Predicting a vehicle's trajectory

Abstract

Methods for predicting the driving paths of a vehicle, including: Providing the current position of the vehicle; Determining a group of future possible routes of the vehicle using a digital map and optionally also the current position and / or a specified maximum size of each route, in particular a maximum length of each route, a maximum travel time for each route, or a maximum number of driving maneuvers per route; Providing measurements of predetermined types of driving-related actions of the vehicle's driver, in particular the speed or acceleration of the vehicle, the setting of a turn signal, and / or the point of view to which a driver looks in the vicinity of the vehicle; Providing, for each driving path of the group of driving paths, one or more models, each providing an occurrence measure for one or more types of driver actions, wherein the occurrence measure provides a measure of the probability that the measurements of the driving-related actions occur while driving the vehicle on the respective driving path; Determine the respective impact factor for each route and for each model; Determining total occurrence metrics for each route from the group of routes, namely a metric based on the totality of occurrence metrics determined for the respective route, in particular the product of the occurrence metrics of a respective route; Determining an error occurrence measure, wherein the error occurrence measure in particular has a predetermined value, wherein the error occurrence measure represents the probability that the set of possible routes is incomplete and / or faulty and in particular represents the totality of probabilities of traveling on such routes that are not included in the set of future possible routes, and takes into account the probability of actions by the driver that are contrary to the traffic rules; Determine, for each route from the group of routes, an outcome performance measure, namely a measure for the probability that the respective route is travelled, based on the total occurrence measure determined for the respective route and the total occurrence measures determined for the routes of the group of routes, including the determined error occurrence measure.

Description

[0001] The invention relates to a method for predicting the driving path of a vehicle and a corresponding driver assistance system.

[0002] Nowadays, vehicles, especially passenger cars, are equipped with systems that support the driver in their driving tasks; these are known as driver assistance systems. Driver assistance systems are known to warn the driver of potential conflicts with other road users in certain situations or, in particularly critical situations, to intervene in the vehicle's dynamics by delaying the driver. It is recognized that potential conflicts must be assessed in terms of their probability of occurrence in order to avoid distracting the driver with numerous unnecessary warnings and thus potentially endangering driving safety. The actual probability of a conflict occurring depends both on the maneuver preferred by the driver and on their situational awareness, i.e., the driver's decision as to whether the preferred maneuver is feasible.This decision may be revised even a few seconds later if the driver receives additional information during this time, for example, through further mirror or shoulder checks. However, for the driver assistance system to decide whether a warning is necessary, it is often sufficient to determine the probability that the driver is currently pursuing a course of action in which this conflict arises. (See literature, in particular Holger Berndt and Klaus Dietmayer, “Driver intention inference with vehicle onboard sensors”, IEEE International Conference on Vehicular Electronics and Safety (ICVES), pages 102-107, 2009; or Georges S. Aoude, Vishnu R. Desaraju, Lauren H. Stephens, and Jonathan P.As described in “Behavior classification algorithms at intersections and validation using naturalistic data”, 2011 IEEE Intelligent Vehicles Symposium (IV), (IV): 601-606, June 2011, methods are known that determine the probability that a driver intends to perform a maneuver of a given type, such as a right turn or a lane change, within a specific foresight horizon. For this purpose, a classifier is trained for each addressed maneuver type, which can estimate the probability of occurrence of a corresponding maneuver based on a fixed set of features. It is known that a Naive Bayesian approach can be chosen for such a classifier.

[0003] An example of the use of a driver assistance system will now be given with reference to Fig.1. The driver of the vehicle under consideration, 1 (a car), is traveling in the right lane of a highway and is approaching a slower-moving truck, 2, in the same lane. There is a probability that the driver will want to change lanes within the next few seconds to overtake the truck. At the same time, a significantly faster motorcycle, 3, is approaching from behind. It is assumed that the driver will overlook the motorcyclist. Depending on the specific timing of the lane change, the following situations can be distinguished: 1. The lane change occurs early enough that the motorcyclist has sufficient time to react and brake (arrow 4). The situation is therefore not critical; a reaction from the driver assistance system is not desired or only desired in a reduced form. 2.The lane change occurs when the motorcyclist is still behind car 1, leaving insufficient time to react and brake (arrow 5). This situation is therefore extremely dangerous. Inadequately secured lane changes are the most frequent cause of accidents on German highways. In this case, a reaction from the driver assistance system is expressly desired. 3. The lane change only occurs after the motorcycle has passed the car (arrow 6). The situation is therefore not critical, and a reaction from the driver assistance system is not desired or only desired in a reduced form. A lane is also referred to as a driving lane.

[0004] To ensure the driver assistance system meets requirements as effectively as possible, the probability of occurrence for each of the aforementioned situations is needed. Typical features for predicting lane changes include the turn signal, the driver's lateral deviation from the lane center, and their mirror checks. However, depending on the installed sensors (lane marking detection camera, eye-tracking system) and their current data quality, only some of these features are available for prediction. A state-of-the-art solution would involve training a classifier during the development phase of the driver assistance system for each of the alternative lane change points and for possible combinations of the relevant features. This classifier would then output a numerical measure of the maneuver's probability of occurrence.

[0005] A disadvantage of existing approaches to predicting driver intent is that they are based on a fixed set of characteristics or types of driving-related actions, such as the vehicle's speed or acceleration, the use of a turn signal, etc. For each new combination of characteristics, resulting, for example, from the current availability of these characteristics, as well as for each type of maneuver and each driving situation, a new classifier must be trained. This results in significantly increased development and validation costs, as well as reduced predictive reliability if the responsible classifier cannot be clearly identified.

[0006] Another example of the use of a driver assistance system will be given below with reference to Fig.Given the following scenario, the driver of vehicle 1 (a car) is approaching an urban intersection in the middle lane. He realizes rather late that he intends to turn right. The lane change therefore only occurs immediately before the intersection. This lane-change maneuver, with its associated mirror checks and the need for simultaneous braking and steering, fully engages the driver's cognitive abilities. Consequently, there is a risk that he will overlook cyclist 7, who is traveling parallel to the road on the cycle path, resulting in a collision. The goal of a corresponding driver assistance system is to prevent this impending collision. This can be achieved by the assistance system initiating a braking maneuver at the last possible second.However, this requires a very high level of reliability in detecting cyclist 7 and predicting their subsequent trajectory, which is currently rarely the case. Instead, the driver of car 1 should be alerted to cyclist 7 so that they can handle the situation independently and confidently. This, however, necessitates that the assistance system recognizes the driver's intended action early on, i.e., in this case, even before they change lanes to the right. In this specific case, the driver's intended action is to change lanes and then turn right.

[0007] Existing approaches to driver intention detection are designed solely to identify the next driving maneuver. In this case, however, a right turn is not a possible first maneuver, as the vehicle is initially in the middle lane, from which a right turn is not possible. Furthermore, even the reliability of predicting the first maneuver, the lane change, using a pure lane-change classifier is compromised by the overlapping of driver behaviors resulting from the individual driving maneuvers.

[0008] This is easily understood by considering, for example, the method disclosed in German patent application DE 10 2006 040 537 A1. This method infers a driver's intention to change lanes based on certain steering movements in conjunction with the turn signal and a constant speed. In the scenario under consideration, the driver is forced to reduce their speed before and during the lane change due to their intention to turn. This contradicts the behavior assumed during the development of the lane-change classifier and can therefore easily lead to an incorrect detection of the driver's intention.

[0009] Furthermore, a disadvantage of existing approaches is that if several maneuvers occur in succession, the second and every subsequent maneuver in that sequence can no longer be predicted. Moreover, the overlapping of driver behaviors typical for the individual sub-maneuvers can lead to a situation where even the first maneuver can no longer be reliably predicted.

[0010] Patent application DE 10 2013 212 359 A1 describes a method for predicting the driving path of a vehicle, which includes: providing the current position of the vehicle; determining a group of future possible driving paths of the vehicle based on a digital map, the current position, and a predetermined maximum size of each driving path, in particular a maximum length of each driving path, a maximum driving time for each driving path, or a maximum number of driving maneuvers per route; providing measurements of predetermined types of driving-related actions of the vehicle's driver, in particular the speed or acceleration of the vehicle, the setting of a turn signal, and / or the point of view at which a driver looks in the vicinity of the vehicle;Provide, for each route of the group of routes, one or more models, each providing an occurrence measure for one or more types of driver actions, wherein the occurrence measure provides a measure of the probability that the measurements of the driving-related actions occur while driving the vehicle on the respective route; Determine, for each route and for each model of the respective occurrence measure; Determine, for each route from the group of routes, an execution measure, namely a measure of the probability that the corresponding route is driven, based on the occurrence measures determined for each route.

[0011] In practice, it can happen that the observed driver behavior, i.e., the measured actions of the driver, cannot be adequately explained by any of the considered hypotheses, i.e., driving paths from the group of possible future driving paths. In the Fig. In the situation depicted in 2a, the driver of vehicle 20, for example, accidentally activated the turn signal, or failed to reverse after changing lanes, even though he intended to continue straight ahead. It is assumed that he overlooked the cross traffic 21, which was partially obscured by trees 22.

[0012] Since the driver intends to drive straight ahead, there is no need for him to reduce his speed. Therefore, he continues driving at a constant speed of, for example, 50 km / h. For two exemplary hypotheses or driving paths, "straight ahead" and "turning right," a discrepancy arises between expected and observed driver behavior (or measured action) for each of the characteristics (turn signal, speed), i.e., the type of measured action. This is shown in the following exemplary overview in Table 1 for the procedure described in DE 10 2013 212 359 A1: Table 1 route Performance figure for model “Blinker” Performance indicator for model "Speed" Total occurrence score Execution factor Straight ahead 0,010 0,900 0,009 0,155 Turn right 0,980 0,050 0,049 0,845

[0013] In accordance with the procedure described in DE 10 2013 212 359 A1, the product of the occurrence scores of the models (total occurrence score) of the respective driving path is calculated as an intermediate result. The comparatively low value of the total occurrence score indicates that the observed driver behavior in both cases does not correspond particularly well with the expected driver behavior.

[0014] Assuming that the two maneuvers (driving paths) have the same probability beforehand, their respective execution scores are derived after observing the driver's behavior as the ratio between the respective total execution score and the sum of all total execution scores, which corresponds to a normalization.

[0015] In general, it can happen that none of the considered hypotheses (travel paths of the group of travel paths) satisfactorily explains the observed driver behavior. This is precisely the case in the example considered above: The ratio between the occurrence score for the "turn signal" model when driving straight ahead and the occurrence score for the "speed" model when turning right results in a large relative difference between the execution scores for the respective travel paths and thus a large discrepancy between the execution scores of the two hypotheses. This occurs even though none of the considered travel paths can satisfactorily explain the driver's behavior (or their actions).

[0016] Typically, a driver assistance system that warns of a hazard (such as a collision) bases the decision for or against issuing a warning on the probability of execution for a given driving path.

[0017] In the example scenario, this would have led to the assistance system refraining from warning the driver due to the high number of right-turn executions, even though the driver will actually be driving straight ahead, and a collision could result if cross traffic misinterprets the driver's turn signal.

[0018] Evaluating additional features or driver actions, such as the driver's gaze patterns, could reduce the risk of such failures. However, this would have no effect if the misjudgment stems from the driver performing a maneuver that is not accounted for in the existing environmental model or the digital map used, and is not represented by the possible future driving paths. Examples include turning into an uncharted driveway or performing a maneuver (activating the turn signal, reducing speed, etc.) contrary to traffic regulations. In such situations, the assistance system might warn the driver about a road user who is irrelevant to them due to the maneuver actually performed.

[0019] The object underlying the invention is to provide more meaningful metrics for the probability of executing driving maneuvers.

[0020] This problem is solved by the method, the apparatus, the vehicle, and the computer program according to the independent claims. Advantageous embodiments are defined in the dependent claims.

[0021] In its first aspect, a method for predicting a vehicle's trajectory comprises: providing the vehicle's current position; determining a set of future possible trajectories of the vehicle based on a digital map and optionally additionally the current position and / or a predetermined maximum size for each trajectory, in particular a maximum length of each trajectory, a maximum travel time for each trajectory, or a maximum number of maneuvers per trajectory; providing measurements of predetermined types of driving-related actions of the vehicle's driver, in particular the vehicle's speed or acceleration, the use of a turn signal, and / or the driver's gaze location in the vehicle's surroundings;Providing, for each driving path of the group of driving paths, one or more models, each providing an occurrence measure for one or more types of driver actions, wherein the occurrence measure provides a measure of the probability that the measurements of the driving-related actions occur while driving the vehicle on the respective driving path; Determining, for each driving path and for each model, the respective occurrence measure; Determining overall occurrence measures for each driving path from the group of driving paths, namely a measure based on the totality of the occurrence measures determined for the respective driving path, in particular the product of the occurrence measures of a respective driving path;Determining an error occurrence measure, wherein the error occurrence measure in particular has a predetermined value, wherein the error occurrence measure represents the probability that the set of possible routes is incomplete and / or faulty and in particular represents the totality of probabilities of traveling on such routes that are not included in the set of future possible routes, and takes into account the probability of actions by the driver that are contrary to the traffic rules;Determine, for each path in the group of paths, an outcome performance measure, namely a measure of the probability that the respective path is travelled, based on the total occurrence measure determined for the respective path and the total occurrence measures determined for all paths in the group of paths, including the determined error occurrence measure. The procedure can be supplemented by determining an error performance measure based on the error occurrence measure and the total occurrence measures determined for all paths in the group of paths, including the determined error occurrence measure.

[0022] Therefore, when calculating the outcome execution measure for each driving path, the error occurrence measure is also taken into account. This represents the probability that none of the driving paths in the group of driving paths (the hypotheses) corresponds to the driver's actions, thus resulting in an error in determining the future driving path. In such a case, as explained above, even marginal differences in the calculations of the occurrence measures would lead to incorrect conclusions regarding the probability of individual future driving maneuvers. The function of a driver assistance system is then based on the outcome execution measure.

[0023] The core problem with state-of-the-art methods is that they only identify possible future driving paths that are feasible based on a digital map and assumptions about driving behavior. The scenario where the driver performs a different maneuver than indicated on the map, or behaves differently than predicted by the model for the actual maneuver (behavior contrary to traffic regulations), is thus excluded from the outset. Instead, state-of-the-art methods calculate products of occurrence metrics (total occurrence metric) for each driving path in the group of possible future driving paths. These metrics represent the degree of agreement between observed and expected driver behavior according to the models. These products are then compared (normalized) to determine an execution metric.In this process, information is lost regarding whether the performance indicators or products are generally rather high or rather low, or whether the considered driving paths of the group (i.e., the hypotheses) provide a sufficiently good explanation for the observed driver behavior.

[0024] The magnitude of these occurrence metrics, or their products, or the total occurrence metrics, can not only be used to calculate the execution metric of the respective hypothesis, but also provides information about the quality of the prediction itself. If particularly low occurrence metrics or products result for all driving paths, this indicates that the current driver behavior cannot be explained by any of the maneuvers or maneuver combinations considered, but is determined by other, unobserved factors.

[0025] It is proposed that the aforementioned cases (route not included in the group of possible routes and / or driving behavior contrary to traffic rules) be explicitly considered in the form of an additional error hypothesis with a fixed a priori probability and a fixed occurrence measure (i.e., predetermined values) for each model.

[0026] In a training course, the same number of models are provided for each route; the error occurrence metric is determined depending on the number of models, in particular assuming a predefined value assigned to each model, and the error occurrence metric takes the respective value(s) into account.

[0027] The determination of the fault occurrence score thus takes into account how the occurrence scores for the driving paths are calculated, namely using models. The fault occurrence score is determined based on the number of models used. This corresponds to the calculation of the total occurrence scores, as these are calculated by multiplying the occurrence scores provided by the models. Depending on the number of models, this results in values ​​of varying absolute magnitudes. The fault occurrence score is adjusted accordingly. It is also possible to consider the model itself when determining the value of the fault occurrence score. In this case, different values ​​are assigned to different models (turn signal activation, speed, etc.).The error execution measure can then be calculated as the ratio of the error occurrence measure to the total occurrence measures determined for the paths of the group of paths, including the determined error occurrence measure.

[0028] In a training course, the respective result execution measure is calculated based on the ratio of the total occurrence measure for the respective route to the total occurrence measure for all routes in the group of routes, including the error occurrence measure.

[0029] In a preferred further development, the procedure further comprises: determining one or more hazard paths, namely those paths for which a warning of a collision hazard is to be issued, in particular based on the evaluation of the vehicle's surroundings by a driver assistance system; determining the hazard-outcome execution score for the hazard paths based on the total occurrence scores of the hazard paths, including the error occurrence score, in relation to the total occurrence scores for all paths in the group of paths, including the error occurrence score; and issuing the warning depending on the hazard-outcome execution score. In one calculation method of this specification, the outcome execution scores of the hazard paths and the error execution score are added together.

[0030] Advanced driver assistance systems monitor the vehicle's surroundings and determine which paths within the group of paths pose a collision risk (hazard paths). If the execution score for the hazard path(s) – i.e., a measure of the probability that a particular path will be followed – exceeds a threshold, a hazard warning is issued. It may also be possible to issue a warning as a precautionary measure if the actual path is not included in the group of paths or if the driver is acting contrary to traffic regulations. For this purpose, the error occurrence score is considered when calculating the hazard outcome execution score. Thus, the probability of errors in determining the path is factored in as a safety precaution. This increases road safety.

[0031] In an advantageous implementation, the procedure further includes: determining, by the driver assistance system, that the collision risk justifies considering the probability that the vehicle will not follow any of the paths within the group of possible future paths, and that the driver will act contrary to traffic regulations; where the determination of the hazard-outcome execution metric is performed in response to the determination by the driver assistance system. Thus, the consideration of the error occurrence metric is decided based on the severity or potential damage resulting from the hazard. In non-critical cases, an unnecessary warning can therefore be avoided.

[0032] In another aspect, a device comprises electronic computing means, wherein the device is configured to perform one of the aforementioned methods. A motor vehicle may comprise the device. The computing means of the device are operationally coupled with the necessary sensors and driver assistance systems. The electronic computing means may be a computer, which in turn may be a microcontroller, an ASIC, or the like.

[0033] In yet another aspect, a computer program causes a computer to execute one of the aforementioned procedures.

[0034] The procedure can also be carried out in combination with the following supplementary procedures. Additional features in the supplementary procedures complement the features of the procedures described above. 1. Additional procedure, comprehensive: Determine, for each driving path of the group of driving paths, the driving maneuvers to be performed in the respective driving path; Providing predetermined probabilities for the execution of a driving maneuver; Provide, for each path of the group of paths, a standard measure, namely a measure of the probability of executing the respective path of the group of paths, based on the predetermined probabilities for the driving maneuvers to be performed in the respective path; Determining the result performance measure also takes into account the standard measure for the corresponding route. 2. Additional procedure following additional procedure 1, furthermore comprehensive: Providing information about the lanes available on the roadway, as indicated by the current position of the vehicle; Determine, for each lane, a lane measure for the probability that the vehicle is in that lane, based on the determination of its current position; Whereby the determination of the standard dimension takes into account the lane dimension, in particular by multiplying it by the lane dimension. 3. Additional procedure following one of the preceding additional procedures, furthermore comprising: For each driving path in the group of driving paths: Determine the driving maneuvers to be performed on the corresponding driving path; Whereby determining one or more models for each driving path takes into account the driving maneuvers to be performed on the respective driving path. 4. Additional procedure following one of the preceding additional procedures, comprising the provision of one or more models: Providing each model in at least two variants, with each variant taking into account a different driving style from a given group of driving styles. 5. Additional procedure following one of the preceding additional procedures, furthermore comprising: Determining the position and movement of other road users or obstacles, in particular using environmental sensing; Determining those travel paths that could lead to a collision with other road users or obstacles; Determining a collision measure that indicates the probability of a collision with another road user or an obstacle, taking into account the outcome execution measures of the paths that could lead to a collision.

[0035] The following terms are preferred here: Occurrence measure: f(O i |H x ); Total occurrence rate: G(O i |H x ); Failure occurrence score: FE; Execution score: P a-posteriori (H x |O); Result execution measure: E - P a-posteriori (H x |O); Failure Execution Score: FE - P a-posteriori (H x |O); Standard measure: P a-priori (Hx ); Route: Trajectory: H x .

[0036] An alternative method incorporates the magnitude of the occurrence metrics themselves into the decision of whether, in a given situation with specific calculated probabilities or execution metrics for the hypotheses under consideration, the assistance system should react, and if so, in what form. For example, the magnitude of the highest overall metric (occurrence metric or its product) can be used as an indicator of the prediction quality. BRIEF DESCRIPTION OF THE DRAWINGS Fig. Figure 1 schematically shows a driving situation in which the use of a driver assistance system may be useful. Fig. Figure 2 schematically shows another driving situation in which the use of a driver assistance system may be useful. Fig.Figure 2a schematically shows a driving situation in which the prediction of the driving path may be incorrect due to the driver's behavior not complying with the traffic rules. Fig. Figure 3 shows a flowchart of the method according to the invention in an exemplary embodiment. Fig. Figure 4 schematically shows a segmented section of a digital map according to an exemplary embodiment. Fig. 5 shows one for Fig. 4 extracted tree graphs for determining driving paths for the starting point Q1 according to an exemplary implementation. Fig. Figure 6 schematically shows examples of the progression of calculation bases for occurrence parameters provided by a model for a direction indicator. Fig. Figure 7 shows an example of the progression of appearance parameters provided by a model for a direction indicator. Fig.Figure 8 illustrates exemplary viewpoints of a driver that are relevant for viewing a section of a digital map as in Fig. 1 were shown and calculated.

[0037] Identical reference symbols refer to corresponding elements across the figures. DETAILED DESCRIPTION OF THE EXECUTION EXAMPLES

[0038] Fig.Figure 3 shows a flowchart of the method according to an exemplary embodiment. The features of the additional method are optional for the method according to the invention. The method begins with determining the current position of the vehicle, for example, using a GPS and / or GLONASS receiver. Using the determined position, the vehicle can then be located on a digital map. For the purposes of the invention, a high-precision digital map is advantageously used, i.e., a digital geographic map that records individual lanes and has an accuracy in the range of 10 cm to 50 cm or up to 1 m. Since positioning using typical GPS and / or GLONASS receivers is often not precise enough to reliably assign the position to a specific lane, it is provided that a lane measure is output for each lane, representing the probability with which the vehicle is located in that lane.

[0039] Next, a group of possible future routes for the vehicle is determined. This is done using the digital map. To determine these possible future routes, the lanes on the digital map are divided into segments; in other words, the lanes are discretized. Fig.Figure 4 schematically shows a section of a high-precision digital map. The section depicts a roadway with two lanes traveling in the same direction and a junction approximately in the middle of the section. The two lanes are divided by the four segments S1, S2, S4, and S5. The junction is represented by segment S3. To further determine the possible routes, traffic management guidelines are considered, such as the designation of a lane as a turning lane or the prohibition of changing lanes from one lane to another. It is also assumed that only one lane change occurs within a segment, or that only one lane change occurs every 500 m. Typically, a segment of a straight-ahead lane is 5, 10, 50, 200, 300, or 500 m long. Junctions are taken into account during the discretization process.In the present example, the vehicle was positioned on one of the two lanes at the points Q1 and Q2 indicated by the circles. The probability P was calculated for each point Q1 and Q2. a- priori (Q1) and P a-priori (Q2) determines that the vehicle is located at it. This probability is the lane measure. In some implementations, an intermediate result from the probability calculation is also used as the lane measure.

[0040] For each location Q1 and Q2, the available routes for the vehicle are determined. These routes are extracted only up to a predefined maximum length, for example, 200 m or 1 km. The result is displayed as a tree diagram for each location (starting point). Fig. 5 shows the one for Fig.4. Extracted tree graph for the starting point Q1 (represented by the thick outline of box S1). Lane changes are shown by dashed arrows. Each path within the tree starting from S1 to one of the leaves of the graph represents a lane, and all lanes together form the group of lanes for Q1. A possible lane can also be called a hypothesis.

[0041] In the next step, each route is assigned a standard measure, in this case the a priori probability that the respective route will be traveled. This assignment is made without knowledge of the driver's actions relevant to the journey. The maneuvers to be performed on the respective route can be taken into account, along with statistical evaluations of the probability of each maneuver being executed. The probability of each route is calculated using the following typical rules: The initially determined probability (in this case, P) a-priori (Q1) or P a-priori(Q2)) is distributed equally among the respective child nodes (subsequent nodes) if the maneuver to be performed is a turn. This probability is further distributed equally among subsequent child nodes, if any, and provided a turn is performed. If a lane change is possible from a node, this node transfers part of its probability to the node to which the lane change leads, for example, 1 / 3. The probability for a path then results from the probability of the leaf node (the last node). For example, the probability P for the path S1-S4-S5 is... a-priori (S1-S4-S5) = P a-priori (Q1) * 2 / 3 / 2 / 3, assuming that the probability of a lane change is 1 / 3. With H x The respective routes are described below, see Fig. 5.

[0042] In the next step, one or more models for the driver's driving-related actions are provided. Based on measurements of these actions, the models calculate probabilities for the occurrence of these actions along a given driving path. The models can be parameterized using the high-precision digital map, for example, the curvature of a curve or straightaway, distances to intersections, or similar parameters. To do this, the properties or parameters of each driving path within the group of driving paths can first be extracted from the digital map. An example of such a model is determining the probability (or occurrence metric) of a turn signal being activated, depending on the distance to a turning point along a driving path. Detailed examples of these models are given below.

[0043] Simultaneously, measurements of the driver's driving-related actions are provided. These measurements concern the current status of a turn signal and, if applicable, the time since its last activation (referred to as Measurement I), the speed profile over the last 1.4 s (referred to as Measurement V), and the driver's gaze direction or point of view within the last 1 s (referred to as Measurement G). The combination of these measurements is called Observation O, O ≈ [I, V, G]. A model is provided for each measurement of I, V, and G. Generally, it is also possible to provide a single model for the entire Observation O, or for subcombinations of the observations, for example, for I and G.

[0044] Using the models and the measured driving-related actions, the occurrence metrics (or just a single occurrence metric) are calculated for each driving path. In this embodiment, the occurrence metric is determined by the likelihood function f(O). n |H) is calculated. For example, if a model combines the measurement for the turn signal I for the first driving path H1 into a probability, this results in f(O1|H1). The measurements of G contained in observation O1 are disregarded by this combination. A model adapted to driving path H2, which takes the turn signal into account, provides the probability f(O1|H2). Intermediate results of the calculation of f(O) can also be used to calculate the occurrence measure. i |H x ) are used, which represent the ratio of all occurrence probabilities to be calculated.

[0045] Furthermore, the error occurrence score (FE) is calculated. This calculation considers the number of models and the model itself used for each driving path, with the same models being used for each path. Predefined values ​​(as occurrence scores) are assumed for the models to calculate the error occurrence score. For example, a value of 0.3 can be assigned to a model for the turn signal and also to a model for speed. The values ​​for the individual models can be obtained through empirical evaluations and / or consideration of typical value ranges for the occurrence scores of the models for correct detection or non-detection of the future driving path. The error occurrence score (FE) is then calculated as the multiplication of these two values, resulting in FE = 0.09.

[0046] Finally, the result performance measure is calculated, which in the present embodiment is calculated as follows: E−Pa−posteriori(Hx|O)=∏if(Oi|Hx)P(O)+FEPa−priori(Hx), where the total occurrence factor G = ∏ i f(O i |H x ). This measure is referred to here as a posterior probability, as it takes into account the driver's driving-related actions.

[0047] Furthermore, the error execution score is calculated: FE−Pa−posteriori(Hx|O)=FEP(O)+FEPa−priori(error),

[0048] For the a priori probabilities P a-priori In some implementations, a uniform distribution is assumed for all paths and for the error. For P a-priori (Error) can also be the mean of the a priori probabilities determined for the routes or predetermined values.

[0049] Table 2 below describes the situation of Fig. 2a an example with uniformly distributed P a-priori given, which is also based on the example in Table 1. Table 2 route Performance figure for model “Blinker” Performance metric for model “speed” Total occurrence rate Result - Execution Measure Straight ahead 0,010 0,900 0,009 0,061 Turn right 0,980 0,050 0,049 0,331 Value for model “Turn Signal” Value for model “Speed” Error occurrence rate Error execution rate Mistake 0,300 0,300 0,090 0,608

[0050] As can be seen, taking the error and its occurrence metric into account reduces the result execution metric compared to the execution metric for the two driving paths. The result metrics reflect the reduced probability of executing the respective maneuver, and the relatively high value of the error execution metric indicates that the prediction of the driving paths is subject to significant errors, which in this case result from the driver's traffic violation.

[0051] Depending on the situation, the assistance system could then decide, based on these modified metrics, whether a reaction is necessary. In the particularly critical situation considered in this example, it would be conceivable, for instance, to issue a warning both when driving straight ahead is recognized as such and when the intended maneuver cannot be unambiguously identified. Thus, the result execution metric for driving straight ahead and the error execution metric must be added together. The result, 0.669, clearly indicates a decision in favor of issuing a warning.

[0052] For comparison: If the driver wanted to turn right and therefore adjusted his speed accordingly, the following figures and probabilities would result, for example, as shown in Table 3: Table 3 route Performance figure for model “Blinker” Performance metric for model “speed” Total appearance measure number Result - Execution Measure Straight ahead 0,010 0,050 0,001 0,001 Turn right 0,980 0,900 0,882 0,906 Value for model “Turn Signal” Value for model “Speed” Error occurrence rate Error execution rate Mistake 0,300 0,300 0,090 0,092

[0053] Since the driver's behavior now complies with the traffic rules and the driver's future route has been correctly included in the prediction, there is no reason for error, resulting in a relatively small value for the error execution score and a high value for the result execution score of the right turn, which is a clear indication of the driver's intention and future route. Direction indicator

[0054] A first model that can be used to calculate the execution metric is a model for the activation of the direction indicator, in particular the turn signal of a passenger car. To use this model, the driving maneuver(s) to be performed on a given route are first determined.

[0055] For driving maneuvers that the model can predict, the corresponding model parameters for the respective driving path and maneuver are extracted from the digital map. The model is parameterized for each individual maneuver. Therefore, multiple differently parameterized models exist on a single driving path if several maneuvers are performed. However, it is also conceivable that a combination of driving maneuvers is represented by a single model, which is then parameterized for that combination.

[0056] In this embodiment, the model makes a prediction for the following driving maneuvers: driving straight ahead, turning right, turning left, changing lanes to the right, and changing lanes to the left. In this embodiment, the model also depends on the status of the turn signal at the time of evaluation. S State of the direction indicator (deactivated, activated to the right, activated to the left).

[0057] The following table details the model used for activating the turn signal according to the exemplary embodiment and provides formulas for calculating the occurrence values. In this exemplary embodiment, the occurrence value corresponds to the corresponding probability. It is also conceivable to use an occurrence value based on an intermediate result of the probability calculation, which still allows the driving maneuvers to be related to each other in a way that corresponds to the probabilities.

[0058] If several driving maneuvers are to be performed on a driving path, the occurrence measure f(O) results. i |H x ) only from the next driving maneuver to be executed (from s C(from), because it can be assumed that the driver only activates the turn signal according to the next maneuver to be performed. Therefore, only one model is used for each driving path. Driving maneuvers S = Left S = Disabled S = Right Drive straight ahead f R (s A , s C ) 1 - 2P R f R (s A , s C ) Turn right f R (s A , s C ) 1 - 2P R - F T (s C , s T ) f T (s A , s T ) +f R (s A , s C ) Turn left f R (s A , s C ) +f T (s A , s T ) 1 - 2P R - F T (s C , s T ) f R (s A , s C ) Change lanes to the right f R (s A , s C ) 1 - 2P R -F L (s C , s0, s1) f L (s A , s0, s1) +f R (s A , s C ) Change lanes to the left f R (s A , s C ) +f L (s A , s0, s1) 1 - 2P R - F L (s C , s0, s1) f R (s A , s C )

[0059] Depending on the model to be determined, the required parameters (formula letters in parentheses) are extracted from the digital map. These mean: s C Distance of the current position to a reference point on the digital map. This reference point can be arbitrarily placed, but is advantageously positioned a few hundred meters in front of the current position. s A Distance to the location where the direction indicator was set to its current state, if it was activated. s T Distance to the fork of the turn s0 Start of the lane change segment (as distance from the reference point), or, if s0 < s C , the current distance sC s1 The end of the lane change segment (as distance from the reference point)

[0060] The probabilities are calculated as follows: fR(s, sC)=PR λ eλ(s−sC)

[0061] Where λ=−log(1−P10) / Δs

[0062] Typical values ​​for P R are 0.015; 0.02; 0.025. Typical values ​​for P 10 are 1 / 150; 1 / 200; 1 / 250. Typical values ​​for Δs are 0.5 m; 1 m; 1.5 m. fT(s, sT)=PT qT2π σTexp[−12(s−sT−μTσT)2]

[0063] Where P T For example, 0.78 µ T = -55.6 m and σ T = 25.3m and qT−1=∫−∞sTfT(s, sT)PT qTds=12[1+erf(−μT2σT)] erf(x)2π∫0xe−t2dt FT(s, sT)=PT qT2[1+erf(s−sT−μT2σT)] fL(s, s0, s1)=PL qL2(s1−s0)erf(s−s'−μL2σL)|s'=s1s'=s0 with P L typically 0.68 µ L = v C ∗ -2.83 s, σL = v C ∗ 0.61 s and erf(...) as well as q L as above (v C (is the current speed). FL(s, s0, s1)= PL qL2[1+2σLs1−s0H(s−s'−μT2σL)|s'=s1s'=s0] with the indefinite integral: H(x)=∫erf(x)dx=x erf(x)+1πe−x2+C.

[0064] Fig. Figure 6 schematically shows examples of the progression of the individual calculation elements f shown above. R , f T and f L .

[0065] Fig. Figure 7 shows an example of the behavior of the activation threshold values ​​for different turn signal models. In the diagram, the dotted curve represents the logarithm of the activation threshold value for a turning maneuver, the dashed curve represents the logarithm of the activation threshold value for a lane change, and the solid curve represents the logarithm of the activation threshold value for driving straight ahead. View direction

[0066] Another model that can be used to calculate the performance index is a model for the driver's gaze point. The driver's gaze point is the point in the environment that the driver views. This can be derived from the driver's gaze direction. Instead of the precise gaze direction, i.e., the orientation of the driver's pupils, the orientation of the head can also be used. Alternatively, for example, the point tangent to the inner lane marking of the curve, which is closest to the driver's actual gaze point, can also be used as the gaze point in the calculation. Systems for determining the gaze direction and head orientation are known in the art.

[0067] The model is parameterized for each route using the digital map. The model assumes a driver's viewpoint, which, according to the model, lies a distance Δs = a0 + a1v. Cin front of the vehicle, where a0 is typically 6 m, a1 is typically 1 s and v C The current speed of the vehicle. Typical viewpoints of a driver, determined by the model, are in Fig. 8 represented by stars. Fig. Figure 8 represents the same section of the segmented digital map as Fig. 4. In Fig. For the path that turns, the star on the far right has been calculated. For the other hypotheses of driving straight ahead, either in the left or right lane, the viewpoints represented by the two closely spaced stars have been calculated. In the Fig. 8 The driver looks at the viewpoint that was calculated for the turning maneuver.

[0068] For each specific model viewpoint, an expected viewing direction can be defined. φ^k(h) derive, where k denotes a time step k and (h) denotes the hypothesis, i.e. the respective route.

[0069] The viewing directions φ are k The driver's gaze direction was measured at time points k. The deviation between the measured gaze direction and the gaze direction determined by the model for a driving path (h) is given by: Δφk(h)=φk−φ^k(h).

[0070] The calculation of the occurrence metrics of this model corresponds to the actual probabilities f(O). i |H x ) = f Gaze However, it may also be possible to use intermediate results from the probability calculations that reflect the ratio of the individual probabilities. Gaze For hypothesis h, the following is calculated: fGaze(h)=exp[1NGaze∑i=0NGaze−1log(fΦ(φk−i−φ^k−i(h)))]

[0071] Where N Gaze = 10 is and fΦ(Δφk(h))=1−PΦ2π σΦexp[−12(Δφk(h)σΦ)2]+PΦ02π,

[0072] A typical value for P Φ0 is 0.11 and for σ Φ 0.13. Vehicle speed

[0073] Another model that can be used to calculate the execution score is a model for the vehicle's speed, particularly when approaching intersections. Using the method described in patent application DE 10 2013 200 724 A1 or patent application DE 10 2013 203 908 A1, incident scores can also be determined by observing the vehicle's speed. These incident scores can then be taken into account when determining the execution score.

Claims

[1] Method for predicting the driving paths of a vehicle, comprising: Providing the current position of the vehicle; Determining a group of future possible routes of the vehicle using a digital map and optionally also the current position and / or a specified maximum size of each route, in particular a maximum length of each route, a maximum travel time for each route, or a maximum number of driving maneuvers per route; Providing measurements of predetermined types of driving-related actions of the vehicle's driver, in particular the speed or acceleration of the vehicle, the setting of a turn signal, and / or the point of view to which a driver looks in the vicinity of the vehicle; Providing, for each driving path of the group of driving paths, one or more models, each providing an occurrence measure for one or more types of driver actions, wherein the occurrence measure provides a measure of the probability that the measurements of the driving-related actions occur while driving the vehicle on the respective driving path; Determine the respective impact factor for each route and for each model; Determining total occurrence metrics for each route from the group of routes, namely a metric based on the totality of occurrence metrics determined for the respective route, in particular the product of the occurrence metrics of a respective route; Determining an error occurrence measure, wherein the error occurrence measure in particular has a predetermined value, wherein the error occurrence measure represents the probability that the set of possible routes is incomplete and / or faulty and in particular represents the totality of probabilities of traveling on such routes that are not included in the set of future possible routes, and takes into account the probability of actions by the driver that are contrary to the traffic rules; Determine, for each route from the group of routes, an outcome performance measure, namely a measure for the probability that the respective route is travelled, based on the total occurrence measure determined for the respective route and the total occurrence measures determined for the routes of the group of routes, including the determined error occurrence measure. [2] Method according to claim 1, where the same number of models are provided for each route; where the error occurrence rate is determined depending on the number of models, in particular, a predefined value is assumed for each model, and the error occurrence measure takes the respective value(s) into account. [3] Method according to claim 2, wherein the respective result execution measure is calculated on the ratio of the total occurrence measure for the respective path to the total occurrence measure for all paths of the group of paths including the error occurrence measure. [4] Method according to any one of the preceding claims, further comprising: Determining a fault execution measure based on the fault occurrence measure and the total occurrence measures determined for the paths of the group of paths, including the determined fault occurrence measure. [5] Method according to any one of the preceding claims, further comprising: Determining one or more hazardous driving paths, namely those driving paths for which a warning of a collision risk is to be issued, in particular based on the evaluation of the vehicle's surroundings by a driver assistance system; Determining the hazard-outcome execution score for the hazard pathways based on the total occurrence scores of the hazard pathways, including the fault occurrence score, in relation to the total occurrence scores for all pathways of the group of pathways, including the fault occurrence score; The warning is issued depending on the hazard-result-execution score. [6] The method of claim 5, further comprising: Determine, in particular through the driver assistance system, that the risk of collision justifies taking into account the probability that the vehicle will not travel on any of the paths of the group of possible future paths, and that the driver is performing actions contrary to the traffic rules; The determination of the hazard-outcome-execution measure is carried out in response to the determination by the driver assistance system. [7] Device comprising electronic computing means, wherein the device is configured to perform one of the foregoing methods. [8] Motor vehicle comprising a device according to claim 7. [9] Computer program which, when executed, causes a computer to perform a method according to any one of claims 1 to 6.

Images