscuration maps 52 can then be used to adjust the decision delay time T. p can be used. In step S13, one or more scenarios 36 can also be excluded from the scenario tree 34 based on the sensor data 24 at the future time t1 > t0.

[0068] Through reassessment at the future time t1 > t0 or through any further reassessment at times t2 > t1, t3 > t2, etc., more and more scenarios 36 can be excluded from the scenario tree 34 until only one scenario 36 remains. Finally, the corresponding movement plan 30 for this remaining scenario 36 to drive agent 10 in traffic situation 12 is determined via the (adjusted) decision delay time T. p selected.

[0069] Fig. Figure 3 shows a second embodiment of the computer-implemented method.

[0070] In step S21, the second embodiment comprises the following: - Creating the current occupancy map 38 based on the current sensor data 24.

[0071] In step S21, the occupancy map 38 is preferably three-dimensional, although the invention is not limited to this. In the occupancy map 38, the sensor field of view 26 can be divided into a plurality of cells. The occupancy map 38 can then contain information for each cell as to whether the cell is occupied or not. Additionally, the occupancy map 38 can contain an input 56 regarding the nature of the occupancy, i.e., how the cell is occupied, e.g., whether the cell is occupied by a building, a vehicle, a pedestrian, a road, and / or a sidewalk. The input 56 can be determined based on a machine learning algorithm, such as a neural network, that is trained for annotation, e.g., for classification and / or segmentation, of traffic situation scenes.

[0072] The occupancy map 39 is constructed based on the current sensor data 24 at time t0. Optionally, the occupancy map 38 can also be constructed based on sensor data 24 at one or more previous times t < t0.

[0073] In step S22, the second embodiment comprises the following: - Determining the current cover map 32 based on the current occupancy map 38.

[0074] In step S22, a sensor model 59 encompassing the maximum sensor detection range 42 can be constructed. For each cell of the occupancy map 38, a three-dimensional beam with a length equal to the maximum sensor detection range 42 can be aligned, originating from the center of the sensor 22. Starting from the center of the sensor 22, each cell can be marked as visible until either the first occupied cell is reached or the end of the beam is reached. After evaluating each sensor 22 as above, the result of this step S22 can be a three-dimensional occlusion map 32, which includes occlusion information about whether any unoccupied cell not marked as visible is occluded. Alternatively, the occlusion map 32 can be two-dimensional.

[0075] In step S23, the second embodiment comprises the following: - Filter and / or sort the current occlusion map 32 by relevance.

[0076] In step S23, the current occlusion map 32 can be filtered and / or sorted by assigning a relevance score to one or more currently obscured areas 28. The relevance score can be based on the navigation map 58. The navigation map 58 can, for example, be a standard definition (SD) map. All obscured areas 28 can be filtered by relevance by projecting each obscured cell onto the SD map and retaining the cell only if it is projected onto areas (such as a road, a sidewalk, intersections, etc.) that may be relevant to the planning phase. Additionally, information such as the input 56 for annotation (segmentation and / or classification) in the occupancy map 38 and / or the occlusion map 32 can be used to detect scenarios 36 and obscured areas 28 that pose a higher risk.The output of step S23 can be a rated occlusion map 60, which includes one or more occluded areas 28 that are filtered and / or sorted by relevance.

[0077] In step S24, the second embodiment comprises the following: - Including one or more worst-case scenarios 48 in the scenario tree 34.

[0078] In step S24, the evaluated obscuration map 60 can be used to calculate possible future positions of the assumed or potential obscured obstacles 44 or objects 20. The word-case scenarios 48 can be included by calculating a forward-reachable set where a possible maximum speed and / or acceleration of the assumed or potential obscured obstacle 44 is / are limited by the most probable object type (nature) extracted from the situation assessment (pedestrian or vehicle). This set can be added to the scenario tree 34 as a crisis scenario.

[0079] In step S25, the second embodiment comprises the following: - Including one or more best-case scenarios 48 in the scenario tree 34.

[0080] In step S26, the second embodiment comprises the following: - Calculating the decision delay time T p .

[0081] To calculate the appropriate decision postponement time T p In step S26, the predicted occlusion map 52 can be used for each scenario 36. The predicted occlusion maps 52 can be determined by an output from a prediction module 62 from one or more preceding time points t < t0.

[0082] The second embodiment further comprises steps S27, S28, and S29, which are performed by the prediction module 62 at time t0. The evaluated occlusion map 60, the one or more worst-case scenarios 48, the one or more best-case scenarios 50, and the decision delay time T p are entered into the prediction module 62.

[0083] In step S27, the procedure can determine an occlusion state 64 by extending a kinematic bicycle model with the evaluated occlusion map 60. In this step, a function can be derived that maps the current position and / or orientation of the agent 10 to the occluded areas 28. This can essentially involve formulating a geometric equation that can calculate the sensor field of view 26 and derives the occluded space relative to a position and / or orientation of the agent 10. For this purpose, the kinematic bicycle model can use the sensor model 59.

[0084] In step S28, the concealment state 64 can be used to formulate the cost function 54 for information gathering. The forecasting module 62 can plan with all scenarios 36, 48, 50 that were added to the scenario tree 34, for example, in steps S24 and S25. The forecasting module 62 can minimize the cost function 54 for information gathering by restricting it to the decision postponement time T. p and fulfills the common movement plan. This can be achieved by the prediction module allowing 62 different strategies (i.e., control inputs), which are restricted in that they are used in the initial time steps, i.e., until the calculated time of the decision shift T. p Once reached, they are identical.

[0085] In step S29, the prediction model 62 iterates to predict the occlusion maps 52 for each scenario 36. The output of step S29 can be the appropriate motion plan 30 for each scenario 36, which includes a current optimal trajectory 66 according to the common motion plan. Furthermore, the predicted occlusion map 52 for each scenario 36 from the previous time t < t0 can be updated based on the current sensor data 24, and the decision delay time T p from the preceding time t < t0 can be adjusted based on the updated predicted occlusion maps 52.

[0086] The current optimal trajectory 66 can be fed back to step S2 as a previously planned optimal trajectory for agent 10 in order to filter the remaining hidden cells using a distance metric; that is, hidden cells that are too far away can be discarded. The adjusted decision delay time T p can be reported back at step S26.

[0087] The procedure can then be carried out at the future time t1 > t0 until the (adjusted) decision postponement time T. p can be repeated, and a decision regarding a suitable movement plan 30 to target agent 10 can be made about the (adjusted) decision delay time T. p beyond being hit, i.e., a single scenario 36 remains.

[0088] Embodiments of the invention may further include the generation of a control signal for controlling the agent 10 based on the selected motion plan 30 (including the common motion plan). Embodiments of the invention further include a data processing device or a control device (not shown) that includes or is designed to carry out embodiments of the described methods. Embodiments of the invention may be implemented in a computer program (not shown) that includes instructions which, when the program is executed by a computer or a control device, cause the computer to carry out the embodiments of the described methods. The computer program may be stored on a computer-readable data carrier (not shown).Embodiments of the invention further include the agent 10, which comprises the data processing device or the control device.

[0089] In summary, the embodiments of the invention relate to a concealment-sensitive, scenario-based motion planning system that couples risk sensitivity and information acquisition. REFERENCE MARK 10 autonomous agents 12 Traffic situation 14 road users 14a first road user 14b second road user 14c third road user 16 Agent route direction 18 Direction of travel of the road user 18a Direction of travel of the first road user 18b Direction of travel of the second road user 18c Direction of travel of the third road user 20 static object 22 Sensor 24 sensor data 26 Sensor field of view 28 hidden area 30 Movement Plan 32 Coverage map 34 Scenario Tree 36 Scenario 38 Occupancy map 40 occupied area 42 maximum sensor detection range 44th obstacle 46 Non-existent obstacle (pedestrian) 47 Direction of travel of the non-existent obstacle 48 Worst-case scenario 50 Best-Case Scenarios 52 predicted occlusion map 54 Cost function 56 entries 58 Navigation map 59 Sensor model 60 rated coverts map 62 Prediction module 64 Concealment state 66 optimal trajectory t0 current time t1 future point in time T p Decision postponement time

Claims

[1] Computer-implemented method for determining a movement plan (30) for controlling an autonomous agent (10) in a traffic situation (12), wherein the agent (10) comprises a sensor (22) for acquiring sensor data (24) in a sensor field of view (26), wherein the sensor data (24) allow the traffic situation (12) to be recognized, and wherein the method comprises: a) Evaluating the traffic situation (12) at a current time t0 by determining, based on current sensor data (24), a current obscuration map (32), wherein the current obscuration map (32) reveals one or more obscured areas (28) in the sensor field of view (26) for which no corresponding sensor data (24) are currently available; b) Based on the current occlusion map (32), create a scenario tree (34), wherein the scenario tree comprises a plurality of scenarios (36) for the traffic situation (12) in a planning horizon, and determine, for each scenario (36), a suitable movement plan (30) to approach the agent (10), wherein the scenario tree (34) is based on a decision shift that includes all suitable movement plans (30) up to a decision shift time T p limited to a common movement plan, whereby the decision delay time T pbased on a covert map (52) predicted for each scenario (36) and which identifies a time at which all but one scenario (36) can be eliminated, the appropriate movement plans (30) are determined based on a cost function (54) that includes the predicted covert maps (52) for each scenario (36); and c) Re-evaluating the traffic situation (12) at a future time t1 > t0 until a single scenario (36) remains, and selecting the corresponding movement plan (30) to approach the agent (10) in the traffic situation (12) via the decision delay time T p out. [2] Method according to claim 1, characterized by , that step b) further includes the following: b1) Determining the appropriate movement plans (30) by minimizing the cost function (54), wherein the cost function (54) includes an expression to maximize an information gain regarding one or more currently hidden areas (28). [3] Method according to any of the preceding claims, characterized by , that step c) further includes the following: c1) Updating the predicted occlusion maps (52) based on the sensor data (24) to the future time t1 > t0 and adjusting the decision delay time T p based on the updated predicted occlusion maps (52); and / or c2) Excluding one or more scenarios (36) from the scenario tree (34) based on the sensor data (24) at the future time t1 > t0. [4] Method according to any of the preceding claims, characterized by , that step b) further includes the following: b2) Creating the scenario tree (34) by including one or more worst-case scenarios (48), each worst-case scenario (48) being one or assumes several obstacles (44) in at least one of the currently obscured areas (28); and / or b3) Creating the scenario tree (34) by including one or more best-case scenarios (50), where each best-case scenario (50) assumes no obstacle (44) in any of the currently obscured areas (28). [5] Method according to any of the preceding claims, characterized by , that step b) further includes the following: b4) Estimating a possible maximum speed and / or acceleration of the assumed obstacle (44) to calculate a reachable area of ​​the assumed obstacle (44) in at least one of the currently obscured areas (28). [6] Method according to any of the preceding claims, characterized by , that step b) further includes the following: b5) Predictions of the occlusion map (52) for each scenario (36) by mapping, for each movement plan (30), a position and / or orientation of the agent (10) on the one or more currently occluded areas (28). [7] Method according to any of the preceding claims, characterized by , that step a) further includes the following: a1) Determine, based on the current sensor data (24), a current occupancy map (38) that identifies one or more areas (40) in the sensor field of view (26) that are occupied; and a2) Determine, based on the current occupancy map (38), the current cover map (32). [8] Method according to claim 7, characterized by , that step a2) further includes the following: a2a) Determining the current occlusion map (32) by extending a straight line from the sensor (22) beyond the one or more occupied areas (40) to a maximum sensor detection distance (42) of the sensor (22). [9] Method according to any of the preceding claims, characterized by , that step a) further includes the following: a3) Filtering and / or sorting the current obscuration map (32) by classifying the one or more currently obscured areas (28) according to a relevance score to assess the traffic situation (12). [10] Method according to claim 9, characterized by , that step a3) further includes the following: a3a) Determining the relevance score based on a navigation map (59), a distance from the agent (10) to the one or more currently obscured areas (28), a reachable area of ​​one or more assumed obstacles (44) and / or a nature of one or more occupied areas (40) in the sensor field of view (26). [11] A method according to any of the preceding claims, further comprising: d) Generating a control signal to control the agent (10) based on the common movement plan and / or the selected movement plan (30). [12] Data processing device comprising means for carrying out the method according to any of the preceding claims. [13] Computer program comprising instructions which, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 11. [14] Computer-readable data carrier on which the computer program according to claim 13 is stored. [15] Autonomous agent (10) comprising a sensor (22) for acquiring sensor data (24) in a sensor field of view (26), wherein the sensor data (24) enable the detection of a traffic situation (12), wherein the agent (10) further comprises a data processing device according to claim 12.

Images