EP1947623B1 - Method and device for the dynamic classification of objects and/or traffic situations - Google Patents

Description

The invention relates to a method and a device for the dynamic classification of objects and / or traffic situations. In particular, the invention relates to a method for classifying objects and / or traffic situations, wherein for the classification of at least one object or a traffic situation, a first classifier and at least a second classifier are provided, wherein the first classifier and the at least second classifier are different or differently trained classifiers are. The invention further relates to a device for object and / or situation classification comprising an assistance system having a first classifier, wherein the assistance system has at least one second classifier different from the first classifier and / or one differently trained compared to the first classifier.

The assistance of the driver of a motor vehicle by means of technical means has become increasingly important in the recent past. Depending on the intended use, such technical aids are also referred to as driver assistance or safety systems.

One group of driver assistance or safety systems is the group of so-called predictive driver assistance or safety systems. With these systems, objects and / or traffic situations are to be detected and classified by technical aids, in particular by cameras or other sensors, in order to be able to make the driver aware of possible dangerous situations at an early stage or initiate countermeasures. Such countermeasures are, for example, the triggering of belt tensioner systems and interventions in brake or steering.

For the classification of objects and / or traffic situations, different classification methods are used, for example based on decision trees, neural networks or support vector machines. Classification software modules created on the basis of such classification methods are referred to as classifiers.

Classifiers can be optimized using training data to increase the number of objects and / or traffic situations to be detected, as well as the success rate. The disadvantage, however, is that training the classifiers is not only costly, but also there is a risk that classifiers are "over-trained". Indeed, over-training of a classifier will result in this classifier having high performance with respect to the trained patterns, whereas performance in classifying non-trained patterns will be significantly reduced. A generalized classifier, on the other hand, has high performance with respect to non-trained patterns, but in special cases does not achieve the performance of a specialized classifier. Therefore, classifiers can not be interpreted as having a high classification performance in nearly all applications, similar to the human brain.

The DE 103 36 638 A1 describes a generic method and a generic device for classifying objects by means of an environment sensor. In this case, the surroundings sensor system records object data which comprise a shape, dimensions and a speed of the respective objects. These object data are fed to different classifiers. A first classifier is here designed to recognize a pedestrian on the basis of the object data made available to him, while a second classifier is designed to recognize a car.

The WO 2005/052883 A1 describes a method for determining a driving situation, for example stop and go traffic, city traffic or highway traffic. Here, a driving speed as well as longitudinal and lateral accelerations of the motor vehicle, an engaged gear and steering wheel movements are detected as input data and evaluated by means of a trained neural network as a classifier. It stores the input data for which Driving situation typical data compared and classified due to a similarity of the data the driving situation.

The WO 2005/064566 A1 describes a method in which for a road type, such as a highway, main road, side road or the like, an expected minimum speed is determined. If this minimum speed is fallen below, the traffic situation is classified as a "traffic jam". Here, typical minimum speeds can be modified by taking into account boundary conditions such as weather and road guidance. The road type is determined by means of a navigation system.

The post-published DE 10 2005 043 471 A1 describes a method for driving a driver assistance system. Here, different observers are used to observe spatially different sections. For example, a first observer can be used in dense traffic and a second observer in quiet traffic. The observers then track vehicles recognized as objects on the respective sections of track.

The object of the invention is to provide a method and a device for classifying objects and / or traffic situations with increased performance.

The object is achieved according to the invention with the features of claims 1 and 13, respectively.

According to the method according to the invention, a first classifier and at least one second classifier are made available for classifying objects and / or traffic situations for classifying at least one object or a traffic situation, wherein the first classifier and the at least second classifier are different or differently trained classifiers , Before a classifier selection, at least one boundary condition is determined in a first step. In a second step, a selection of at least one classifier to be used from the classifiers provided is carried out, adapted to the determined boundary conditions. In a third step, the at least one selected classifier is used to classify objects and / or traffic situations. boundary conditions may be in this context all detectable parameters in the environment of the execution of the method. If the method is carried out in a motor vehicle, such boundary conditions may be, for example, information on the outside temperature, the position of the motor vehicle, the light and / or road conditions, individual vehicle parameters, etc. By providing at least two different or differently trained classifiers and the consideration of boundary conditions of the classifier to be used is not rigidly predetermined, but it can be adapted to the corresponding boundary conditions Klassifikatorauswahl. With the method according to the invention can therefore be effectively avoided that under certain conditions unsuitable classifiers are used.

In a preferred embodiment of the method according to the invention, the boundary conditions are determined periodically so that a classifier selection can be made dynamically adapted to the boundary conditions. A periodic determination of the boundary conditions has the advantage, in particular if there are only small time intervals between the individual determinations, that changes in the boundary conditions are detected promptly and can be taken into account directly in the classifier selection. If, for example, a vehicle drives into a tunnel on a sunny day, the lighting conditions change within a very short time, while the other boundary conditions remain essentially constant. If the boundary conditions are determined periodically, this change is immediately recognized within a cycle and taken into account in such a way that - if a classifier designed especially for darkness or better suited for darkness is available - this classifier is selected directly.

In a further preferred embodiment of the method according to the invention, the determination of the boundary conditions is carried out with the aid of auxiliary means arranged in a motor vehicle. Most modern motor vehicles already have, in their basic configuration, a multiplicity of aids which are suitable for providing useful information in relation to a classifier selection. Such tools can be used with almost no additional technical and financial overhead to determine constraints.

A first example of such an aid is a possibly existing state and environment sensor system, in particular an electronic stability program (ESP), a camera, a radar system, the information of a Global Positioning System (GPS) etc. With a state and environment sensor system, for example current position (coordinates) of a vehicle and thus the country are determined in which a vehicle is located. Furthermore, the speed and direction of movement and the current traction of the vehicle can be determined. Such information may be used to the extent that any existing country-specific classifiers are selected based on country-specific markings, left- or right-hand traffic, a country-specific arrangement of traffic signs (eg traffic signs arranged predominantly on the right-hand side of the road), country-specific traffic signs, etc. are.

Further examples of aids arranged in a motor vehicle are telematics and weather services as well as date and / or time information, wherein the term telematics services also includes the use of GPS and digitized maps. With such aids particular features with respect to the weather or in relation to the current day and / or season can be considered.

In a particularly preferred embodiment of the method according to the invention, a classifier selection takes place on the basis of a correlation table, wherein it is determined in the correlation table under which boundary conditions which classifier has the highest performance. This can be done, for example, by establishing a quality measure for the performance in advance and by using tests for each combination of boundary conditions to determine a classifier which has the highest performance under the given boundary conditions. An example of a usable measure of merit is the hit rate of the classifiers, i. the probability that a classifier performs a correct classification under given boundary conditions. The assignment between the detected boundary conditions and the most powerful classifier does not necessarily have to be done by means of a correlation table. Alternative assignment models can also be used.

According to a further embodiment of the method according to the invention, a classification of an object or a traffic situation takes place on the basis of two or more classifiers, wherein the selected classifiers are used sequentially. By the sequential use of classifiers, for example, the advantages of a generalized classifier can be combined with the advantages of specialized classifiers by performing a rough classification by means of a first classifier and then refining the classification with the aid of a second, downstream classifier.

The classifier selection can either be made strictly deterministic or based on a higher-level classification method, such as with the aid of a decision tree or a neural network. In this case, according to a first alternative, it can be specified whether the classifier selection should be made by a superordinate classification method or is performed according to a second alternative depending on the determined boundary conditions.

The second alternative is particularly suitable when it is to be feared that individual boundary conditions may not be clearly determined by the system or can be ascertainable. If, for example, due to contradictory information from two devices, reliable information about the weather situation is not available, a weather situation can be determined on the basis of a neural network, which seems obvious on the basis of the other available information.

On the other hand, neural networks are seldom used in connection with applications involving safety-relevant aspects, since decisions based on neural networks are not clearly predictable, especially in the medium to long term. Accordingly, it is preferred for such applications that a classifier selection should be based on a decision tree or on another deterministic method.

In an alternative embodiment of the method according to the invention, the determined boundary conditions are checked and / or processed before the selection of a classifier to be used. In particular, a control unit may be provided for this purpose which partially or completely checks information about determined boundary conditions and, in particular in the case of contradictory or incomplete information, plausibility of the boundary conditions. For a plausibility check also "blurred" methods, such as. a classification using neural networks.

The invention is also reflected in an inventive device for object and / or situation classification comprising an assistance system with a first classifier and a data input for boundary conditions, wherein the assistance system at least one different from the first classifier and / or one compared to the first classifier differently trained second classifier, wherein a Klassifikatorauswahleinheit is provided, which is adapted to make on the basis of detected via the data input boundary conditions adapted to the detected boundary conditions selecting at least one classifier from the provided classifiers and wherein the at least one selected classifier is usable to perform an object and / or situation classification. On the related to the above The advantages set forth in the inventive method described are hereby referred to.

Further advantageous embodiments and modifications of the invention will become apparent from the dependent claims and from the description in conjunction with the drawings.

Fig. 1

ein System umfassend eine erfindungsgemäße Vorrichtung in einer ersten Ausführungsform sowie

Fig. 2

ein System umfassend eine erfindungsgemäße Vorrichtung in einer zweiten Ausführungsform.

Show it:

Fig. 1

a system comprising a device according to the invention in a first embodiment and also

Fig. 2

a system comprising a device according to the invention in a second embodiment.

Fig. 1 shows a system 100 comprising a first assistance system 120, a second assistance system 140, further assistance systems (in Fig. 1 indicated by four points) as well as an mth assistance system 160. Each of the assistance systems 120, 140, 160 is a device according to the invention for object and / or situation classification.

The invention is explained in more detail below with reference to the second assistance system 140, wherein the second assistance system 140 is used for traffic sign recognition and is part of a motor vehicle (not shown).

The second assistance system 140 has a data input for constraints 142, a Klassifikatorauswahleinheit 144 and a first classifier (K 21) 146, a second classifier (K 22) 148, a third classifier (K 23) 150 and a fourth classifier (K 24) 152 on. The first classifier 146 is a generalized classifier, the second classifier 148 is a classifier specialized in daytime traffic in Germany, the third classifier is a classifier specialized in night traffic in Germany, and the fourth classifier is a specialized classifier for the classifier Traffic outside Germany. The classifier selection unit 144 has an interrogation unit (not shown) via which the second assistance system 140 can permanently query data of the GPS system present in the vehicle as well as date and time information.

If the second assistance system 140 is activated, an interrogation of the GPS about the position of the vehicle and a query about the current date and the current time are carried out with the aid of the interrogation unit. The boundary conditions thus queried are transmitted via the data input 142 to the classifier selection unit 144, which compares the determined boundary conditions with a correlation table.

The correlation table contains for each complete constellation of boundary conditions, i. for each combination of location of the vehicle, date and time an assignment of a particularly suitable for this combination classifier. With the aid of a comparison module (not shown), the classifier selection unit 144 decides based on this correlation table which classifier is used for the classification of the traffic sign.

If a constraint is not available, for example because the GPS does not provide values over the position of the vehicle (within or outside Germany) at the time of determination, the classifier selection unit 144 determines that the first generalized classifier 146 is used in a first step to provide a first coarse classification and using a neural network to determine which position of the vehicle is most likely. On the basis of this first classification, a second classification with the classifier is then made in a second step, wherein the classifier is selected, which is the most powerful due to the information available after the first classification.

As in Fig. 1 The remaining assistance systems 120, 160 are constructed analogously, with the number of classifiers integrated into the respective assistance systems being variable depending on the desired power, as indicated by the dotted lines between K 12 and K 1 n or K m1 and K mn , What is common to the assistance systems, however, is that the first and the m-th assistance systems 120, 160 also have a data input 122 or 162, the data transmission to all assistance systems 120, 140, 160, in particular if they are arranged in spatial proximity to one another a common data line 180, in particular a CAN bus can be made.

This in Fig. 2 System 200 shown comprises a first assistance system 220, a second assistance system 240, further assistance systems (in Fig. 2 indicated by four points) as well as an m-th assistance system 260 and is analogous to that in Fig. 1 shown system 100 constructed. For the same elements are therefore in the in Fig. 2 shown second assistance system 200 by 100 increased reference numerals used.

This in Fig. 2 System 200 shown differs from that in Fig. 1 shown system in that the system 200 additionally includes a control unit 290. This first receives all information about the prevailing boundary conditions and plausibility of these, if contradictory or incomplete information is available. For plausibility check in the in Fig. 2 used system neural networks used. This ensures that the classifier selection units 224, 244, 264 of the assistance systems 220, 240, 260 are always supplied with complete information, since missing information is supplemented with the aid of the control unit 290. Accordingly, with the in Fig. 2 shown assistance systems 220, 240, 260 are dispensed generalized classifiers, if for each constellation of boundary conditions specific classifiers are available and it is ensured that missing by the control unit missing boundary conditions with high probability properly.

Claims (19)

1. Method for the classification of objects and/or traffic situations wherein
   a first classifier (146; 246) and at least one second classifier (148; 248) are provided for the classification of at least one object or traffic situation,
   the first classifier (146; 246) and the at least second classifier (148; 248) being different or differently-trained classifiers,
   characterized by the following steps:

   a) at least one-time determination of initial conditions,

   b) selection of at least one classifier adapted to the determined initial conditions from the classifiers provided (146; 246; 148; 248), and

   c) use of the at least one selected classifier for the classification of objects and/or traffic situations.
2. Method according to Claim 1,
   characterized in that
   the initial conditions are determined periodically, and classifier selection is adapted dynamically to the periodically determined initial conditions.
3. Method according to Claim 1 or Claim 2,
   characterized in that
   the determination of initial conditions is made with the aid of resources present in a motor vehicle.
4. Method according to Claim 3,
   characterized in that
   status and/or surround-field sensing systems of a motor vehicle are used as a resource.
5. Method according to Claim 3 or Claim 4,
   characterized in that
   telematics and/or weather services are used as a resource.
6. Method according to any one of Claims 3 to 5,
   characterized in that
   date and/or time information is used as a resource.
7. Method according to any one of Claims 1 to 6,
   characterized in that
   classifier selection is made on the basis of a correlation table laying down which classifier performs best for given initial conditions.
8. Method according to any one of Claims 1 to 7,
   characterized in that
   classification of an object or traffic situation is made on the basis of two or more classifiers, the selected classifiers being used in sequence.
9. Method according to any one of Claims 1 to 8,
   characterized in that
   classifier selection is made on the basis of an imposed classification method such as a decision tree and/or on the basis of a neural network and/or on the basis of a support vector machine and/or of some other kind of rule-based system.
10. Method according to Claim 9,
    characterized in that
    the choice of whether the classifier selection should be made on the basis of a decision tree and/or on the basis of a neural network is made as a function of the determined initial conditions.
11. Method according to any one of Claims 1 to 10,
    characterized in that
    the determined initial conditions are checked and/or edited prior to selection of the classifier to be used.
12. Method according to Claim 11,
    characterized in that
    a fuzzy method, in particular a classification by means of neural networks, is used for the editing process.
13. Device for object and/or situation classification comprising an assistance system (120; 140; 160; 220; 240; 260) with a first classifier (146; 246), wherein
    the assistance system (120; 140; 160; 220; 240; 260) comprises at least one second classifier (148; 248) differing from the first classifier (146; 246) and/or trained differently from the first classifier,
    characterized in that
    a data input (122; 142; 162; 222; 242; 262) for initial conditions and a classifier selection unit (124; 144; 164; 224; 244; 264) are provided, the classifier selection unit (124; 144; 164; 224; 244; 264) being configured to carry out, on the basis of initial conditions acquired via the data input (122; 142; 162; 222; 242; 262), selection of at least one classifier adapted to the acquired initial conditions from the classifiers provided (146; 246; 148; 248), the at least one selected classifier being usable for an object and/or situation classification.
14. Device according to Claim 13,
    characterized in that
    the classifier selection unit (124; 144; 164; 224; 244; 264) is configured to periodically check the data input (122; 142; 162; 222; 242; 262) for any change in the initial conditions acquired.
15. Device according to Claim 13 or Claim 14,
    characterized in that
    the classifier selection unit (124; 144; 164; 224; 244; 264) comprises an interrogator unit which is configured to automatically access resources for the detection of initial conditions of a motor vehicle and/or the initial conditions acquired by resources of a motor vehicle.
16. Device according to any one of Claims 13 to 15,
    characterized in that
    the classifier selection unit (124; 144; 164; 224; 244; 264) comprises a memory for a correlation table, and a comparison module.
17. Device according to any one of Claims 13 to 16,
    characterized in that
    the classifier selection unit (124; 144; 164; 224; 244; 264) caters for the implementation of a decision tree and/or neural network and/or support vector machine and/or some other kind of rule-based system.
18. Device according to any one of Claims 13 to 17,
    characterized in that
    the classifier selection unit (124; 144; 164; 224; 244; 264) comprises a control unit for checking and/or editing initial conditions acquired.
19. System for object and/or situation classification comprising a number of devices according to any of Claims 13 to 18.

Images