DE102016212700A1 - Method and system for controlling a vehicle - Google Patents

Abstract

A method is provided for controlling a vehicle, comprising the steps of: when the vehicle is approaching a predetermined location (S1), providing (S2) a location-based model that depicts a typical behavior of a vulnerable road user at the predetermined location to a control unit the vehicle (3), determining (S4) a context in which the vulnerable road user is currently located, the context being formed from measured values of an environment detection (S3) with respect to the vulnerable road user, determining (S5) a prediction of the behavior of the vulnerable road user with the location-based model using the determined context, and determining (S6) a response to the determined prediction by the control unit of the vehicle. Furthermore, a corresponding system for carrying out the method and a correspondingly equipped vehicle are provided.

Description

The invention relates to a method and a system for controlling a vehicle. In particular, the invention relates to a method and a system for controlling a vehicle, in which traffic safety can be improved by using models for describing the behavior of vulnerable road users.

State of the art

It is becoming apparent that, due to the increasing automation of vehicles, technologies will be increasingly used in the future to analyze the behavior of other road users. Here, computers will take on more and more tasks that can be done so far only one person (here: the driver). In particular, the analysis of the behavior of pedestrians poses complex challenges, for example because they are able to change their direction of movement and speed very quickly.

The publication DE 10 2013 202 463 A describes a method for determining a movement model of a vulnerable road user. In this case, motion influencing parameters are determined that are representative of a given situation and / or a predefined device and / or a predetermined significant point and that can also have an influence on a movement of the detected vulnerable road user. Depending on the determined movement influencing parameters and the determined position of the vulnerable road user, a movement model of the vulnerable road user is determined by means of which a position of the vulnerable road user is predictable.

There is a need to continue to increase road safety with regard to the behavior of vulnerable road users such as pedestrians, and to provide methods and systems for improved prediction of the behavior of the vulnerable road users.

Disclosure of the invention

A method for controlling a vehicle according to independent claim 1 and a system according to claim 12 are provided.

Preferred developments of the invention will become apparent from the dependent claims.

According to one aspect of the invention, there is provided a method of controlling a vehicle, wherein when the vehicle approaches a predetermined location, a location-based model that depicts a typical behavior of a vulnerable road user at the predetermined location is provided to a control unit of the vehicle , Furthermore, a context in which the vulnerable road user is currently located is determined from measured values of an environment detection with regard to the vulnerable road user. With the location-based model, a prediction of the behavior of the vulnerable road user is determined using the determined context, and a response to the determined prediction is determined by the control unit of the vehicle.

The method provided can advantageously increase a prediction accuracy with respect to a behavior of a vulnerable road user. In particular, a prognosis of a movement behavior and the course of a movement trajectory of the vulnerable road user can be improved.

This is made possible in particular by the provision of the location-based model and the use of the respective context. Thus, according to the method, a model is used which is specifically intended for the specific, determined location. Due to the fact that the model is practically only applicable at this location or applies only to this particular place and is made available accordingly, in particular local features such as fences, walls, traffic, etc. can be clearly better represented. The respective individual starting situation can be taken into account by the respective context, which in particular can be included as a parameter in the form of a vector in a respective evaluation or prediction. It is thus possible to make the greatest possible preparation for efficient prediction at the particular location and to use a prepared model, the context being able to be composed of any measurements which may depend in particular on the particular model. The computational effort at runtime when passing through the site can be minimized or better utilized.

For example, the term "vulnerable road user" refers to a pedestrian. The procedure is not limited to pedestrians. Rather, the method is applicable to any vulnerable road users, such as bicycles and other light or non-motorized vehicles, or to animals, such as dogs.

The method makes it possible, in particular, to model the typical behavior of pedestrians so accurately that predictions with a comparison with the prior art significantly increased time horizon of several seconds can be performed, compared to recent few hundred milliseconds for reliable statements about pedestrian behavior.

The method can be applied particularly advantageously at a road intersection or a pedestrian crossing. The pedestrian crossing can be characterized for example by a crosswalk or a traffic island. Thus, according to the method, traffic safety can be increased, especially at these locations, but the method is not limited to such locations.

In one embodiment, the location-based model is transmitted from a server to the vehicle in response to a request from the vehicle. For this, the vehicle contacts the server. However, according to a variant of the embodiment, it is also conceivable that the server or an associated communication unit initiates a transmission of models available to the vehicle.

In this way, current, location-based movement and behavioral models can be made available to all nearby or interested vehicles. In particular, learned location-based models can be stored on the server and later retrieved, evaluated, and / or expanded or replaced. In doing so, models that concern a specific location can be used locally. This applies in particular to models that have been learned by a comparatively high amount of computation, or that have been adapted particularly finely or tediously to the place under consideration and are now prepared in accordance with the method. The method can be compared to previous pedestrian predictions with the previous very simple models that allow only a low time horizon, a high improvement and increase in accuracy can be achieved.

Furthermore, the models may additionally or alternatively also be stored in the vehicle or may be available for reuse, for example during a repeated passage through the intersection. In such a case, a synchronization between the server and the vehicle can be carried out when approaching or passing by at the predetermined location.

The server can be a local unit on site, which can be communicated via WLAN, for example, or a central server. The server can thus be installed as a local entity at the given, special location. In this case, the communication between server and vehicle, for example by means of WLAN or Car2X technology done. The Car2X communication generally allows the exchange of any data between a vehicle and other objects, such as other vehicles, servers, pedestrian smartphones, etc. Alternatively or additionally, a central server may be present, for example, via radio, for example via LTE can. For example, a local unit may have a respective model for a pedestrian crossing or an intersection, whereas a central server or a corresponding network may for example manage the associated data for an entire city.

The environment detection can be done by the vehicle and / or by locally installed sensors. This can be done for example by stereo cameras, laser scanners, infrared cameras, ultrasonic sensors, radar, lidar, etc. Furthermore, the local unit may be equipped with its own sensors, such as stationary cameras. Furthermore, the entire environment of the vehicle or the entire traffic scenario can also be detected on site.

When the environment detection is performed by the local unit, vehicles may also receive information about participants that they themselves can not see. As a result, for example, covers or missing or defective sensors can be compensated. Even vehicles that do not have their own sensors can benefit from this information.

The determination of the prediction can be carried out by the control unit of the vehicle.

Alternatively or supportively, the determination can also be made by the server. For example, the complete behavioral analysis can be taken over by the server.

In this way, in particular complex calculations can only be performed once on the server, so that the respective in-vehicle computer requires less computing power. For this purpose, for example, when approaching an overpass / an intersection, the vehicle can periodically report all observed road users to the server. The server may then, for example, determine a prediction and report back to the vehicle.

Further, the model may describe likely motion trajectories, and based on the particular context, a most probable motion trajectory is determined.

In one embodiment, the control unit is connected to a driver assistance system of the vehicle.

According to another preferred embodiment, the vehicle is an automatically moving vehicle.

Thus, according to the method above all driver assistance systems and automated vehicles can be an ideal decision making at the specific location possible.

In summary, the process may reinforce the fact that the behavior of pedestrians is highly dependent on local circumstances.

The method can also advantageously be extended or supported by a corresponding learning method, and it can in particular be used to learn current, location-based movement and behavior models for road users. In particular, it can be made possible to constantly expand the model by using learning methods and thus to map any changes, which are caused, for example, by an increase / decrease in local traffic, in the behavior of the vulnerable road users.

According to one embodiment, during movement detection, movement trajectories of all road users in the respective local traffic scenario are continuously recorded.

In this case, the detected movement trajectories are preferably automatically marked after leaving the field of view of the surroundings detection. In other words, there is an automatic labeling of the trajectories, in particular the pedestrian movement trajectories. In this way, a continuous collection of measurement data for the generation of ordered data can be made. These can be used as training and / or test data for learning and / or for categorization.

Depending on the measurement, the marking or the attached label or the marking can be stored as a binary value and, for example, state whether in a detected traffic scenario a pedestrian has crossed a respective road or not, or whether a vehicle has stopped at a certain stop line or not , The place and time can also be saved. Labeling can also serve as a basis for creating a model.

Furthermore, by labeling for measured movement trajectories, a subsequent assessment or assignment to a rule-compliant behavior or non-rule-compliant behavior can be done and a corresponding model can be learned. The assessment as to whether or not there is a non-compliant behavior can, for example, be made taking into account associated light signal statuses or according to other criteria which are present during the detection of the movement trajectory.

In this way, the method can advantageously be used to classify completed trajectories as a function of the respectively recorded status of the relevant traffic signal system either as standard behavior or non-conforming behavior. This classification can now be used to improve an automatic learning process, which maps the situation-specific norm behavior mentioned above. For this purpose, all non-compliant trajectories can be ignored when learning the standard behavior.

According to a preferred refinement, a first and a second location-related model are used in parallel, wherein the first location-related model is used for the precise prediction of movement trajectories of the vulnerable road user and the second location-related model for mapping serves a norm-compliant behavior of the vulnerable road user. With a significant mutual deviation of a prediction of the first model and the second model, a non-compliant behavior can be detected by the control unit of the vehicle.

As a result, advantageously based on a trajectory prediction, atypical or non-compliant or illegal, i. potentially dangerous behavior of the vulnerable road user can be reliably detected. The atypical or illegal behavior can be reflected, for example, in the fact that a pedestrian crosses the street despite a red traffic light or, for example, crosses the pedestrian area next to the designated areas. Driver assistance systems and systems for automatic driving are thus in particular made possible to particularly take into account a pedestrian malfunction in their action planning and / or to warn a human driver of a potential danger.

The detection of atypical or illegal road users can be equated with the early detection of a potential risk of collision. This allows, for example, a driver assistance system to generate a corresponding warning and thus to sensitize a human driver for this danger.

Special benefit is given to systems of automatic driving. These can react to a potential danger situation, for example by adjusting parameters. A possible reaction of the vehicle may, for example, lead to an increased safety distance, in particular with regard to the curb, or a more defensive driving style, such as premature braking when approaching a pedestrian crossing.

In a further advantageous embodiment, the system according to the invention also permits the learning of standard behavior at each crossing situation, for example by observing the trajectories of the pedestrians and the association with known light signal states. As a result, the data set can be distinguished according to "standard behavior" and "non-conforming behavior" and subsequently used to adapt the classification and prediction models.

According to one variant, if a deviation of an instantaneous movement of a vulnerable road user in relation to a movement trajectory predicted by the model is detected in a currently present traffic scenario, the movement can be interpreted as non-compliant behavior. In particular, the current behavior of a pedestrian can be observed and predicted and compared downstream with a trajectory database or a model of situation-specific standard behavior. For example, a current status of all relevant traffic lights can be explicitly taken into account. If there are differences, atypical behavior can be assumed.

Thereupon, for example, a corresponding warning to a driver and / or a command to a control unit of a driver assistance system can be output or an autonomous vehicle can be controlled accordingly.

Furthermore, according to a preferred embodiment, after passage of the location with the vehicle, data about a currently detected movement trajectory of the vulnerable road user can be transmitted to the server.

According to a preferred embodiment, measurements for detecting movement trajectories of road users that can be assigned to the traffic scenario are continuously carried out in a respective traffic scenario, wherein the detected movement trajectories are respectively assigned to a training data record or a test data record.

As a result, the data basis can be continually extended and / or updated, which further increases the reliability of the forecasts made possible. Through an appropriate distribution of the old and the new data in training data and test data, an optimized training run is made possible at any time.

According to a preferred embodiment, in a respective traffic scenario, measurements are continuously made for determining current values for features used in the selected model used and for detecting motion trajectories of road users that can be assigned to the traffic scenario, wherein the detected motion trajectories are respectively assigned to a training data set Test data record to be assigned.

Preferably, the location-based model is automatically generated, for example, by a collaboration of a plurality of vehicles takes place.

For this purpose, basically steps may be carried out in which a given feature set of features that are suitable for describing a local traffic scenario at the location is determined, a subset of the complete feature set to be selected is determined, and the model with the selected subset is learned.

As a result, the prediction accuracy can be further increased. This may be due, for example, to the fact that, owing to the reduced number of features or features on which the model is based, an available computing power is better utilized and a faster calculation can be carried out, the complexity of the algorithms used being advantageously reduced. Advantageously, it is possible to explicitly use not all available information, but only the information that is currently of actual relevance. An over-adaptation through the use of too many features can be avoided, whereby the accuracy continues to increase comparatively. The method is particularly applicable to any location in urban areas.

Preferably, the location-related model is evaluated based on the environment detection, wherein in particular a detected movement trajectory of the vulnerable road user can be consulted. In this case, the size of a deviation of a currently detected movement of a vulnerable road user with respect to a prediction based on the location-based model can be determined.

According to a particularly advantageous embodiment, a training run for adapting the location-related model is performed when based on the evaluation, an error rate exceeding a predetermined lower threshold is determined. In this case, to update the data, a plurality of detected movement trajectories of vulnerable road users can be divided into training data and test data, and the step of identifying a subset to be selected with the current data can be repeated.

According to another development of the method, an iteration is carried out for identifying the subset to be selected by vehicles and / or the local unit repeatedly performing comparative measurements according to a specific method. In particular, an iteration can be carried out by doing so, wherein in each iteration step a test-wise subset is determined in which a predefined number of features are removed and / or features are changed. From the test subset, a preliminary model is learned, which is evaluated on the basis of a predefined quality measure.

To determine the context, measurements can be taken to determine current values for the features of the selected subset in predefined time steps. Furthermore, historical values from a predefined number of previous time steps can also be used.

According to further aspects of the invention, a corresponding system for carrying out the method and a correspondingly equipped vehicle are provided,

Brief description of the drawings

Preferred embodiments of the present invention will be explained in more detail with reference to the accompanying drawings. Show it:

1 a schematic diagram of a situation at a pedestrian crossing, illustrating an embodiment of the invention,

2 a schematic diagram of a situation at a pedestrian crossing, illustrating an embodiment of the invention, wherein compared to 1 an additional obstacle that makes it difficult to see between the vehicle and the pedestrian,

3 a schematic diagram of a situation at a pedestrian crossing, illustrating an embodiment of the invention, wherein compared to 2 a local unit is equipped with its own stationary sensors,

4 a flowchart of a method for controlling a vehicle according to an embodiment of the invention,

5 a block diagram of a system for controlling a vehicle, according to an embodiment of the invention,

6 a schematic diagram of a situation at a pedestrian crossing and a related system structure, illustrating an embodiment of the invention, wherein a non-compliant behavior of a vulnerable road user can be considered, and

7 a flowchart of a method for controlling a vehicle according to an embodiment of the invention.

Description of embodiments

In the figures, like reference numerals designate the same or functionally identical elements.

The 1 Figure 12 is a schematic diagram illustrating a situation of a typical traffic scenario at a pedestrian crossing 1 to illustrate an embodiment of the invention. The pedestrian crossing 1 is marked by a crosswalk. The environment / area of the pedestrian crossing 1 is according to the aspects of the invention as a predetermined location 2 designated. Like in the 1 is shown approaching a vehicle 3 the pedestrian crossing 1 at. This takes the vehicle 3 Contact to a server 4 and asks for the movement / behavioral models 5 for pedestrians 6 At this place 2 at. Thereupon is through the server 4 a local model 5 selected based on the nature of the situation at the specific location 2 is adjusted. It can be a stored model 5 also be divided into several parts. For example, in a variant of the system and method according to the invention, a country-specific submodel can also be provided, which already generally depicts the behavior in a particular country, so that country-specific differences are included in the submodel. This sub-model can then be supplemented by a local model, which in particular depicts the local structural peculiarities, from which the overall model 5 is formed. According to a variant of the embodiment, the models 5 instead of in the server 4 , the on-site local as a unit 8th is also stored on a remote server (not shown). In such a variant, the remote server can use the models 5 manage for a variety of predefined locations of an entire region or city.

In general, the server submits 4 the required models 5 to the passing vehicles of interest. The model 5 It depicts the typical behavior of pedestrians in this location. The pedestrian behavior, for example, by so-called approach trajectories 9 or movement trajectories 9 to be discribed. Thus, a model 5 a bevy of movement trajectories 9 describe. In the embodiment shown, the sum of all approach trajectories describes 9 all previously observed paths on which the pedestrian crossing 1 or the crossing has been crossed so far.

Because of the special location-based model 5 to the vehicle 3 is transmitted to the vehicle 3 allows the behavior of the vehicle sensors 10 recorded pedestrian 6 to determine.

In the 2 becomes one too 1 similar situation, but with the most probable movement trajectory 9 opposite to the 1 shown differs as an obstacle 11 , For example, a wall exists.

The context, which is usually represented as a vector, can be composed of arbitrary measurements. The vehicle sensors 10 may include, for example, cameras, laser scanners, radar, etc. The context vector can, for example, have measurements on the speed and acceleration of all participants, here: the instantaneous speed and acceleration of the vehicle 3 and the pedestrian 6 , the relative distance between pedestrians 6 and vehicle 3 , the distance of the pedestrian 6 to the curb 12 , and the distance to the pedestrian crossing 1 or to the intersection for all participants 3 . 6 ,

Depending on the model 5 The context may be composed of other measurements. For example, the context may be composed of metrics for particular features or features to which the particular model pertains 5 based. It can be considered in particular that the pedestrian 6 the street 7 traversing in a direction that runs in the drawing from top to bottom, wherein the determination of the context, a distance to the nearby curb 12 is determined. The model 5 to 2 is generally different from the model 5 to 1 , due to the presence of the wall 11 ,

In response to the prediction of the movement trajectory, the control unit selects 13 of the vehicle 2 a suitable response, such as the initiation of a braking process.

After crossing the way 1 or the intersection can be the vehicle 3 optional trajectories 9 all watched pedestrians 6 to the server 4 to transfer.

The server 4 can then, if necessary, using a learning process, the movement / behavioral model 5 To update.

The original movement / behavior model 5 was created using the same learning method from a given training record. Alternatively, the commissioning at a new pedestrian crossing / a new intersection, but also without such a training record done. In this case, the system can only collect the measurement data of all vehicles in the first few days. If a certain amount of pedestrian trajectories has been observed, for example a few hundred, the first models will be 5 generated and transmitted.

In the 3 will be a similar situation as in 2 shown, the vehicle 3 but coming from the other direction and the pedestrian 6 can not see. The local unit 8th to 3 is with own sensors 14 , such as cameras, equipped and can be the pedestrian 6 within the field of view 15 see. The local unit 8th may for example be attached to a traffic light. From the perspective of the vehicle 3 Be missing information through the additional camera 14 the local unit 8th provided. In this way, even road users, especially the pedestrian 6 , are observed when no vehicle equipped with sensors is nearby or in visual contact.

The 4 shows a flowchart of a method for controlling a vehicle according to an embodiment of the invention. In step S1, a vehicle approaches a predetermined location, such as a road intersection or a pedestrian crossing. In step S2, a location-related model is provided which depicts a typical behavior of a vulnerable road user at the given location. The model is transmitted to the control unit of the vehicle. In step S3, an environment detection is performed at the predetermined location. This can be done by locally installed sensors or by in-vehicle sensors. The processing of the measured data can be carried out in this embodiment, in particular by a control unit of the vehicle and / or by a locally installed control unit. In step S4, with respect to the road user, the current context in which the road user is located at the location is determined. This happens for example by a selective selection of measurement data or a selective measurement of predetermined Features / characteristics. In step S5, a prediction of the behavior, in particular the movement behavior of the vulnerable road user with the location-related model based on the particular context is performed. Finally, in step S6, a response of the vehicle to the determined prediction is determined by the control unit and provided to the driver or a vehicle control system.

The 5 shows a system 20 to the controller and a vehicle. The system 20 has a measurement data acquisition system 21 with a variety of sensors 22 to capture the environment in a local traffic scenario. Further, the system rejects 20 one with the measurement data acquisition system 21 coupled measurement data analysis unit 23 on, which is used to assign the acquired measurement data to defined characteristics of a predefined complete feature set for the description of the local traffic scenario. The measurement data analysis unit 23 also has a movement trajectory identification unit 24 for identifying movement trajectories from the acquired measurement data and a subset selection unit 25 for identifying a subset of the complete feature set to be selected. Further, in the system 20 a control unit 26 arranged with the measurement data analysis unit 23 and a control system 13 of a vehicle 3 is coupled.

The 6 shows a schematic diagram of a situation at a pedestrian crossing and an associated system structure according to an embodiment of the invention. In this case, according to the embodiment shown here, a non-compliant behavior of a vulnerable road user is explicitly taken into account.

The system has a central control unit 8th On the one hand, status information of all traffic lights relevant to the current route 16 receives (red / yellow / green) and further equipped with sensors, such as cameras, with which a detection of all present road users is enabled. The control unit 8th is here as a local stationary unit 8th but may also be at least partially in the vehicle 3 be arranged. The cameras are not explicitly shown in the drawing, instead only two exemplary viewing areas 17 of possible at the traffic lights 16 attached cameras sketched. In addition, a Funktionsausprägung is conceivable, in addition to the in a vehicle 3 installed sensors are used. An indicated movement trajectory of a curb-enduring pedestrian is denoted by the reference numeral 18 and an indicated movement trajectory of a crossing pedestrian is denoted by the reference numeral 19 designated.

The starting point for the method is the current position or the previously observed trajectory of the relevant vulnerable road users 6 , Assuming a red pedestrian traffic light 16 is designed with the system between pedestrians stopping at the curb 6 and crossing pedestrians 6 be differentiated. The latter would in this case represent the "illegal" behavior.

Two systems are combined for this purpose: A first system is used to predict pedestrian behavior, especially pedestrian trajectories 9 to recognize atypical pedestrian behavior. This is initially a precise prediction of pedestrian trajectories 9 based on the trajectory of the pedestrian, for example. The second system has a movement model 5 on which the location-related, typical behavior / situation-specific standard behavior, taking into account the relevant traffic lights 16 , pictures.

If significant differences are identified when comparing the two predictions, it can be assumed that the pedestrian 6 atypical behaves and, for example, in the red street 7 crossed. This can be recognized, for example, by the fact that the pedestrian 6 with constant or even increasing speed of the red traffic light 16 approaches, although the standard behavior provides a sinking speed. This knowledge can now be used either to generate a warning for the driver with a driver assistance system or to prepare emergency braking, for example by applying the brake discs. In highly or fully automatic systems, the parameterization of relevant subsystems can be adapted, for example by increasing the safety margin to the curb 12 , early deceleration when approaching the overpass 1 , etc.

The method is not limited to detection of atypical procedures with respect to traffic light states. Instead, other atypical behaviors, such as a drunken pedestrian's behavior, may also be detected. Under certain circumstances, this case would be recognizable by a serpentine form of a currently detected trajectory. Furthermore, distracted pedestrians can be recognized, who are busy with a smartphone, for example, and pursue a trajectory that is too straightforward instead of the trajectory that is usually at least slightly curved when approaching an overpass. If necessary, it is possible to recognize playing and therefore inattentive children. This can be reflected, for example, in many short changes of direction and speed changes. Expanded conceivable are scenarios with alone children who may inadvertently participate in road traffic without a caretaker.

The 7 shows a flowchart of a method for controlling a vehicle according to a preferred embodiment of the invention.

In step S11, the process starts. In step S12, movement trajectories are continuously detected by all road users and subsequently each automatically labeled, i. described or marked. The required trajectories can be provided both by stationary sensors, such as surveillance cameras, and by vehicles equipped with different sensors. The trajectories can then be collected and evaluated on a server, for example. Automatic labeling / marking is possible as soon as the road user has left the viewing area of all the sensors, i. as soon as the movement is completed from the system viewpoint. Depending on the model to be created, such a label may include, for example, the information as to whether the pedestrian has crossed the street or not, as a binary value, and if so, where and when the road user leaves the road, and if the vehicle is at a particular location Stop line stopped or not, as a binary value.

In step S13, a continuous evaluation of the models is performed based on the trajectories taken at runtime. It checks whether the error rate exceeds a predefined threshold. If not, the process restarts and begins again with step S11.

If the error rate exceeds the threshold in step S13, a new training run is performed in steps S14 to S16. For this purpose, the detected trajectories are meaningfully divided into a training and a test data record in step S14. In step S15, a feature subset or feature subset is identified that best describes both the local conditions and the interactions between multiple road users and between a road user and the environment.

By identifying the best possible feature set, a small subset is selected from a large number of features of the given feature set, which best describes the current, local situation. To do this, the predefined feature quantity must first be defined. This can have the following features / features: General characteristics, such as speed in 2D / 3D coordination with amount and direction and / or. Further, the predetermined feature set may include features related to interaction with the environment, such as given by a map, such as distance to predefined points, for example. Objects, such as a distance to the curb, for pedestrians, cyclists, etc., a distance to a defined official pedestrian crossing, which is for example characterized by a crosswalk, a traffic light or a traffic island, or a distance to a real or imaginary stop line The distance can be defined in each case by a Euclidean distance and taking into account possible known obstacles as a minimum remaining distance up to the considered point / object, but it is also conceivable to use all distances taking into account the current or theoretical, for example maximum speed to portray times. Furthermore, pedestrian-specific features can be used. These may include, for example, a line of sight or a movement of individual body parts, such as the upper body, the arms, legs, etc. It is also possible to take into account features relating to an interaction between at least two road users. These may, for example, relate to a distance between the participants, such as a Euclidean distance, taking into account known obstacles as minimum remaining distance, or a relative speed in 2D / 3D coordinates, an amount, a direction, or a so-called time-to-distance. Collision. The list of possible predefined features or features can be extended as desired in a suitable manner.

Normally, measured values for the features or selected features are respectively determined for the time step currently to be calculated. To extend the feature space, the history of all objects can also be taken into account by the additional determination of all the features mentioned above for any number of preceding time steps.

To select a subset from the assigned set of features, i. In particular, an iteration can be carried out to determine the smallest possible feature set with the best possible modeling result. In each iteration step, a model is generated and evaluated on the basis of a defined quality measure. This quality measure allows a ranking of all features. Now a defined number of features, at least one, with the lowest ranking from the feature set can be removed and a new iteration can be started.

For each iteration step, an evaluation is performed using the test data set, which examines the predictive accuracy of each model.

After elimination of all features, the subset that achieved the highest prediction accuracy can be selected based on the evaluation results.

In step S16, a new motion / behavior model is learned with the selected feature set. Taking into account possibly limited computing capacities for the analysis of trajectories at runtime ("online"), it is also conceivable to select a minimally worse model, if significant computing power can be saved as a result In step S17 the method ends and goes back to the start in step S 11. According to the method, the model can thus be constantly checked at runtime If a deterioration of the prediction accuracy is detected, a new training run can be initiated at any time, including a renewed determination of a best associated feature set, ie a suitable feature subset.

QUOTES INCLUDE IN THE DESCRIPTION

This list of the documents listed by the applicant has been generated automatically and is included solely 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.

Cited patent literature

• DE 102013202463 A [0003]

Claims (13)

Method for controlling a vehicle ( 3 ), comprising the steps: when the vehicle ( 3 ) a given location ( 2 ) (S1), providing (S2) a location-related model ( 5 ), which is a typical behavior of a vulnerable road user ( 6 ) at the given location ( 2 ) to a control unit ( 13 ) of the vehicle ( 3 ), Determining (S4) a context in which the vulnerable road user ( 6 ), wherein the context of measured values of an environment detection (S3) with respect to the vulnerable road user ( 6 ), determining (S5) a prediction of the behavior of the vulnerable road user ( 6 ) with the location-related model ( 5 ) using the determined context, and determining (S6) a response to the determined prediction by the control unit ( 13 ) of the vehicle ( 3 ). Method according to claim 1, wherein the location-related model ( 5 ) from a server ( 4 ) to the vehicle ( 3 ), whereby the server ( 4 ) to a local unit on site or to a central server ( 4 ), and wherein the surroundings detection (S3) by the vehicle ( 3 ) and / or locally installed sensors ( 14 ) and the determination of the prediction by the server ( 4 ) and / or by the control unit ( 13 ) of the vehicle ( 3 ) and where by the model ( 5 ) Movement trajectories ( 9 ) and, based on the particular context, a most probable movement trajectory ( 9 ), and after passing through the place ( 2 ) a currently detected movement trajectory ( 9 ) of the vulnerable road user ( 6 ) to the server ( 4 ) is transmitted. Method according to claim 1 or 2, wherein the control unit unit ( 13 ) with a driver assistance system of the vehicle ( 3 ) or the vehicle ( 3 ) an automatically moving vehicle ( 3 ), in response to the prediction by the control unit ( 13 ) a control command for the vehicle ( 3 ) is produced. Method according to one of the preceding claims, wherein the surroundings detection (S3) continuously detects movement trajectories (S3) 9 ) of road users ( 6 ) are detected in the respective local traffic scenario, wherein after leaving the field of view of the surroundings detection (S3) the detected movement trajectories ( 9 ) are automatically marked. Method according to claim 4, further comprising an identification of the detected movement trajectories ( 9 ) is carried out as a rule-compliant or non-compliant behavior. Method according to claim 5, wherein the model ( 5 ) has a first and a second location-related model, wherein the first location-related model for precise prediction of movement trajectories ( 9 ) of the vulnerable road user ( 6 ) and the second location-related model for mapping a norm-compliant behavior of the vulnerable road user ( 6 ), wherein in the case of a significant mutual deviation of a prediction of the first model and of the second model, non-compliant behavior by the control unit (FIG. 13 ) of the vehicle ( 3 ) is recognized. Method according to one of the preceding claims, comprising, for the automatic generation of the model ( 5 ), Determining a predetermined feature set from features that are used to describe a local traffic scenario at the location ( 2 ), identifying (S15) a subset of the complete feature set to be selected, and learning (S16) the model ( 5 ) with the selected subset. Method according to one of the preceding claims, comprising evaluating (S13) the location-related model ( 5 ) based on the environment detection (S3), in particular based on a detected movement trajectory ( 9 ) of the vulnerable road user ( 6 ), wherein the size of a deviation of a currently detected movement of a vulnerable road user ( 6 ) with respect to a location-based model ( 5 ) based prediction is determined. The method of claim 8, comprising performing a training run to adapt the location-related model ( 5 ), if, based on the evaluation, an error rate which exceeds a predetermined threshold value is determined, wherein a multiplicity of detected movement trajectories ( 9 ) vulnerable road users ( 6 ) is divided into training data and test data and the step of identifying a subset to be selected is repeated with the current data. Method according to one of claims 7 to 9, wherein an iteration is carried out for identifying the subset to be selected, wherein in an iteration step a test subset is determined in which a predefined number of features is removed and / or features are changed, and wherein respective test ways Subset a provisional model is learned and evaluated on the basis of a predefined quality measure. Method according to one of claims 7 to 10, wherein for determining the context measurements are made for determining current values for the features of the selected subset in predefined time steps, and further wherein historical values from a predefined number of previous time steps are used. System ( 20 ) for controlling a vehicle ( 3 ), comprising: a measurement data acquisition system ( 21 ) with sensors ( 22 ) for detecting (S3) the environment in a local traffic scenario, one with the measurement data acquisition system ( 21 ) coupled measurement data analysis unit ( 23 ) for assigning the acquired measurement data to defined features of a predefined complete feature set for describing the local traffic scenario, wherein the measurement data analysis unit ( 23 ) further comprises a movement trajectory identifying unit ( 24 ) for the identification of movement trajectories ( 9 ) from the acquired measurement data and a subset selection unit ( 25 ) for identifying a subset of the complete feature set to be selected, and a control unit ( 25 ) with the measurement data analysis unit ( 23 ) and a control unit ( 13 ) of the vehicle ( 3 ), the system ( 20 ) is adapted to carry out the method according to one of claims 1 to 11. Vehicle ( 3 ) with a control unit ( 13 ), which the control unit ( 8th ) of the system ( 20 ) according to claim 12 or connected to it and the vehicle ( 3 ), participants in the system ( 20 ) according to claim 12.

Images