Methods for training a probabilistic estimation model, prediction methods and systems

Abstract

The invention relates to a prediction method for predicting a Reachable-Set-Based Time to Collision (RSB-TTC) using a trained probabilistic estimation model, the prediction method comprising the steps: - Providing (S1) a sensor scan captured by an in-vehicle sensor, at least one movement information of the vehicle, and a selection of at least one collision-relevant road user in the sensor scan; - probabilistic prediction and / or estimation (S2) of the RSB-TTC based on the sensor scan, which contains at least one movement information of the vehicle, and the selection of at least one road user in the sensor scan; and - Determine (S3) a prediction uncertainty for the probabilistic prediction and / or estimate.

Description

[0001] The invention relates to a method for the at least partially supervised training of a probabilistic estimation model. The invention relates to a prediction method and a system for predicting a reachable set-based time to collision (RSB-TTC). The invention further relates to a computer program with program code and a computer-readable data carrier with program code of a computer program. State of the art

[0002] Road safety is crucial, especially in complex urban scenarios where pedestrians, cars, and other road users converge in close quarters. Concepts such as "reachable sets" have been developed to better understand potential hazards and develop preventive measures.

[0003] Reachable sets describe the range a pedestrian, car, or other road user can occupy in a worst-case scenario under certain assumptions. Such assumptions can include maximum acceleration, current speed, and a specific motion model.

[0004] This is described, for example, in the scientific publication by P. Schneider, M. Butz, C. Heinzemann, J. Oehlerking and M. Woehrle, “Towards threat metric evaluation in complex urban scenarios,” 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), Indianapolis, IN, USA, 2021, pp. 1192-1198, doi: 10.1109 / ITSC48978.2021.9564790.

[0005] Based on these reachable sets, a given situation can be assessed for its criticality. Taking into account the movement uncertainty of individual road users, an assessment and / or prediction can be made, for example, of when the reachable sets of two road users intersect. The reachable sets serve as the basis for calculating a robust time-to-collision (TTC) as a proxy for the criticality of a given traffic situation.

[0006] Driver assistance systems should be able to handle and evaluate complex environmental situations. For example, it must be ensured that all relevant objects, especially those currently or potentially currently participating in traffic, are detected in the vehicle's vicinity. One factor determining the relevance of an object is the criticality of a perceived scene. Evaluating and / or testing the safe functionality of driver assistance systems requires a large amount of relevant data, i.e., data on critical traffic situations. The TTC can be used as a proxy for this.

[0007] In addition to reachable sets, there are other methods for calculating criticality based on a TTC metric. The scientific publication by Badki, Abhishek et al., "Binary TTC: A Temporal Geofence for Autonomous Navigation." 2021 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021): 12941-12950, ​​describes how to calculate binary TTC (time-to-contact) maps based on two consecutive images from a monocular camera. A series of binary, per-pixel classifications determines whether a road user will collide with an image plane at a given time interval. The authors also use optical flow maps to refine the binary TTC maps.

[0008] From the scientific publication H. Kataoka, T. Suzuki, K. Nakashima, Y. Satoh and Y. Aoki, "Joint Pedestrian Detection and Risk-level Prediction with Motion-Representation-by-Detection," 2020 IEEE International Conference on Robotics and Automation (ICRA), Paris, France, 2020, pp. 1021-1027, doi: 10.1109 / ICRA40945.2020.9197399, a pedestrian near-miss detector is known that includes the prediction of bounding boxes with the corresponding criticality. For this purpose, the authors created their own Pedestrian Near-Miss (PNM) dataset, which categorizes the near-miss scenarios based on their risk level. To define a near-miss scenario, the authors apply thresholds based on TTC. Using an adapted SSD architecture (SSD-MR), the authors combine temporal information of a pedestrian sequence to predict the collision risk (high-, low-, or no-risk) using temporal activations.

[0009] From the scientific publication M. Wolf, L.R. Douat, and M. Erz, "Safety-Aware Metric for People Detection," 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), 2021, pp. 2759-2765, doi: 10.1109 / ITSC48978.2021.9564734, an assessment of the criticality for vulnerable pedestrians in off-road scenarios (e.g., construction sites) is known. This assessment is based on the distance and the collision time. The authors define a safety-critical zone around the emergency vehicle and determine so-called frames-to-collision (FTTC), measured in pixels / frame, as a proxy for TTC for vulnerable pedestrians moving toward the criticality zone.

[0010] From the scientific publication "Evaluating the Consequences of Object (mis)Detection from a Safety and Reliability Perspective: Discussion and Measures" (2022) by Ceccarelli, Andrea, and Leonardo Montecchi, it is known to determine a potential collision point based on the current trajectory of road users and to assess the criticality based on TTS (time-to-stop) or "collision time criticality," as a proxy for TTC.

[0011] The advantages of reachable set based TTC are discussed in detail in the scientific paper M. Lyssenko, C. Gladisch, C. Heinzemann, M. Woehrle and R. Triebel, “Towards Safety-Aware Pedestrian Detection in Autonomous Systems,” 2022 IEEE / RSJ International Conference on Intelligent Robots and Systems (IROS), Kyoto, Japan, 2022, pp. 293-300, doi: 10.1109 / IROS47612.2022.9981309.

[0012] The key difference lies in the consideration of movement uncertainties of road users and the flexible model parameterization. Therefore, this methodology is also used as a training basis. In order for the TTC calculated using the reachable sets (hereinafter also referred to as reachable-set-based (RSB) TTC) to deliver meaningful values, it is advantageous to make further assumptions in addition to maximum accelerations, etc. These include, for example, the assumption that vehicles stay in their respective lanes, people walk past obstacles, etc. In order to be able to include all of these values ​​in the calculation of the TTCs, a precise map of the environment is necessary, which includes, for example, information on the exact road geometry and lane geometry, etc.

[0013] The need for all this information limits the practical application of reachable sets. For example, if only a camera image of a test vehicle is recorded and no map information is available, a TTC cannot be calculated using reachable sets. This problem requires a solution.

[0014] In light of the prior art, the invention is therefore based on the object of providing an improved method and / or system for specifying a reachable set-based TTC and a training method for training a machine learning model.

[0015] The problem is solved by a method for at least partially supervised training of a probabilistic estimation model according to the features of patent claim 1. The problem is solved by a prediction method for predicting a reachable-set-based time to collision, RSB-TTC, using a trained probabilistic estimation model according to the features of patent claim 7. The problem is solved by a system for predicting a reachable-set-based time to collision, RSB-TTC, using a trained probabilistic estimation model according to the features of patent claim 12. Disclosure of the invention

[0016] According to a first aspect, a method is provided for the at least partially supervised training of a probabilistic estimation model for predicting a reachable-set-based time to collision (RSB-TTC). The method comprises the steps of: providing a training data set, at least comprising a plurality of training data triplets, each comprising a sensor scan, a selection of at least one collision-relevant road user in the respective sensor scan, and each comprising a predetermined RSB-TTC; at least partially supervised training of the probabilistic estimation model based on the training data set by optimizing a model-specific loss function; and providing the trained probabilistic estimation model.

[0017] In a training phase, also referred to as the offline phase, data is collected and / or provided that is used to train a machine learning method, in this case the probabilistic estimation model. The goal is to train the probabilistic estimation model in such a way that, in subsequent inference, it can directly predict the RSB-TTC based on a sensor scan and a selected region or selection in the sensor scan that indicates a potentially collision-relevant road user. Supervised learning techniques for probabilistic prediction models are preferably used to train the probabilistic estimation model. This requires data triplets, each comprising a sensor scan, a selection of the (collision-)relevant road user in the respective sensor scan, and the corresponding RSB-TTC.The RSB-TTC for training the probabilistic estimation model is preferably determined based on a map with annotated road users. The present solution for probabilistic estimation of the RSB-TTC is therefore essentially based on machine learning methods.

[0018] Reachable-Set-Based TTC (Time to Collision) is a method for calculating a point in time at which two road users could be at a potentially critical distance from each other. The reachable set describes the range in which a road user can move under certain physical constraints (such as maximum acceleration and speed). The Time to Collision (TTC) is a parameter used to assess potential collisions or hazardous situations in traffic. It indicates how much time remains until two road users would collide at a certain point if they continued to move at their current speeds and in the same direction.A small TTC value indicates an imminent collision and signals a potentially dangerous situation, while a larger TTC value indicates that there is sufficient time to avoid a collision. Using reachable-set-based TTC allows for a more accurate and realistic assessment of the collision risk, as the movement restrictions of the respective road user are incorporated into the calculation. This method is particularly important in complex urban traffic scenarios, as it can help prevent accidents and improve road safety.

[0019] It is understood that the steps according to the invention, as well as other optional steps, do not necessarily have to be performed in the order shown, but can also be performed in a different order. Furthermore, additional intermediate steps can be provided. The individual steps can also comprise one or more substeps without thereby departing from the scope of the method according to the invention.

[0020] In a preferred embodiment, the respective RSB-TTC in the respective training data triplet is determined based on a map with annotated road users. The "map with annotated road users" preferably refers to a map or an image of a street scene on which the positions or movements of road users (such as cars, pedestrians, or other objects) have been manually or automatically annotated or marked.

[0021] In a preferred embodiment, the probabilistic estimation model comprises a probabilistic neural network for estimating the RSB-TTC and / or a mixture density network and / or a quantile regression neural network. In particular, a quantile regression neural network is designed to directly estimate the quantiles and thus allows direct derivation of statements about the approximation quality of the RSB-TTC, e.g., the 99% quantile. A probabilistic neural network is an extension of a conventional neural network that, in addition to the normal forward calculations, also takes into account the uncertainty in its predictions. While conventional neural networks provide deterministic outputs that can be interpreted as fixed values, a probabilistic neural network enables the estimation of the uncertainty or probability of its predictions.A mixture density network (MDN) is a neural network used to model the probability distribution of a dependent variable as a function of several independent variables. It is a special type of probabilistic model that is particularly used in modeling complex, multidimensional data and estimating uncertain predictions. The advantage of a mixture density network is that it allows for the modeling of complex relationships between the input variables and the target variable while simultaneously accounting for the uncertainty in the predictions. The network can generate a series of probability distributions representing various possible outcomes, thus quantifying the epistemic uncertainty in the predictions. A quantile regression neural network (QRNN) is a special type of neural network used in quantile regression.In contrast to traditional regression, which attempts to estimate the expected or mean value of a dependent variable based on independent variables, quantile regression aims to model different quantiles of the distribution of the dependent variable.

[0022] In a preferred embodiment, the probabilistic estimation model is combined with a sensor domain-specific backbone network. The probabilistic estimation model is particularly preferably trained separately from the backbone network or together with the backbone network. The method described here is general and also includes other network architectures and backbones, such as feature pyramid networks. The probabilistic estimation model is preferably trained holistically, i.e. the weights of the backbone network are optimized together with the weights of the probabilistic estimation model. It is also conceivable to carry out a different training method, e.g. exclusive fine-tuning by adjusting the weights and / or optimizing a cost function of the probabilistic estimation model. For the optimization, i.e. the training of the probabilistic estimation model, a loss function orA loss function (also called loss) appropriate to the probabilistic network type is used. In the case of quantile regression, for example, this is the pinball loss. In the case of mixture density networks, it is cross-entropy. In this case, the probabilistic network or model is preferably combined with a backbone network, especially an existing one. The backbone network is preferably selected to match the sensor domain. In the case of a camera, the backbone network can, for example, comprise a ResNet.

[0023] In a preferred embodiment, the backbone network is trained to extract image attributes from the respective sensor scan. A backbone network preferably refers to the main part of a deep neural network, which is responsible for extracting features from the input data. It is the core or "backbone" component of such a network and contributes significantly to the performance and capability of the entire model. The backbone network comprises a number of layers that perform various types of convolution and pooling operations. These operations enable the network to hierarchically process the information in the input data and extract abstract features at different levels. In image processing applications, for example, a backbone network can extract visual features such as edges, texture, shapes, and objects from the raw pixels.The trained backbone network can then be used as a feature extractor or preprocessing operation before the output is passed on to a task such as classification, object detection, segmentation, or another specific application. It is common to use the pre-trained backbone network as a basis for transfer learning, reusing the learned features for related tasks or fine-tuning them for specific datasets. Examples of backbone networks in image processing include the convolutional neural network (CNN), ResNet (Residual Network), VGG (Visual Geometry Group), Inception, and many others.

[0024] In a preferred embodiment, the respective sensor scan comprises at least one image and / or video datum. Alternatively or additionally, the selection of the at least one road user in the sensor scan is carried out using a bounding box, particularly if the sensor scan comprises image and / or video datum. A bounding box, also called a limiting box or enclosing box, is a rectangular or cuboid-shaped geometric construct that serves to describe the extent of an object or a group of objects in space. It is frequently used in image processing, computer vision, 3D graphics, machine learning, and other fields to capture the position, size, and orientation of an object. A bounding box is defined by the coordinates of two diagonally opposite corners, which define the enclosed area.In two-dimensional (2D) space, a bounding box typically has four parameters: the x- and y-coordinates of the upper left corner and the x- and y-coordinates of the lower right corner. In three-dimensional (3D) space, a bounding box can have six parameters: the x-, y-, and z-coordinates of the leading edge and the x-, y-, and z-coordinates of the trailing edge. Bounding boxes are used to enclose objects in an image or 3D model to facilitate the detection, tracking, or analysis of those objects. In machine learning, they are used as annotations / selections to teach the model to identify and / or classify specific objects in images or videos.

[0025] According to a second aspect, a prediction method for predicting a reachable-set-based time to collision (RSB-TTC) using a trained probabilistic estimation model is proposed. The prediction method comprises the steps of: providing a sensor scan acquired by an in-vehicle sensor, at least one piece of vehicle movement information, and a selection of at least one collision-relevant road user in the sensor scan; probabilistically predicting and / or estimating the RSB-TTC based on the sensor scan, the at least one piece of vehicle movement information, and the selection of the at least one road user in the sensor scan by the trained probabilistic estimation model; and determining a prediction uncertainty for the probabilistic prediction and / or estimate by the trained probabilistic estimation model.

[0026] In contrast to “deterministic” prediction methods, i.e. for example regression methods which optimize a mean square error, probabilistic prediction methods, as used here, enable an estimate of the prediction uncertainty. The statement about the prediction uncertainty can be used to provide a statistical estimate of a prediction error. It is also possible to make a statement about an expected worst case scenario based on this. By applying these techniques, a statement can preferably be made about the approximation quality of the RSB-TTC estimated by the prediction method. For this purpose, the quantiles of a predicted probability distribution are preferably considered, e.g. a 90%, a 95% or a 99% quantile. Preferably, any quantile between 0% and 100% (exclusively 100%) is possible. If, for example, one considersthe 99% quantile, this would mean that in 99% of all cases, the RSB-TTC is not greater than the value estimated for the quantile. This can be understood as a kind of statistical guarantee regarding the approximation quality.

[0027] The present prediction method can be used, in particular, to approximate a reachable-set-based TTC from a single vehicle to other road users. This requires, in particular, information about the vehicle itself (e.g., its speed and / or acceleration and / or yaw rate) and information from an on-board sensor, in particular an imaging sensor, as well as the selection of at least one other road user. The information from the sensor can, for example, be image and / or video data from a camera included in the vehicle. The selection can, for example, be a bounding box.

[0028] The present prediction method thus estimates the respective RSB-TTC. The method is also capable of providing statistical guarantees regarding the quality of the estimate. To achieve this, methods of probabilistic prediction and machine learning are used.

[0029] The trained probabilistic estimation model preferably receives image and / or video data from a sensor and motion information about the vehicle as inputs. In the preferred backbone network, features are preferably extracted from the image and / or video data and / or from the motion information. In addition, the probabilistic estimation model is provided with the section or a selection of the sensor scan (e.g., a corresponding image section in a camera image) comprising the object for which the RSB-TTC is to be calculated. Based on these features, the probabilistic estimation model then probabilistically predicts the RSB-TTC.

[0030] In a preferred embodiment, the determined prediction uncertainty is used to provide a statistical estimate of a prediction error.

[0031] In a preferred embodiment, determining the prediction uncertainty involves determining a probability distribution. The determined probability distribution is preferably calibrated. To ensure that the statistical guarantees are usable and meaningful (i.e., that the 90% quantile truly indicates that 90% of the RSB-TTC is smaller than the estimated value for the quantile), it is preferable to perform an additional calibration of the output probability distributions if necessary. Probability calibration techniques are preferably used for this purpose. However, this step is not absolutely necessary.

[0032] In a preferred embodiment, the trained probabilistic estimation model has been trained according to the training method in one of its embodiments. After training and any probability calibration (both of which are preferably performed offline), the trained probabilistic estimation model can be operated in online mode by inputting sensor data.

[0033] In a preferred embodiment, the respective sensor scan is captured and / or provided by a camera and / or a lidar sensor and / or an ultrasonic sensor and / or a multidirectional camera and / or another imaging sensor. In principle, other sensor types not explicitly listed here are also usable.

[0034] According to a third aspect, a system for predicting a reachable-set-based time to collision (RSB-TTC) using a trained probabilistic estimation model is proposed. The system comprises at least one evaluation and / or computing device configured to perform the following steps: providing a sensor scan acquired by a sensor, at least one piece of movement information of the vehicle, and a selection of at least one road user in the sensor scan; probabilistically predicting and / or estimating the RSB-TTC based on the sensor scan, the at least one piece of movement information of the vehicle, and the selection of the at least one road user in the sensor scan by the trained probabilistic estimation model; and determining a prediction uncertainty for the probabilistic prediction and / or estimate by the trained probabilistic estimation model.The evaluation and / or computing device may further comprise a provision device. The provision device may also be provided independently of the evaluation and / or computing device and serve to provide the aforementioned data for further processing.

[0035] According to the invention, a control device is also claimed which is included in an autonomous vehicle and / or a robotic system and / or an industrial machine and on which the present prediction method can be executed.

[0036] The invention also claims a computer program with program code for executing at least parts of the method according to the invention in one of its embodiments when the computer program is executed on a computer. In other words, the invention provides a computer program (product) comprising instructions that, when executed by a computer, cause the computer to execute the method / steps of the method according to the invention in one of its embodiments.

[0037] According to the invention, a computer-readable data carrier with program code of a computer program is also proposed for executing at least parts of the method according to the invention in one of its embodiments when the computer program is executed on a computer. In other words, the invention relates to a computer-readable (storage) medium comprising instructions which, when executed by a computer, cause the computer to execute the method / steps of the method according to the invention in one of its embodiments.

[0038] The described designs and further training courses can be combined as desired.

[0039] Further possible embodiments, developments and implementations of the invention also include combinations of features of the invention described previously or below with regard to the embodiments that are not explicitly mentioned. Short description of the drawings

[0040] The accompanying drawings are intended to provide a further understanding of embodiments of the invention. They illustrate embodiments and, in conjunction with the description, serve to explain principles and concepts of the invention.

[0041] Other embodiments and many of the aforementioned advantages will become apparent upon review of the drawings. The elements illustrated in the drawings are not necessarily drawn to scale.

[0042] They show: Fig. 1 is a flowchart of an embodiment of the present prediction method; Fig. 2 is a block diagram of an embodiment of the present training method; and Fig. 3 a block diagram of an embodiment of the present prediction method during inference.

[0043] In the figures of the drawings, the same reference symbols designate the same or functionally equivalent elements, parts or components, unless otherwise stated.

[0044] Fig. Figure 1 shows a schematic flow diagram of a prediction method for predicting a Reachable-Set-Based Time to Collision, RSB-TTC, using a trained probabilistic estimation model.

[0045] In any embodiment, the method can be carried out at least partially by a system 1, which for this purpose can comprise several components not shown in detail, for example, one or more provision devices and / or at least one evaluation and computing device. It is understood that the provision device can be designed jointly with the evaluation and computing device or can be different from it. Furthermore, the system can comprise a storage device and / or an output device and / or a display device and / or an input device.

[0046] The prediction method, in particular computer-implemented, comprises at least the following steps: In a step S1, a sensor scan is provided which is detected by a sensor, at least one piece of movement information of the vehicle, and at least one road user is selected in the sensor scan.

[0047] In a step S2, probabilistic prediction and / or estimation of the RSB-TTC is carried out on the basis of the sensor scan, the at least one piece of movement information of the vehicle, and the selection of the at least one road user in the sensor scan.

[0048] In a step S3, a prediction uncertainty for the probabilistic prediction and / or estimation is determined.

[0049] Fig. 2 shows a block diagram of an embodiment of the present training method. The method for at least partially supervised training of a probabilistic estimation model 200 for predicting a reachable-set-based time to collision (RSB-TTC). The method comprises at least the following steps: A training data set is provided, at least comprising a plurality of training data triplets, each comprising a sensor scan 202, a selection 203 of at least one road user in the respective sensor scan 202, and a respective predetermined RSB-TTC 204. The respective RSB-TTC 204 is preferably determined in the respective training data triplet based on a map 206 with annotated road users. In this case, the reachable sets are preferably used based on the map 206 to estimate a potential collision between at least two road users.The map 206 preferably includes marked road users, in particular with information about a type of user (person, car, motorcycle, bicycle, etc.), a dimension, a position, and / or a speed. Furthermore, information 208 about a movement status of the vehicle for which collision monitoring is to be performed can be provided. The respective RSB-TTC 204 is preferably associated together with the respective sensor scans 202 and / or the information 208 to provide the training data set 210. The probabilistic estimation model 212 is then trained at least partially supervised on the basis of the training data set 210 by optimizing a model-specific loss function. If necessary, the probability distribution resulting from the estimation model 212 can also be calibrated, as indicated by reference numeral 214. The estimation model 212 is trained in offline mode.

[0050] In Fig.3 shows a block diagram of an embodiment of the present prediction method during inference. At least one sensor scan 300, which is detected by a sensor, at least one piece of movement information 302 of the vehicle, and a selection 304 of at least one road user in the sensor scan 300 are provided S1. Furthermore, a probabilistic prediction and / or estimation S2 of the RSB-TTC 306 is performed based on the sensor scan 300, the at least one piece of movement information 302 of the vehicle, and the selection 304 of at least one road user in the sensor scan 300 by the trained probabilistic estimation model 212. Furthermore, the determination S3 of a prediction uncertainty 308 for the probabilistic prediction and / or estimate is performed by the trained probabilistic estimation model 212.The inference of the trained estimation model 212 takes place in online mode, for example by providing sensor scans 300 in real time and / or directly while the vehicle is driving. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited non-patent literature

[0000] P. Schneider, M. Butz, C. Heinzemann, J. Oehlerking and M. Woehrle, “Towards threat metric evaluation in complex urban scenarios,” 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), Indianapolis, IN, USA, 2021, pp. 1192-1198, doi: 10.1109 / ITSC48978.2021.9564790

[0004] Badki, Abhishek et al. „Binary TTC: A Temporal Geofence for Autonomous Navigation.“ 2021 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021): 12941-12950

[0007] H. Kataoka, T. Suzuki, K. Nakashima, Y. Satoh and Y. Aoki, „Joint Pedestrian Detection and Risk-level Prediction with Motion-Representation-by-Detection,“ 2020 IEEE International Conference on Robotics and Automation (ICRA), Paris, France, 2020, pp. 1021-1027, doi: 10.1109 / ICRA40945.2020.9197399

[0008] M. Wolf, L. R. Douat and M. Erz, „Safety-Aware Metric for People Detection,“ 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), 2021, pp. 2759-2765, doi: 10.1109 / ITSC48978.2021.9564734

[0009] Ceccarelli, Andrea and Leonardo Montecchi. „Evaluating the Consequences of Object (mis) Detection from a Safety and Reliability Perspective: Discussion and Measures.“ (2022)

[0010] M. Lyssenko, C. Gladisch, C. Heinzemann, M. Woehrle and R. Triebel, „Towards Safety-Aware Pedestrian Detection in Autonomous Systems,“ 2022 IEEE / RSJ International Conference on Intelligent Robots and Systems (IROS), Kyoto, Japan, 2022, pp. 293-300, doi: 10.1109 / IROS47612.2022.9981309

[0011]

Claims

[1] Method for at least partially supervised training of a probabilistic estimation model for predicting a Reachable-Set-Based Time to Collision, RSB-TTC, the method comprising the steps: - Providing a training data set (210) comprising at least a plurality of training data triplets, each comprising a sensor scan (202), a selection (203) of at least one road user in the respective sensor scan (202) and a predetermined RSB-TTC (204), - at least partially supervised training (S2) of the probabilistic estimation model (212) on the basis of the training data set (210) by optimizing a model-specific loss function; and - Providing the trained probabilistic estimation model. [2] Method according to claim 1, wherein the respective RSB-TTC (204) in the respective training data triplet is determined on the basis of a map (206) with annotated road users. [3] The method according to claim 1 or 2, wherein the probabilistic estimation model (212) comprises a probabilistic neural network for estimating the RSB-TTC and / or a mixture density network and / or a quantile regression neural network. [4] Method according to one of the preceding claims, wherein the probabilistic estimation model (212) is combined with a backbone network, and wherein the probabilistic estimation model (212) is trained separately from the backbone network or jointly with the backbone network. [5] The method of claim 4, wherein the backbone network is trained to extract image attributes from the respective sensor scan (202,300). [6] Method according to one of the preceding claims, wherein the respective sensor scan (202, 300) has at least one image and / or video datum, and wherein the selection (203) of the at least one road user in the sensor scan (202, 300) is carried out by means of a bounding box. [7] Prediction method for predicting a Reachable-Set-Based Time to Collision, RSB-TTC using a trained probabilistic estimation model (212), the prediction method comprising the steps: - Providing (S1) a sensor scan (300) detected by a sensor, at least one item of movement information (208, 302) of the vehicle, and a selection of at least one road user in the sensor scan (300); - probabilistic prediction and / or estimation (S2) of the RSB-TTC based on the sensor scan (300), the at least one piece of movement information (208, 302) of the vehicle, and the selection of the at least one road user in the sensor scan (300) by the trained probabilistic estimation model (212); and - Determining (S3) a prediction uncertainty about the probabilistic prediction and / or estimation by the trained probabilistic estimation model (212). [8] The method of claim 7, wherein the determined prediction uncertainty is used to provide a statistical estimate of a prediction error. [9] A method according to claim 7 or 8, wherein determining the prediction uncertainty comprises determining a probability distribution, and wherein the determined probability distribution is preferably calibrated. [10] A method according to any one of claims 7 to 9, wherein the trained probabilistic estimation model is trained according to the method according to any one of claims 1 to 6. [11] Method according to one of the preceding claims, wherein the respective sensor scan (202, 300) is captured and / or provided by a camera and / or a lidar sensor and / or an ultrasonic sensor and / or a multidirectional camera and / or another imaging sensor. [12] System (1) for predicting a Reachable-Set-Based Time to Collision, RSB-TTC using a trained probabilistic estimation model (212), the system comprising: - an evaluation and / or computing device designed to carry out the following steps: ◯ Providing a sensor scan (300) detected by a sensor, at least one piece of movement information (208, 302) of the vehicle, and a selection of at least one road user in the sensor scan (300); ◯ probabilistic prediction and / or estimation of the RSB-TTC based on the sensor scan (300), the at least one piece of movement information (208, 302) of the vehicle, and the selection of the at least one road user in the sensor scan by the trained probabilistic estimation model (212); and ◯ Determining a prediction uncertainty for the probabilistic prediction and / or estimation by the trained probabilistic estimation model (212). [13] Computer program with program code to carry out at least parts of a method according to one of claims 1 to 10 when the computer program is executed on a computer. [14] Computer-readable data carrier with program code of a computer program for carrying out at least parts of a method according to one of claims 1 to 10 when the computer program is executed on a computer.

Images