Computer-implemented method and system for behavioral planning of participants of traffic scene

By quantifying and evaluating the cognitive uncertainty of traffic participants' behavior in traffic scenario prediction and choosing to consider prediction steps based on decision-making criteria, the problem of unreliable prediction results in the prior art is solved, which significantly improves the safety of traffic scenario planning and computational efficiency.

CN118898901APending Publication Date: 2024-11-05ROBERT BOSCH GMBH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410538748.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2023-05-03
Filing Date
2024-04-30
Publication Date
2024-11-05

AI Technical Summary

Technical Problem

Existing traffic scenario prediction methods are difficult to effectively consider the cognitive uncertainty of traffic participants' behavior, resulting in unreliable and error-prone prediction results, especially due to inadequate performance in perception and limitations of training data.

Method used

By determining the cognitive uncertainty of the predicted behavior in each prediction step and using decision criteria to evaluate whether the behavior of another participant in the prediction step is considered in behavior planning, its prediction is only considered when the decision criteria are met, Otherwise, it will not be considered, thereby improving the safety of maneuvering operation planning.

Benefits of technology

The safety and reliability of maneuvering action planning is significantly improved, by quantifying cognitive uncertainty and setting decision standards, ensuring behavioral planning is carried out only in reliable prediction steps, saving computing power and avoiding unnecessary prediction steps.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118898901A_ABST
    Figure CN118898901A_ABST
Patent Text Reader

Abstract

The invention relates to a computer-implemented method for behavior planning for a first participant of a traffic scene, in which a predicted behavior of at least one further participant of the traffic scene is taken into account, in particular as a function of a cognitive uncertainty of the predicted behavior. For this purpose, the behavior of the further participant is predicted at regular time intervals in each case for a prediction interval (21) of a predetermined length, each prediction interval (21) comprising a predetermined number of successive prediction steps (31 to 35; 41 to 43). According to the invention, a cognitive uncertainty of the predicted behavior is determined for each prediction step of a prediction interval, and the predicted behavior is evaluated by means of a decision criterion in order to decide whether to take into account the behavior of the further participant predicted for the respective prediction step in the behavior plan of the first participant. The invention further relates to a system for carrying out the above method.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a computer-implemented method for planning the behavior of a first participant in a traffic scenario, in which the predicted behavior of at least one further participant in the traffic scenario is taken into account, in particular based on the epistemic uncertainty of the predicted behavior. The behavior of the further participant is predicted at regular time intervals for prediction intervals of a predetermined length, each prediction interval comprising a predetermined number of successive prediction steps. Background Art

[0002] Therefore, the prediction for the prediction interval consists of the predictions for the individual prediction steps of the prediction interval. Here, the prediction step can either correspond to a single time step or also to a time segment comprising a plurality of time steps. Usually, successive prediction intervals overlap with each other, wherein the degree of overlap is related to the time interval or frequency of the predictions and to the length of the prediction interval. Usually, the result of the behavior planning for the first participant of the traffic scene and also the predicted behavior of other participants of the traffic scene are described by trajectory data, which include position data and / or stop probability, movement data and / or orientation data of the corresponding participant for each prediction step of the prediction interval.

[0003] In order to be able to plan safe, understandable maneuvers, an automated vehicle must know how the current traffic scene will develop and in particular how other participants in the traffic scene, such as motor vehicles, cyclists and pedestrians, will behave in the future. To this end, the future behavior of other traffic participants is predicted and forwarded to the planning component of the automated vehicle. Traditional prediction methods usually perform dynamic-based predictions. The prediction methods only model the interactions between traffic participants to a limited extent. For this reason, in recent years, the use of machine learning, in particular deep learning (DL), has been established as the de facto standard for prediction. In addition, compared to traditional prediction methods, machine learning methods offer the possibility of taking extensive and diverse contextual information into account in the prediction.

[0004] In any case, the prediction of the behavior of the participants in the traffic scene is based on the scene-specific information collected at the current point in time. Based on this information, the DL-based prediction component generates a predetermined number of behavior options for the prediction interval, that is, for a defined duration of a future time segment. Within the scope of such a prediction, a plurality of different behavior options are usually predicted for the participants. However, applications are also conceivable for which only one behavior option is sufficient for the prediction.

[0005] Depending on how the prediction results are output, a distinction is made between so-called one-shot prediction methods and autoregressive prediction methods. In a one-shot prediction method, the behavior of the participant is predicted in a single step based solely on the aggregated scenario-specific information for the entire prediction interval, i.e. for all prediction steps of the prediction interval. In contrast, in the case of an autoregressive prediction method, in each prediction step of the prediction interval, in addition to the aggregated scenario-specific information, the prediction results of at least one previous prediction step of the prediction interval are also taken into account. Therefore, the behavior of the participant within a prediction interval is predicted by linking multiple single-step predictions for the individual prediction steps of the prediction interval. In this case, each prediction step affects the subsequent prediction steps of the prediction interval.

[0006] Predictions about the future are inherently unreliable and error-prone, wherein one must distinguish between the unknown nature of the future itself, known as aleatory uncertainty, and the uncertainty in predictions about the future, known as epistemic uncertainty.

[0007] The main cause of epistemic uncertainty in predictions is an inadequate performance of the perception and prediction components, which results in the epistemic uncertainty of predictions being subject to context-related fluctuations. Examples of perception defects include: inadequate sensor equipment, which results in insufficient collection of sufficient, convincing scene-specific data; imperfect sensor performance, which may be due to component conditions, aging conditions or ambient conditions; or incorrect interpretation of scene-specific data within the perception component. Inadequate performance of DL-based prediction components can usually be traced back to the training data used, which is reflected in the weights of the DL components, but can also be traced back to an architecture of the DL components that is unsuitable for the prediction problem. Summary of the invention

[0008] The present invention proposes measures that make it possible to take into account the epistemic uncertainty of the predictions and the temporal fluctuations of the epistemic uncertainty over the surroundings in the action planning. These measures can significantly increase the safety of maneuver planning overall.

[0009] According to the invention, this is achieved by determining the epistemic uncertainty of the predicted behavior for each prediction step of the prediction interval and evaluating the predicted behavior with the aid of a decision criterion in order to decide whether the behavior of the other participant predicted for the corresponding prediction step should be taken into account in the behavior planning of the first participant.

[0010] According to the invention, the measures are also implemented by a computer-implemented system for carrying out the method according to the invention.

[0011] Since the cognitive uncertainty of the predicted behavior is determined anew in each prediction step of the prediction interval according to the present invention, the time fluctuation of the cognitive uncertainty occurring in the prediction interval can also be taken into account in the behavior planning. For this reason, the cognitive uncertainty is compared with the decision criterion in each prediction step to decide whether the behavior of the other participants predicted for the prediction step can be considered in the behavior planning of the first participant. That is, the satisfaction of the decision criterion is a prerequisite for the following: the behavior of the other participants predicted in the prediction step is considered in the behavior planning of the first participant. Therefore, if the decision criterion is not met, the behavior of the other participants predicted by the prediction step is not considered in the behavior planning of the first participant. In this case, although other participants are also considered in the behavior planning of the first participant, the method is, for example, to vacate a safety area that covers all the movement possibilities of the other participants. But this measure is irrelevant to the behavior predicted for the other participants. If the decision criterion is met for a prediction step, the planning component can trust the prediction and thus determine the space that the other participants may occupy in the time period of the prediction step. In the behavior planning of the first participant, the space determined separately can be regarded as a non-travelable space.

[0012] Since the behavior planning according to the invention is essentially based only on prediction steps that are classified as reliable based on an assessment of the cognitive certainty of the predicted behavior, the proposed measures contribute significantly to the safety of maneuver planning.

[0013] In an advantageous embodiment of the method according to the invention, the epistemic uncertainty of the predicted behavior of the other participants is quantified, so that at least one threshold value can be selected as a decision criterion for the epistemic uncertainty.

[0014] In principle, there are various possibilities for determining or quantifying the epistemic uncertainty of the predicted behavior of the other participants. Thus, for example, various causes of the epistemic uncertainty can be monitored and evaluated.

[0015] Another possibility is to evaluate the predicted behavior itself. For this purpose, for example, variances can be introduced into the DL-based predictions in order to generate a predetermined number of different predictions for each behavior option for the given scenario-specific input data. The epistemic uncertainty of the predicted behavior can then be derived based on the deviations between the different predictions. In the case of DL-based autoregressive prediction methods, the epistemic uncertainty of the predicted behavior can be evaluated based on the reconstruction error, i.e., based on the deviation between the predicted behavior and the participant behavior reconstructed based on the predicted behavior. This proves to be advantageous regardless of the metric used in each case, because it is easy to handle if a common epistemic uncertainty value is determined for all predicted behavior options of a prediction step.

[0016] In an advantageous development of the method according to the invention, the decision criterion is context-dependent, i.e., is dependent on the respectively current traffic situation. For this purpose, the at least one threshold value may be dependent on the position and / or movement state of a participant in the traffic scene, in particular its speed. Alternatively or in addition thereto, the at least one threshold value may also be dependent on the current surrounding conditions of the traffic scene, in particular the geographical conditions, weather conditions, visibility conditions and / or road state conditions.

[0017] As already mentioned, according to the invention, the epistemic uncertainty of the predicted behavior is determined for each prediction step of the prediction interval in order to check with the aid of a decision criterion whether the behavior predicted for the other participant can be taken into account in the behavior planning of the first participant.

[0018] In a first variant of the method according to the invention, for each prediction step of the prediction interval a decision is made independently of the remaining prediction steps of the prediction interval whether the predicted behavior is actually taken into account in the behavior planning of the first participant. In this case, all prediction steps of the prediction interval that meet the decision criterion are taken into account in the behavior planning of the first participant.

[0019] In a second variant of the method according to the invention, in the behavior planning of the first participant, only those prediction steps of the prediction interval are taken into account for which the decision criterion is satisfied and for all preceding prediction steps, the decision criterion is also satisfied. That is, the behavior predicted for the other participants of the prediction interval is only taken into account in the behavior planning of the first participant as long as the decision criterion is satisfied. Starting from the first prediction step for which the decision criterion is not satisfied, the consideration of the predicted behavior of this prediction interval is terminated. This measure makes it possible to save computing power in the planning without thereby impairing the planning reliability.

[0020] If the autoregressive prediction method is used for the prediction of the behavior of other participants, it is also possible to save computing power within the scope of the prediction. For this purpose, in a third variant of the method according to the invention, the prediction of the prediction interval is interrupted as long as the epistemic uncertainty determined in the prediction step does not meet the decision criterion "taken into account when planning the behavior of the first participant". Since in the case of the autoregressive prediction method, the prediction of a prediction step is always also related to the prediction of the previous prediction step of this prediction interval, it should be assumed that the prediction step of the prediction interval that follows a prediction step with high epistemic uncertainty also has high epistemic uncertainty. Therefore, this prediction suspension does not have a negative impact on the planning reliability, especially because this suspension is limited to the current prediction interval and has no effect on the subsequent prediction intervals. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The measures according to the invention and preferred implementation options are explained in more detail with reference to the drawings.

[0022] Figure 1 : The functional mode of a computer-implemented system for executing a method for behavior planning for an automated vehicle according to the present invention is explained based on a block diagram.

[0023] Figure 2 : Two preferred embodiments of the method for behavior planning according to the present invention are drawn using a single prediction example in the upper half of the image and an autoregressive prediction example in the lower half of the image. DETAILED DESCRIPTION

[0024] Figure 1 The block diagram shows a perception component 1 of an at least partially automated vehicle, which aggregates scene-specific information 10 from different information sources inside the vehicle and, if necessary, outside the vehicle. Typically, the scene-specific information 10 is data detected by a camera, a lidar and / or a radar sensor. In addition to this, the scene-specific information 10 can also be detected with the aid of inertial sensors of the vehicle. The scene-specific information 10 is also often supplemented by GPS data and map data as well as weather data and data about road conditions. The scene-specific information 10 aggregated at a given point in time typically includes current sensor data and other data at that point in time. However, the scene-specific information may also include sensor data and other data collected within a predetermined time period up to a given point in time.

[0025] The collected scene-specific information 10 is evaluated by the perception component 1 in order to recognize and locate objects and participants in the current traffic scene. In the exemplary embodiment described here, the perception component 1 is also equipped with a neural network in order to generate a potential feature set for each participant based on the collected scene-specific information 10.

[0026] In the exemplary embodiment described here, this potential feature set is used as input for a DL-based prediction component 2 of the vehicle. This can be a one-shot prediction component or also an autoregressive prediction component. The neural network of the prediction component 2 generates a predetermined number of behavior options 12 for the participant at regular time intervals, specifically using the potential feature set 11 generated for the participant in each case. The prediction is performed for a predetermined time interval, which in the context of the present invention is referred to as a prediction interval or also as a prediction horizon. The totality of the predicted behavior options is referred to below as predicted behavior. Each prediction interval includes a predetermined number of prediction steps, each of which can correspond to one or more time steps. Therefore, the prediction is always based at least indirectly on scene-specific information 10 aggregated at the current point in time. For example, the future behavior of a participant can be predicted in the form of a trajectory, i.e. in the form of position data x, y of the participant at successive time points in the prediction interval. In addition to this, the trajectory information can also include movement information of the participant, such as speed data and acceleration data, and / or orientation information of the participant or information about spatial orientation, for example in the form of steering angle information.

[0027] In the exemplary embodiment described here, the prediction component 2 is designed to determine an uncertainty value for each prediction step, which quantifies the epistemic uncertainty of the predicted behavior 12. This uncertainty value is compared with a decision criterion in order to decide whether the behavior of another participant predicted for the corresponding prediction step should be taken into account in the participant's own behavior planning. In this case, the decision criterion can simply be a predetermined threshold value. However, the decision criterion can also be selected in a situation-dependent manner, for example, by modifying the decision criterion based on the current position of the vehicle and based on movement information or updating it at regular time intervals.

[0028] In any case, in the exemplary embodiment of the invention described here, predicted behavior 12 is only transmitted to planning component 3 when a decision criterion is met for the corresponding prediction step, ie, for example, when the epistemic uncertainty of predicted behavior 12 is below a predefined threshold value.

[0029] Therefore, in the behavior planning, the planning component 3 only considers those prediction steps that have been graded as less unreliable. For example, in this case, the planning component 3 can trust the prediction and regard the space that can be determined by the prediction and occupied by other participants in the future as undriveable. If the cognitive uncertainty value in the prediction step is higher than the threshold, the predicted behavior is not passed to the planning component 3 and therefore cannot be considered by the planning component 3. Instead, the planning component 3 assigns a safety area to the corresponding participant, which covers all the space that can potentially be driven by the participant in the prediction step. The safety area is considered to be undriveable by the planning component 3. This has nothing to do with the predicted behavior of the participant.

[0030] This method contributes significantly to the planning reliability of the planning component 3 .

[0031] It should be noted here that the determination of the epistemic uncertainty and its evaluation according to the decision criteria can alternatively also take place in a separate component connected downstream of the prediction component, which then makes a decision on forwarding the predicted behavior to the planning component.

[0032] However, the method according to the invention can also be implemented if the planning component determines and / or evaluates the epistemic uncertainty of the predicted behavior. In this case, the planning component will decide whether to consider a certain prediction step in the behavior planning.

[0033] for Figure 2 , an abstract representation is chosen in order to illustrate the single prediction method in the upper image part and the autoregressive prediction method in the lower image part. The current time point 20 and the subsequent prediction interval 21 with five prediction steps 31 to 35 or 41 to 43 are respectively shown. Each prediction step 31 to 35 or 41 to 43 can correspond to a time step, or the prediction interval corresponds to a time segment with a plurality of time steps. In the case of the single prediction shown in the upper image part, the predictions for the individual prediction steps 31 to 35 are generated independently of each other, specifically only based on the scene-specific information collected up to the current time point 20, which is indicated by the arrows, which start from the current time point 20 and point to the prediction steps 31, 32, 33, 34 or 35. In contrast, in the case of the autoregressive prediction shown in the lower image part, the predictions of the individual prediction steps 41, 42, 43 are based not only on the scene-specific information collected up to the current time point 20, but also on the predictions of the previous prediction steps of the prediction interval 21. This is likewise symbolized here by arrows which each start from a prediction step of the prediction interval and run into the next prediction step of the prediction interval.

[0034] In both cases - single-shot forecast and autoregressive forecast - according to the invention, the epistemic uncertainty σ of the predicted behavior is determined for each forecast step 31 to 35 or 41 to 43 of the forecast interval 21 t , and evaluates this epistemic uncertainty with the aid of a threshold τ in order to decide whether the behavior predicted for the corresponding prediction step should be forwarded to the planning component for behavior planning.

[0035] In the single prediction method of the embodiment described here, the epistemic uncertainty of prediction steps 33 and 34 is above the threshold value τ, and the epistemic uncertainty for the remaining prediction steps 31, 32 and 35 is below the threshold value, so that the decision criterion is satisfied here. However, in the present embodiment, only the predicted behavior of prediction steps 31 and 32 is forwarded to the planning component. In addition to the decision criterion τ, as a further requirement, all earlier prediction steps of the prediction interval should also be classified as reliable. Therefore, neither the predicted behavior of "unreliable" prediction steps 33 and 34 nor the predicted behavior of "reliable" prediction step 35 is forwarded to the planning component, although prediction step 35 satisfies the decision criterion τ.

[0036] In the case of the autoregressive forecasting method, it is checked in each of the consecutive forecasting steps 41, 42, ... whether the epistemic uncertainty is below the threshold value τ. Here, the decision criterion is met for the first two forecasting steps 41 and 42 of the forecasting interval 21. Therefore, the predicted behavior of these forecasting steps 41 and 42 will be passed to the planning component. However, in the case of forecasting step 43, this criterion is not observed, so that this forecasting step and all other forecasting steps of the forecasting interval 21 should no longer be considered in the behavior planning. Therefore, the forecasting step of the forecasting interval 21 is terminated here. As a result, two forecasting steps are saved in this case. This measure can significantly improve the efficiency of the behavior planning as a whole.

[0037] The following notation is used in the description of the two method variants discussed above:

[0038] t Index of the prediction step

[0039] σ t Epistemic uncertainty value at the prediction step

[0040] τUncertainty threshold

[0041] N is the number of forecast steps for the forecast interval

[0042] Single prediction method:

[0043] 1. Forecast all N forecast steps

[0044] 2. Determine the epistemic uncertainty value σ for all N prediction stepst

[0045] 3. Set all uncertainty values ​​σ t Compare with the threshold τ

[0046] 4. Set its uncertainty value below the threshold (σ t <τ) are passed to the planning component if the uncertainty values ​​of all previous forecast steps are also below the threshold (σ t-i <τ),

[0047] Autoregressive forecasting method:

[0048] 1. Prediction prediction step t

[0049] 2. Determine the epistemic uncertainty value σ for this prediction step t

[0050] 3. Set the uncertainty value σ t Compare with the threshold τ

[0051] a. σ t <τ: Repeat steps 1 to 3 for the next prediction step

[0052] b. σ t ≥τ: prediction ends

[0053] 4. Pass all previous forecast steps to the planning component

Claims

1. A computer-implemented method for behavior planning for a first participant in a traffic scenario, in which a predicted behavior of at least one further participant in the traffic scenario is taken into account, in particular based on an epistemic uncertainty of the predicted behavior, in, predicting the behavior of the further participant at regular time intervals for prediction intervals (21) of a predetermined length, wherein each prediction interval (21) comprises a predetermined number of successive prediction steps (31 to 35; 41 to 43); Characterized in that the epistemic uncertainty of the predicted behavior is determined for each prediction step (31 to 35; 41 to 43) of the prediction interval (21), and the predicted behavior is evaluated with the aid of a decision criterion in order to decide whether the predicted behavior of the other participant for the corresponding prediction step (31 to 35; 41 to 43) is to be taken into account in the behavior planning for the first participant.

2. The method according to claim 1, characterized in that The epistemic uncertainty of the predicted behavior of the further participant is quantified, and at least one threshold value is selected for the epistemic uncertainty as a decision criterion.

3. The method according to claim 2, characterized in that At least one context-related threshold is selected as a decision criterion, wherein the threshold is particularly related to the position and / or movement state of the participants of the traffic scene and / or to the current surrounding environmental conditions of the traffic scene, in particular the geographical conditions, weather conditions, visibility conditions and / or road state conditions.

4. The method according to any one of claims 1 to 3, characterized in that In the behavior planning for the first participant, the predicted behavior of the further participant for all prediction steps from the prediction interval that satisfy the decision criterion is taken into account.

5. The method according to any one of claims 1 to 3, characterized in that In the behavior planning for the first participant, the predicted behavior of the other participant from all prediction steps (31, 32; 42, 42) of the prediction interval (21) that meet the decision criterion is taken into account if the decision criterion is also met for all previous prediction steps of the prediction interval (21).

6. The method according to any one of claims 1 to 3, wherein: An autoregressive prediction method is used for predicting the behavior of the further participant, characterized in that the prediction in the prediction interval is terminated as soon as the epistemic uncertainty determined in the prediction step does not meet the decision criterion "taken into account when planning the behavior of the first participant".

7. A computer-implemented system for behavior planning, the system being configured to perform the method according to any one of claims 1 to 6.