DE102018222542A1 - Motion prediction for controlled objects - Google Patents

Abstract

Method (100) for predicting the trajectory (1a) of at least one controlled object (1) with the steps: • a current position (1b) of the object (1) determined by physical measurement and a current speed (1c) determined by physical measurement Object are provided (110); • a probable target (2) of the movement of the object (1) is provided (120); • using physical observations (3) of the object (1) and / or the environment in which if the object (1) moves, at least one probable preference (4) is determined (130), which is tracked when the object (1) is controlled in the direction of the probable target (2); • using at least the current position ( 1b) and speed (1c) of the object in combination with the target (2) and the at least one preference (4), a cost function (5) is determined (140), which is in a spatial area (6) which shows the current position (1b ) of the object (1) and the expected destination (2) assigns to each location (6a) that can be reached for the object (1) a cost value per unit of time spent there and / or distance traveled there per unit; • at least one of the current position (1b) of the A spatial and / or temporal integral of the cost function (5) along this candidate trajectory (1a ') is assigned to the object (1) leading to the likely destination (2) as total costs (5') (150) .

Description

The present invention relates to the prediction of the movement of objects, for example with the aim of coordinating the behavior of one's own vehicle with this movement.

State of the art

Traffic accidents in which vehicles hit pedestrians or other weaker road users often lead to serious or even fatal injuries. A collision often occurs because pedestrians are not only less visible than vehicles, but can also change their behavior so suddenly that there is no time left to react. This is also a challenge for at least partially automated vehicles, which are expected to move much more safely in traffic than human drivers.
( V. Karasev, "Intent-aware long-term prediction of pedestrian motion", 2016 IEEE International Conference on Robotics and Automation (ICRA), ISBN: 978-1-4673-8026-3 , doi: 10.1109 / ICRA.2016.7487409) proposes to model the movement of a pedestrian as a Markov decision-making process. The path of a pedestrian to an expected destination is assumed to be a stochastic process, in which the pedestrian tries to minimize the cost function determined by the current traffic situation. This cost function rewards, for example, the use of pedestrian crossings to cross lanes and, if there are traffic lights, only to cross them when the light is green.

Disclosure of the invention

Within the scope of the invention, a method for predicting the trajectory of at least one controlled object was developed.

A “controlled object” is understood to mean any object whose movement is influenced by one's own or someone else's drive in order to achieve a goal. In addition to people and animals, these can also be other vehicles or mobile robots, for example, which can be controlled manually, partly automatically or also completely automatically towards the target.

In the method, a current position of the object determined by physical measurement is provided. Optionally, a current speed of the object determined by physical measurement can also be provided. The speed can then also be taken into account wherever the position is processed. Any measurement methods can be used for the physical measurement, such as optical observation, radar or LIDAR. The physical measurement also does not necessarily have to be performed by the entity that predicts the object's trajectory. For example, a position and speed that a third-party vehicle outputs via a Vehicle-to-Infrastructure (V2I) interface or via a Vehicle-to-Vehicle (V2V) interface can be used by a separate, following vehicle to track the trajectory to predict the foreign vehicle.

At least one prospective target of moving the object is provided. This information can also come from any source. For example, any physical observation of the object can be used to determine the likely target. A third-party vehicle can also announce its destination via a V2I or V2V interface, for example.

Using physical observations of the object and / or the environment in which the object is moving, at least one prospective preference is determined, which is pursued when the object is steered towards the at least one prospective target. This is based on the insight that there are usually a large number of trajectories of the object that lead to the probable destination, and that these trajectories are nevertheless not equivalent from the point of view of this object or its owner. Rather, the trajectory specifically selected for controlling the object is generally determined by at least one preference. Such a preference can be, for example, to reach the goal by the shortest route or as quickly as possible. For a pedestrian, however, it can also be a preference, for example, to cross roadways as safely as possible, to have to overcome as few obstacles as possible with a stroller or shopping trolley, or to be exposed to the rain for only a short time. For example, for a vehicle or a robot, it may be a preference that wear is minimized and / or that the range is maximized with one tank fill or battery charge.

A cost function is determined using at least the current position of the object in combination with the at least one prospective target and the at least one prospective preference. In a spatial area which contains the current position of the object and the at least one expected destination, this cost function arranges a cost value per unit spent there for the object Time, and / or distance traveled per unit there.

An evaluation of the cost function along this candidate trajectory is assigned as a total cost to at least one candidate trajectory leading from the current position of the object to the at least one prospective target. This evaluation can be, for example, a spatial and / or temporal integral of the cost function. As an alternative or in combination with this, at least one distribution of candidate trajectories can be assigned an expected value of such an evaluation via the candidate trajectories in the distribution as total costs. In the distribution, the candidate trajectories can in particular be weighted according to the probabilities with which the object tracks these candidate trajectories.

These total costs can be used to order various candidate trajectories leading from the current position of the object to the expected destination in the order of the probabilities with which the controlled object will continue on the respective candidate trajectory. This probability is lower for a candidate trajectory with high total costs than for a candidate trajectory with higher total costs.

In particular, in a particularly advantageous embodiment, a candidate trajectory with the lowest total costs is selected from a plurality of candidate trajectories as the expected trajectory of the object. The indefinite article "a" expresses that there can be multiple candidate trajectories with the same minimum total cost. For the determination of the candidate trajectory with the lowest total costs, methods can be used that only deliver this optimal candidate trajectory itself, but not the concrete value of the lowest total costs.

In a further particularly advantageous embodiment, a distribution of candidate trajectories, the probabilities of which are weighted on the basis of their respective total costs, is output as a prediction of the trajectory. This leaves full flexibility and accuracy open for further evaluation, for example for determining the next sensible action with which a vehicle should react to the movement of the object.

The time horizon over which the trajectory is predicted can be freely selected. This time horizon can depend in particular on the intended use of the prediction, which in turn determines the accuracy of the prediction required. Typically, the accuracy decreases as the time horizon increases.

The total costs, and in particular also the optimal candidate trajectory with the lowest total costs, can be determined in any way, depending on the accuracy of the prediction required, the speed with which this accuracy is required and the computing capacity available. For example, the Markov model for the behavior of the object can be expanded using the cost function determined on the basis of the preference. However, the optimal candidate trajectory can also be obtained, for example, by successively testing a large number of candidate trajectories or, for example, using a machine learning method.

It is crucial that the expected preference is included in the determination of the cost function. It was recognized that in many cases this makes the prediction of the trajectory significantly more accurate. Ultimately, the statement that the stay of an object in a certain location is more or less advantageous is always based on some preference, since there is no generally applicable definition of "advantageous". The better the cost function reflects the preference underlying the control of the object, the more accurate the prediction of the movement of the object based on it.

For example, a sensible pedestrian may prefer to strictly adhere to the traffic regulations. For example, a cost function that reflects this preference always forces a pedestrian crossing to be used to cross a lane, even if this means a detour, whenever a pedestrian crossing is in sight.

On the other hand, for example, a hurried pedestrian may have the preference of reaching their destination as quickly as possible. If, for example, he sees a bus or a tram that he threatens to miss, then the road traffic regulations move into the background and he is only fixated on the preference to still reach the bus or the tram. This can happen especially if it is the last bus or the last train of the evening and there is a risk of a long walk or expensive taxi costs.

In such situations, the prediction of the movement quickly becomes incorrect if a standard cost function that is independent of the preference is used for this. Since the vast majority of pedestrians abide by the traffic regulations, such a standard cost function assumes that it complies with the rules, at least for the most part Behavior. If the pedestrian has entered the lane at an unintended or even dangerous point to cross it, the said standard cost function can, in extreme cases, steer the prediction of the movement in the direction that the pedestrian immediately returns to the sidewalk and on the Makes way to the next approved pedestrian crossing. Such a sudden change of heart can rarely be observed in reality.

The cost function can be selected, for example, from a predetermined catalog of cost functions. In the latter example, for example, the forecast can be significantly improved with a catalog of only two different cost functions. For example, a first cost function for sensible pedestrians, the use of which is mandatory, can assign significantly lower cost values than other areas of the road that pedestrians are not allowed to enter. A second cost function for hurried pedestrians who do not care about pedestrian crossings, on the other hand, does not distinguish pedestrian crossings from other areas of the road.

However, the cost function can also be parameterized, for example, with at least one variable that is used for its determination. In this way, cost functions for many different preferences can be summarized in a clear manner, and if the value of the size used changes, smooth transitions can be implemented.

In a particularly advantageous embodiment, on the basis of a semantic segmentation of the spatial area, each location is assigned at least one contribution to the cost function, which depends on its semantic meaning. For example, if a road is segmented into the carriageway and sidewalk, the cost function for a pedestrian may prefer walking on the sidewalk, while the cost function for a vehicle prefers driving on the carriageway.

According to what has been explained above using the example of a pedestrian, at least one contribution to the cost function is allocated to each location, which depends on the extent to which the presence of the object at this location is permitted in accordance with predetermined traffic regulations. This contribution can also be time-dependent. For example, the validity of traffic signs can be limited to certain periods of time, and driving over a traffic light is permitted or prohibited depending on the traffic light phase.

In a further particularly advantageous embodiment, each location is assigned at least one contribution to the cost function, which depends on how likely the collision of the object with other objects and / or another event potentially harmful to the object is at this location. For example, for a pedestrian, crossing a freeway can cost significantly more than crossing a normal street in an urban area. For a vehicle, staying in the opposite lane (for example, to overtake) in a confusing place where the vehicle may not be noticed and rammed in time by oncoming traffic can be subject to higher costs than in other places.

According to what has been described above, in a particularly advantageous embodiment, a pedestrian is selected as the controlled object. Pedestrians in particular can change their preferences particularly spontaneously and implement them with a very short delay in changing their movements.

In a further particularly advantageous embodiment, in response to the finding that the pedestrian has entered a lane or is about to enter such a lane, it is concluded that the pedestrian prefers to cross the lane starting from the point of entry. The cost function can then be adjusted accordingly. As explained above, this can counteract the tendency that an incorrect reversal of the pedestrian back onto the sidewalk is incorrectly predicted.

In a further particularly advantageous embodiment, depending on the weather and / or depending on a loading condition of the pedestrian determined by physical observation, it is concluded that the pedestrian has a preference for minimizing the distance to the expected destination. For example, if it rains, pedestrians who normally use the prescribed pedestrian crossings may be tempted to deviate from this and cross the street by the shortest route. The same applies if the pedestrian carries heavy luggage, for example.

In a further particularly advantageous embodiment, in response to the finding that the pedestrian is separated from another pedestrian by a lane and reacts to the presence of this other pedestrian, a conclusion is drawn to a preference of the pedestrian to cross the lane. The reaction to the presence of the other pedestrian can be, for example, that the pedestrian stops or changes his direction of movement. Children, in particular, can be distracted from the rest of the traffic by a familiar person who perceives them on the other side of the road in such a way that they are over walk the street without first looking left or right.

In a further particularly advantageous embodiment, starting from a plurality of candidate positions adjacent to the current position of the object, a trajectory leading to the expected destination is obtained with minimal total costs, i.e. with a minimal integral of the cost function. The value of this minimum total cost is assigned to the respective candidate position as a cost value. A trajectory of the object, which leads from the current position of the object to a candidate position with the lowest cost value, is selected as the expected trajectory.

In this way, the determination of the probable trajectory can be accelerated by pre-calculation. For a certain expected target of the object, for example, starting from every possible starting position of the object, trajectories with a minimum spatial and / or temporal integral of the cost function on the way from this starting position to the target can be calculated in advance. The candidate positions can then be checked by simply calling up the pre-calculated integrals, from where the path to the goal can be covered with the lowest total costs.

In particular, the expected trajectory can be calculated iteratively. For example, in response to the fact that a trajectory leading over a certain candidate position has been selected as the expected trajectory, only the step to this candidate position can be recorded. The said candidate position can then serve as a starting point for a new calculation. This means that, based on this, neighboring candidate positions can in turn be determined and evaluated with regard to the total costs for reaching the goal. By lining up such iterations in a runtime that grows linearly with the length of the path, a complete predicted trajectory is created from the current position of the object to its expected destination.

As explained above, the starting point for determining the preference, and thus also the applicable cost function, is a probable goal of the movement of the object. In order to determine this probable target, several candidate targets for the movement of the object can be provided, for example. A measure for the probability and / or confidence that the respective candidate target is actually the target of the movement of the object can then be determined for each candidate target using physical observations of the object and / or the environment. The candidate target for which the greatest probability and / or confidence has been determined can then, for example, be evaluated as the expected target of the movement of the object.

The candidate goals do not necessarily have to be discrete. Rather, the candidate goals can also be in the form of a distribution. In such a distribution, for example, each candidate target can be weighted with a probability and / or confidence that it is actually the target of the movement of the object.

Alternatively or in combination with this, several candidate targets for the movement of the object can be provided, and the trajectories of the object predicted for each candidate target can be aggregated to form an overall prediction. This can increase the significance of the future of the movement in particular. For example, if the trajectories that were predicted for a pedestrian to move to many candidate destinations begin to cross a lane, the likelihood that the pedestrian will actually cross the lane is comparatively high.

If, for example, the candidate goals are in the form of a distribution, the prediction of at least one trajectory for each candidate goal automatically results in a distribution of such trajectories.

In a further particularly advantageous embodiment, in addition to a plurality of candidate goals, a number of candidate preferences can also be provided. The trajectories predicted for each combination of a candidate target and a candidate preference are then aggregated to form an overall prediction.

The candidate preferences, as well as the candidate goals, can again be present as distributions. Such a distribution expresses, for example, that a pedestrian with a first probability intends to cross a lane on a pedestrian crossing, while with a second probability intends to cross the lane at another point and thus shorten his path.

In a further particularly advantageous embodiment, combinations of candidate goals and candidate preferences, and / or the trajectories predicted for these combinations, are evaluated and / or weighted to the extent to which they have a history of Movement of the object are plausible. This evaluation and / or weighting can be added seamlessly in particular if there are distributions of candidate goals and candidate preferences. The aggregation can then be carried out with maximum accuracy, ie information is not lost before the actual aggregation by discretization.

The history of the movement of the object can in turn come from any source, such as from a history of physical observations of the object or from information that the object itself has made public.

As explained at the beginning, an important useful application of the described method is to drive vehicles safely in traffic and, in particular, to avoid collisions with weaker road users. The invention therefore also relates to a method for controlling and / or monitoring a vehicle.

In this method, measurement data from the surroundings of the vehicle are recorded by physical measurement with at least one sensor. The sensor can be, for example, a camera, a radar sensor or a LIDAR sensor. The trajectory of at least one controlled object is predicted on the basis of the measurement data using the previously described method.

The predicted trajectory of the object is compared with the currently tracked trajectory and / or with the planned trajectory of the vehicle. In response to the fact that the predicted trajectory of the object affects the currently tracked trajectory and / or the planned trajectory of the vehicle, at least one physical warning device and / or at least one actuator of the vehicle is activated.

The warning device can be, for example, a noise generator with which the noise level of an at least partially electrically driven vehicle is raised in order to make the vehicle more perceptible (Acoustic Vehicle Alert System, AVAS). A horn, for example, can be used as a further warning device that can be perceived outside the vehicle if the warning to a pedestrian, for example, is more urgent. In the context of a driver assistance system, for example, a warning device that is perceptible in the vehicle can also be activated in order to urge a driver of the vehicle to perform a braking and / or evasive maneuver. The actuator can, for example, at least preload a braking system of the vehicle, so that braking that may still be necessary in the further course of the events quickly reaches the full braking pressure.

In a further particularly advantageous embodiment, at least one actuator acting on a drive system, a steering system, and / or a brake system of the vehicle is controlled in such a way that the trajectory of the vehicle that is then tracked is no longer affected by the predicted trajectory of the object. This can be achieved, for example, by braking, an evasive maneuver or a combination of both maneuvers. To avoid a collision, it may be sufficient, for example, to decelerate the vehicle so that it only arrives at a location when the object has already left that location.

The method can be implemented in whole or in part in software which, for example, gives the functionality of one of the methods described to a control unit or a computer for a vehicle. The software then has the direct benefit of the customer, that the prediction of the trajectory of objects becomes more precise and collisions of a vehicle with these objects can be avoided with a higher probability. The invention therefore also relates to a computer program with machine-readable instructions which, when executed on a computer and / or on a control device, cause the computer and / or the control device to carry out one of the methods described. The invention also relates to a machine-readable data carrier or a download product with the computer program.

Furthermore, the invention also relates to a computer and / or a control device with said computer program and / or with said machine-readable data carrier. Alternatively or in combination with this, the computer, or the control device, can also be designed in another way specifically for executing the described method. This specific training can be brought about, for example, by embodying the functionality of the method in the form of application-specific integrated circuits (ASICs).

Further measures improving the invention are described in more detail below together with the description of the preferred exemplary embodiments of the invention with reference to figures.

Figure list

• 1 Erstes Ausführungsbeispiel des Verfahrens 100;
• 2 Weiteres Ausführungsbeispiel des Verfahrens 100 auf der Basis von Verteilungen;
• 3 Beispielhafte Verkehrssituation, in der das Verfahren 100 anwendbar ist;
• 4 Beispielhafte iterative Vorhersage einer Trajektorie 1a auf der Basis jeweils benachbarter Kandidaten-Positionen 1b1-1b8;
• 5 Ausführungsbeispiel des Verfahrens 200.

It shows:

• 1 First embodiment of the method 100 ;
• 2nd Another embodiment of the method 100 based on distributions;
• 3rd Exemplary traffic situation in which the procedure 100 is applicable;
• 4th Exemplary iterative prediction of a trajectory 1a based on neighboring candidate positions 1b1-1b8;
• 5 Embodiment of the method 200 .

To 1 are in step 110 of the procedure 100 the current position 1b and the current speed 1c of an object 1 each determined by physical measurement, for example from camera images, radar data and / or from LIDAR data.

In step 120 becomes a prospective target 2nd the movement of the object 1 provided. For example, according to block 121 multiple candidate goals 21-23 can be provided, and it can then according to block 122 for each of these candidate goals 21-23 using physical observations of the object 1 or its surroundings a measure 21'-23 ' for the likelihood and / or confidence that the respective candidate goal will be determined 21-23 actually the goal 2nd the movement of the object 1 is.

In step 130 is based on physical observations of the object 1 , and / or its surroundings, an expected preference 4th determined when controlling the object 1 towards the expected goal 2nd is being followed. That preference 4th can in turn depend on the expected destination 2nd be (for example: to reach a certain bus or a certain tram), but also a general preference 4th be that of the unlikely goal 2nd depends (for example: minimize the distance covered in the rain). In 1 are given various examples, such as preference 4th can be determined.

For example, according to block 131 been found to be a pedestrian 7 as an object 1 a roadway 8th has entered or is about to enter, according to block 132 on a preference 4th of the pedestrian 7 be closed, the roadway 8th cross from the place of entry.

According to block 133 can depend on the weather, and / or depending on a load condition of the pedestrian determined by physical observation 7 on a preference 4th of the pedestrian 7 be closed, the route to the expected destination 2nd to minimize.

For example, according to block 134 found that the pedestrian 7 through the road 8th from another pedestrian 7 ' is separated and reacts to its presence, according to block 135 on a preference 4th of the pedestrian 7 be closed, the roadway 8th to cross.

In step 140 using the current position 1b and speed 1c of the object 1 in combination with the expected goal 2nd and preference 4th a cost function 5 determined. This cost function 5 occupied in a spatial area 6 which is the current position 1b of the object 1 and the expected destination 2nd contains, each of the object 1 accessible place 6a with a cost value per unit of time spent there and / or distance traveled per unit there.

It is now in step 150 at least one candidate trajectory 1a ' by the current position 1b of the object 1 to the expected goal 2nd leads, a spatial and / or temporal integral of the cost function 5 along this candidate trajectory 1a ' as total cost 5 ' this candidate trajectory 1a ' assigned. The candidate trajectory can be used 1a ' come from any source. For example, candidate trajectories 1a ' in step 149 from the current position 1b and the expected destination 2nd be determined.

With the total cost 5 ' the candidate trajectory 1a ' can be continued in any way. For example, the candidate trajectory 1a ' be refined with any optimization algorithm with the optimization goal, the total cost 5 ' to minimize.

In the in 1 The example shown is in step 160 from several candidate trajectories 1a ' a candidate trajectory 1a ' with the lowest total cost than expected, that is, predicted, trajectory 1a of the object 1 selected.

Like inside the box 160 is sketched, this selection can be made iteratively. According to block 161 can start from the current position 1b of the object 1 neighboring candidate positions 1b1-1b8 each trajectories 1a1-1a8 are determined based on the expected target 2nd lead and each minimal total cost 5 ' to have. According to block 162 can this minimum total cost 5 ' the respective candidate position 1b1-1b8 as a cost value 5 " be assigned. According to block 163 can be a trajectory of the object 1 that have a candidate position 1b1-1b8 with the lowest cost value 5 " leads as an expected trajectory 1a to be selected. That is, as the next step in the trajectory 1a a step to candidate position 1b1-1b8 with the lowest cost value 5 " planned. This process is in 3rd explained in more detail.

According to block 164 can start from a fixed position 1b of the object 1 Trajectories 1a * are determined on different candidate goals 21-23 lead and, for example, minimum total costs 5 ' to have. These trajectories 1a * can make an overall prediction 1a be aggregated.

Alternatively or in combination, a distribution of candidate trajectories can be used 1a ' , their probabilities based on their respective total costs 5 ' are weighted as a prediction of the trajectory 1a be issued.

2nd shows a further embodiment of the method 100 . In contrast to 1 become the candidate targets in this embodiment 21-23 not to a single prospective goal 2nd aggregated, but there is a distribution D (21-23) of many such candidate targets 21-23 provided for further processing. According to block 136 a distribution D (41-43) of many candidate preferences 41-43 provided. The candidate goals 21-23 and the candidate preferences 41-43 in the respective distributions D (21-23) or D (41-43) can optionally step in early 139 evaluated and / or weighted to what extent they have a previous history 1d the movement of the object 1 are plausible. Through this evaluation and / or weighting, the distribution D (21-23) of the candidate goals can 21-23 and / or the distribution D (41-43) of candidate preferences 41-43 be modified accordingly.

In to 1 analogous results in step 140 for a given current position 1b of the object 1 to each pair from a candidate target 21-23 on the one hand and a candidate preference 41-43 on the other hand, a cost function 51-53 . Because both the candidate goals 21-23 as well as the candidate preferences 41-43 when distributions D (21-23) or D (41-43) are present, a distribution D (51-53) of many possible cost functions arises 51-53 .

The determination of candidate trajectories 1a ' in step 149 provides a distribution D (1a ') of many possible candidate trajectories 1a ' because the candidate goals 21-23 available as distribution D (21-23). In step 150 becomes every possible candidate trajectory 1a ' any possible amount from the distribution D (1a ') 5 ' the total cost associated with the candidate trajectory 1a ' with one of the cost functions 51-53 from distribution D (51-53) is evaluated. A distribution D (5 '| 1a') of the total costs arises.

According to block 160a for each combination of candidate goals 21-23 and candidate preferences 41-43 Trajectories 1a * predicted. The distribution D (21-23) of the candidate goals is effective 21-23 in two ways, namely when determining the cost function 51-53 in step 140 and once when determining the candidate trajectories 1a ' in step 149 . The distribution D (41-43) of candidate preferences 41-43 however, only works in step 140 .

At the exit from block 160a there is a distribution D (1a *) of possible trajectories 1a *. For every combination of a specific candidate goal 21-23 and a concrete candidate preference 41-43 this distribution D (1a *) can in principle contain any possible trajectory 1a * that somehow differs from the current position 1b of the object 1 to the concrete candidate goal 21-23 leads. However, the distribution D (1a *) assigns a possible probability to each possible trajectory 1a *, which is higher the lower its total cost 5 ' are. So there are also clearly more "expensive" trajectories 1a * in the distribution D (1a *).

The trajectories 1a * contained in the distribution D (1a *) are in block 164 aggregated. That means in the 2nd shown example that the distribution D (1a *) is then evaluated with the choice of which trajectory 1a the slightest mistake is expected to be made. This can be done in any statistically motivated manner, for example by selecting that trajectory 1a who are most likely to meet the condition that the true goal 2nd just that candidate target 21-23 is and that is the real preference 4th just that candidate preference 41-43 is for the trajectory 1a is cost optimal.

In a special case, each combination can have a specific candidate goal 21-23 and a concrete candidate preference 41-43 targeted cost-optimal trajectories 1a * can be determined. Then each trajectory 1a * is a cost-optimal trajectory on the condition that a specific candidate goal 21-23 the real goal 2nd the movement of the object 1 is and that a certain candidate preference 41-43 the real preference 4th is that in controlling the object 1 is being followed. The trajectories 1a * can be weighted in the distribution D (1a *) with a measure of the probability that this condition applies. In the said special case, there is at least one candidate target for each predicted trajectory 1a * 21-23 and at least one candidate preference 41-43 , for which the trajectory 1a * is actually cost-optimal. In the general case, this is not guaranteed.

3rd shows an exemplary traffic situation in which the related 1 described method is applicable. In the in 3rd illustrated snapshot strives for a pedestrian 7 as a controlled object 1 a goal 2nd to being on the other side of a lane 8th located. The pedestrian 7 therefore the lane 8th somehow cross over to the destination 2nd to reach. At the same time approaching the road 8th a vehicle 60 that its environment 61 with at least one sensor 62 supervised.

From the pedestrian 7 would normally be expected to drive the pavement as required 8th on the over the crosswalk 81 leading candidate trajectory 1a ' crosses. This candidate trajectory 1a ' would neither the trajectory currently driven 60a still the planned trajectory 60b of the vehicle 60 affect.

In the in 3rd presented situation, however, provides another pedestrian 7 ' the pedestrian on the other side of the road 7 an occasion, his preference 4th surprisingly change. On the way to his goal, he is now more important than following the rules 2nd on the other pedestrian 7 ' come over to talk to him. Then, however, is a prediction of pedestrian behavior with a standard cost function, the locations 6a on the crosswalk 81 towards other places 6a on the road 8th clearly preferred, no longer adequate. According to the procedure described above 100 exactly this is taken into account and the cost function 5 adjusted accordingly.

In the end, this leads to a directly on the road 8th leading candidate trajectory 1a ' as an expected trajectory 1a of the pedestrian 7 is predicted. This prospective trajectory 1a affects the planned trajectory 60b of the vehicle 60 , so countermeasures are required. So can warning devices 63a , 63b of the vehicle 60 to be activated. In the in 2nd The example shown is also activated by driving dynamic systems 65a-65c the vehicle 60 brought to evasive and simultaneous braking, so that it is on a trajectory 60c comes to a stop and not with the pedestrian 7 collided.

4th illustrates the iterative prediction of the trajectory 1a . The current position 1b of the object 1 in this example has eight direct neighboring positions 1b1-1b8, which act as candidate positions for the next step based on the current position 1b come into question. Starting from each candidate position 1b1-1b8, the trajectory 1a1-1a8 is determined, which has the lowest total costs 5 ' on the expected goal 2nd leads there. The respective total cost value becomes the respective candidate position 1b1-1b8, which was the starting point, as a cost value 5 " assigned. In the in 3rd Example shown is the candidate position 1b5 the lowest cost value 5 " assigned so that the expected trajectory 1a of the object 1 based on the current position 1b next leads over the candidate position 1b5.

5 shows an embodiment of the method 200 to control and / or monitor the vehicle 60 . In step 210 be with the at least one sensor 62 Measurement data 62a from the environment 61 of the vehicle 60 detected. Based on this measurement data 62a is done using the procedure previously described 100 the trajectory 1a at least one controlled object 1 predicted. This predicted trajectory 1a will in step 230 with the trajectory currently being tracked 60a , and / or with the planned trajectory 60b , of the vehicle 60 compared.

If in step 240 it is found that the predicted trajectory 1a of the object 1a the trajectory currently being tracked 60a , and / or the planned trajectory 60b , of the vehicle 60 affected (truth value 1 ), are in step 250 Countermeasures taken. Specifically, for example, one in the vehicle 60 perceptible warning device 60a and / or an outside of the vehicle 60 perceptible warning device 60b can be controlled. Alternatively or in combination, at least one actuator can be used 64a-64c of the vehicle can be controlled. Actuators 64a-64c can according to block 251 for example on a drive system 65a , on a steering system 65b or on a braking system 65c of the vehicle 60 Act.

QUOTES INCLUDE IN THE DESCRIPTION

This list of documents listed by the applicant has been generated automatically and is only included for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Non-patent literature cited

• V. Karasev, "Intent-aware long-term prediction of pedestrian motion", 2016 IEEE International Conference on Robotics and Automation (ICRA), ISBN: 978-1-4673-8026-3 [0002]

Claims (22)

Method (100) for predicting the trajectory (1a) of at least one controlled object (1) with the steps: A current position (1b) of the object (1) determined by physical measurement is provided (110); • at least one prospective target (2) for the movement of the object (1) is provided (120); • Taking into account physical observations (3) of the object (1) and / or the environment in which the object (1) moves, at least one prospective preference (4) is determined (130) when controlling the object (1) is pursued in the direction of the at least one prospective target (2); • Using at least the current position (1b) of the object in combination with the at least one prospective target (2) and the at least one prospective preference (4), a cost function (5) is determined (140), which in a spatial area (6) , which contains the current position (1b) of the object (1) and the at least one prospective destination (2), each location (6a) accessible for the object (1) a cost value per unit of time spent there and / or per unit there distance traveled, assigned; • At least one candidate trajectory (1a ') leading from the current position (1b) of the object (1) to the at least one likely destination (2) is an evaluation of the cost function (5) along this candidate trajectory (1a') as total costs (5 '), and / or at least one distribution of candidate trajectories (1a') is assigned an expected value of such an evaluation via the candidate trajectories (1a ') in the distribution as total costs (5') (150). Method (100) according to Claim 1 , whereby a candidate trajectory (1a ') with the lowest total costs (5') is selected from 160 candidate trajectories (1a ') as the expected trajectory (1a) of the object (1). Procedure according to one of the Claims 1 to 2nd , whereby a distribution of candidate trajectories (1a '), the probabilities of which are weighted on the basis of their respective total costs (5'), is output (170) as a prediction of the trajectory (1a). Method (100) according to one of the Claims 1 to 3rd , wherein the cost function (5) is selected (141) from a predetermined catalog of cost functions (5), and / or wherein the cost function (5) is parameterized (142) with at least one variable that is used for its determination. Method (100) according to one of the Claims 1 to 4th , on the basis of a semantic segmentation of the spatial area (6) at least one contribution (5a) to the cost function (5) is assigned (143) to each location (6a), which depends on its semantic meaning. Method (100) according to one of the Claims 1 to 5 , whereby each location (6a) is assigned (144) at least one contribution (5b) to the cost function (5), which depends on the extent to which the presence of the object (1) at this location (6a) is permissible according to predetermined traffic regulations. Method (100) according to one of the Claims 1 to 6 Each location (6a) is assigned (145) at least one contribution (5c) to the cost function (5), which depends on how likely the collision of the object (1) with other objects at this location (6a), and / or another event that is potentially harmful to the object. Method (100) according to one of the Claims 1 to 7 A pedestrian (7) is selected as the controlled object (1). Method (100) according to Claim 8 , in response to the determination (131) that the pedestrian (7) has entered a lane (8) or is about to enter it, a conclusion (132) is drawn (132) to a preference (4) of the pedestrian (7), 8) cross from the place of entry. Method (100) according to one of the Claims 8 to 9 , whereby depending on the weather and / or on a loading condition of the pedestrian (7) determined by physical observation, a preference (4) of the pedestrian (7) is inferred (133) to minimize the distance to the expected destination (2) . Method (100) according to one of the Claims 8 to 10 , in response to the determination (134) that the pedestrian (7) is separated from another pedestrian (7 ') by a roadway (8) and reacts to the presence of this other pedestrian (7'), to a preference ( 4) the pedestrian (7) is closed (135) to cross the roadway (8). Method (100) according to one of the Claims 1 to 11 , whereby, starting from several candidate positions (1b1-1b8) adjacent to the current position (1b) of the object (1), a trajectory (1a1-1a8) leading to the expected destination (2) is determined with minimal total costs (5 ') (161), the value of this minimum total cost (5 ') being assigned to the respective candidate position (1b1-1b8) as the cost value (5 ") (162) and a trajectory of the object (1) being derived from the current position (1b) of the object (1) leads via a candidate position (1b1-1b8) with the lowest cost value (5 "), is selected as the expected trajectory (1a) (163). Method (100) according to one of the Claims 1 to 12th , wherein a plurality of candidate targets (21-23) of the movement of the object (1) are provided (121) and wherein for each candidate target (21-23) using physical observations of the object (1), and / or the Environment, a measure (21'-23 ') for the probability and / or confidence is determined (122) that the respective candidate target (21-23) is actually the target (2) of the movement of the object (1). Method (100) according to one of the Claims 1 to 13 , wherein a plurality of candidate targets (21-23) of the movement of the object (1) are provided (121) and wherein the trajectories (1a *) of the object (1) predicted for each candidate target (21 *) in one Overall forecast (1a) can be aggregated (164). Method (100) according to Claim 14 , wherein a plurality of candidate preferences (41-43) are additionally provided (136) and wherein the trajectories (1a *) predicted for each combination of a candidate target (21-23) and a candidate preference (41-43) can be aggregated to a total forecast (1a) (164). Method (100) according to Claim 15 , combinations of candidate goals (21-23) and candidate preferences (41-43), and / or the trajectories (1a *) predicted for these combinations, being evaluated and / or weighted accordingly (139, 164a), to what extent they are plausible with a history (1d) of the movement of the object (1). Method (100) according to one of the Claims 1 to 16 , another vehicle, an animal, and / or a mobile robot being selected as the controlled object (1). Method (200) for controlling and / or monitoring a vehicle (60) with the following steps: • Measurement data (62a) from the surroundings (61) of the vehicle (60) are recorded (210) by physical measurement with at least one sensor (62) ); • On the basis of the measurement data (62a), the method (100) according to one of the Claims 1 to 14 the trajectory (1a) predicts (220) at least one controlled object (1); • the predicted trajectory (1a) of the object (1) is compared (230) with the currently tracked trajectory (60a) and / or with the planned trajectory (60b) of the vehicle (60); • in response to the fact that the predicted trajectory (1a) of the object (1) affects (240) the currently tracked trajectory (60a) and / or the planned trajectory (60b) of the vehicle (60), at least one physical warning device (63a, 63b), and / or at least one actuator (64a-64c) of the vehicle (60) controlled (250). Method (200) according to Claim 18 , wherein at least one actuator (64a-64c) acting on a drive system (65a), a steering system (65b), and / or a brake system (65c) of the vehicle is controlled such that the trajectory (60c) then traced Vehicle (60) is no longer affected by the predicted trajectory (1a) of the object (1). Computer program containing machine-readable instructions which, when executed on a computer and / or on a control device, cause the computer and / or the control device to carry out a method (100, 200) according to one of the Claims 1 to 19th to execute. Machine-readable data carrier or downloadable product with the computer program. Computer and / or control device with the computer program Claim 20 , and / or with the machine-readable data carrier and / or the download product Claim 21 , and / or in another way specifically designed to carry out the method (100, 200) according to one of the Claims 1 to 19th .

Images