DE102005043838A1 - Motor vehicle driving maneuver e.g. lane change, detecting method, involves defining bayes networks for specifying probability of presence of maneuver based on preset state variables, and calculating probability of presence of maneuver - Google Patents

Abstract

The method involves defining multiple bayes networks for specifying the probability of the presence of a certain driving maneuver in dependence of predetermined state variables e.g. vehicle speed. The probability of the presence of the driving maneuver is calculated with the bayes network through input of the detected or derived state variables in the bayes network. A self-learning adaptation of a local probability distribution taken place based on a respective driver of a vehicle. An independent claim is also included for a device for automatic detection of driving maneuver of a vehicle.

Description

The The invention relates to a method and a device for automatic Detection of driving maneuvers of vehicles by means of detected state variables.

In Today's vehicles are a variety of sensors for available with which state variables are detected can be which are sometimes unsafe or invalid values. The sensors will for example, for track recognition, detection of line of sight of the driver, operation a direction indicator, brake pedal operation, accelerator operation, etc. used.

On the basis of these and possibly further internal or external state variables with known methods driving maneuvers detected to control a driving simulator or in the vehicle control by the driver intervene. A classic example of this is the anti-lock braking system.

In US 2002/0128751 is the use of neural networks for Classification of the driving situation described. This is the current Driving situation (large-scale situation), d. H. driving on the Highway, a federal highway, a bypass or recognized in city traffic, but not current driving maneuvers. It can also the driver state, such. B. tired distracted, etc. recognized who the. The classification scheme is learned through empirical data. Previous Know about the detection of driving maneuvers can not be programmed. The classified driving situations can subsequently be used in an overarching Bayesian network for to be further processed for global purposes.

Out Nuria Oliver and Alex P. Pentland: "Graphical Models for Driver Behavior Recognition in a Smart Car. "In: Proceedings of SPIE Aerosense 2000, Enhanced and Synthetic Vision ', Orlando, Florida, April 2000 is the use of a hidden Markov model for detection of driving maneuvers described. The data in hidden Markov models are processed sequentially. Contextual information, such as viewing direction, relative lane position, speeds and directions of traffic have to manually for every frame of a video file can be noted down to some driving maneuvers Turn off and recognize lane change. This is a real-time application hardly possible.

In Kari Torkkola, Srihari Venkatesan and Huan Liu "Sensor Sequence Modeling for Driving ", In: Proceedings of the 18th International Florida Artificial Intelligence Research Society Conference FLAIRS, Clearwater Beach, FL, USA, 2005. Pages 721-727 describes the use of hidden Markov models for detecting maneuvers. The system is for used a driving simulator, with the entered into the system State variables of a set of current and future come from simulated sensors. Simulated sensors usually have perfect values. A system that expects perfect values is for real vehicle data not suitable.

task The present invention is an improved method and an improved device for automatically recognizing driving maneuvers create even with missing and / or unsafe information in real time an analysis of the available state variables under consideration different driving styles of different drivers perform.

• – Definieren eines aus Knoten und Kanten gebildeten Bayesschen Netzwerks zur Beschreibung der Wahrscheinlichkeit des Vorliegens bestimmter Fahrmanöver in vorgegebener Abhängigkeit von Zustandsgrößen, bei dem mit einer lokalen Wahrscheinlichkeitsverteilung initialisierte Knoten eine Zustandsgröße und die Kanten kausale Abhängigkeiten zwischen den Knoten repräsentieren, und
• – Berechnen der Wahrscheinlichkeit des Vorliegens von Fahrmanövern mit dem Bayesschen Netzwerk durch Eingabe detektierter oder davon abgeleiteter Zustandsgrößen in das Bayessche Netzwerk.

The object is achieved by the method of the type mentioned, by

• Defining a Bayesian network formed by nodes and edges for describing the probability of the existence of certain driving maneuvers in predetermined dependence on state variables, in which nodes initialized with a local probability distribution represent a state variable and the edges represent causal dependencies between the nodes, and
• Calculating the probability of the presence of driving maneuvers with the Bayesian network by entering detected or derived state variables in the Bayesian network.

With the aid of the Bayesian network predefined for detecting individual driving maneuvers as a function of detectable state variables, current driving maneuvers can be automatically recognized in real time, even if the state variables are uncertain due to uncertain and / or missing values of sensors. Using a Bayesian network specialized in recognizing individual driving maneuvers also has the advantage of pre-programming dependencies into the network can. On this basis, the probability of a correct detection of the driving maneuver is provided on the basis of state variables determined with or derived from the sensors and possibly their quality together with the recognized maneuver.

Especially It is advantageous if a self-learning adaptation of the local Probability distributions occur. By the possibility a Bayesian network for continuous adaptation can this learn from existing driving situations. If the Bayesian network Training with different drivers can have different parameters for different Driver to be learned. It is advantageous if the probability distributions, the dependent continuously adapted by the respective driver of the vehicle be stored as driver-related parameters.

The Bayesian network preferably has main nodes, also parent nodes called, for the selection of state variables for factors which influence the driver when selecting a driving maneuver. Farther can Secondary nodes, also called child nodes, are provided for additional state variables, as a result of driving maneuvers arise. Other parent and child nodes may be the original one Be added to the definition of the Bayesian network, for example if further sensors in the future available be made.

When maneuvers can for example, lane change, overtaking, Turn off, motorway exit and / or highway entrance are detected.

State variables can, for example the operating condition of direction indicators, Vehicle speed, acceleration, road type, number of available lanes etc. be. Other state variables may be e.g. derived from the current vehicle position, such as the street type, Number of lanes, etc.

• – Sensoren zur Erfassung von Zustandsgrößen,
• – Datenspeicher zur Speicherung eines aus Knoten und Kanten bestehenden Bayesschen Netzwerkes zur Beschreibung der Wahrscheinlichkeit des Vorliegens bestimmter Fahrmanöver in vorgegebener Abhängigkeit von Zustandsgrößen, bei dem mit einer lokalen Wahrscheinlichkeitsverteilung initialisierte Knoten eine Zustandsgröße und die Kanten kausaler Abhängigkeiten zwischen den Knoten repräsentieren, und
• – eine Auswerteeinheit zur Berechnung der Wahrscheinlichkeit des Vorliegens von Fahrmanövern mit dem Bayesschen Netzwerk durch Eingabe detektierter oder davon abgeleiteter Zustandsgrößen in das Bayessche Netzwerk.

The object is further achieved by the device of the type mentioned by

• - sensors for the acquisition of state variables,
• Data storage for storing a node-and-edge Bayesian network for describing the probability of having certain driving maneuvers in predetermined dependence on state variables, where nodes initialized with a local probability distribution represent a state variable and the edges of causal dependencies between the nodes, and
• - An evaluation unit for calculating the probability of the existence of driving maneuvers with the Bayesian network by entering detected or derived state variables in the Bayesian network.

The Evaluation unit is preferably for self-learning adaptation of Probability distributions established in relation to the Nodes are stored in the Bayesian network in the data store.

The Device can be used in driver assistance systems or in these to get integrated. By detecting current or upcoming maneuvers can provide the driver with adequate support for. B. by intervention be given to the vehicle control. If the driver is a complicated one or critical maneuver performs, can unimportant indications of on-board devices in the vehicle suppressed or be moved.

The Recognition of current and upcoming maneuvers allows maneuver-specific assistance, eg. B. overtaking assistance, either as partial automation or in the form of a warning can be offered to the driver, for. B. when carrying out a maneuver can lead to a critical situation.

to Derivation of further state variables, such as z. Road type, Number of lanes, speed restrictions, etc., can be data storage be provided with a database in which the state variables passable streets are recorded with assigned location positions. The evaluation unit can then with the help of the current vehicle position, the z. B. with Satellite navigation was determined, and the database from the Vehicle position to close the other state variables.

The Invention is described below with reference to the accompanying drawings with a Embodiment explained in more detail. It demonstrate:

1 Block diagram of a device for the automatic detection of driving maneuvers of vehicles;

2 Diagram of an exemplary Bayesian network;

3 - Probability table of the Bayesian network for the relationship between driving maneuver and state of a turn signal;

4 - Section of a probability table of the Bayesian network for the relationship between the last detected maneuver, vehicle position and driving maneuver.

The 1 leaves a block diagram of a device 1 to detect driving maneuvers of vehicles using a Bayesian network. The device has an evaluation unit 2 that with a data store 3 communicates. In the data memory, the Bayesian network for determining a probability of the existence of certain driving maneuvers in dependence on state variables is stored. The state variables are via a vehicle bus 4 from sensors 5a . 5b , ... 5n to the evaluation unit 2 transfer. As state variables it is also possible to use state variables not detected by sensors or derived from other state variables. The state variables can be detected in the vehicle itself internally or externally. The term "detected state variables" thus means all data which are directly or indirectly determined or derived internally or externally by sensors and which describe any state of the vehicle or of the driver (eg viewing direction, fatigue, etc.).

To derive further state variables, such as type of road, number of lanes, speed restrictions, etc., the data memory 6 be provided with a database in which the state variables of drivable roads are recorded with assigned location positions. With the help of the current vehicle position and the database in the data memory 6 can then from the current vehicle position using the evaluation 2 be concluded on the other state variables. These are thus determined indirectly. The data store 6 can be separate to the data store 3 running or simply a data storage area in the data store 3 be.

The evaluation unit 2 For example, by suitable programming of a processor system, it is set up in such a way that, depending on the state variables and the Bayesian network, the probability of the existence of defined driving maneuvers is recognized. This will be explained in more detail below.

The Data read from the sensors provides a set from incomplete and uncertain evidence. This data set must be evaluated to the current driving maneuvers to recognize.

Bayesian Methods provide a formal framework for inference partial evidence. The validity A hypothesis can be used with Bayesian methods in the evaluation be tested by uncertain evidence.

Accepted becomes a hypothesis "A": "The current one maneuvers is a lane change "and a set of evidences "e" that are used can, about the validity hypothesis "A" Evidence "e" represents for example the fact "The Turn signal is on ".

The conditional probability P (A / e) of hypothesis A given evidence "e" can be calculated using the formula of Thomas Bayes:

Conversely, it is known:

umgeschrieben werden.The equation (1) can therefore be considered

be rewritten.

An example is the recognition of the driving maneuver "lane change" on the basis of the likelihood of the presence of a lane change upon actuation of the turn signal (state variable, left turn signal on / off). This is expressed by:

The two terms on the right side of the equation are both now statistically determined values.

Assume a set of independent evidence e1, e2 ... en. Then the following equation system can be defined for each driving maneuver:

This represents the theoretical basis of a Bayesian network. It will be a Model of the domain maneuvers as a directed, acyclic graph. Every node in represents the network a variable of interest in the domain. The edges of the graph represent causal dependencies between the nodes.

Everyone Node is initialized with a local probability distribution. If evidence over one or more nodes is obtained, the global probability distribution of the model recalculated.

The Network is capable of learning. Any spread of information through the network can lead to a Update the probability distribution. Even if the original one Probability distribution of only a rough estimate of the true distribution provides, can the real probability distribution be learned quickly from the model.

The Network can for different drivers are trained. The different ones Distributions are learned.

The System for the automatic detection of driving maneuvers has as its core a Bayesian Network. The recognized driving maneuver is represented by one / more nodes. Supportive evidence of The sensors will be through children's nodes represents. Relevant factors for the selection of driving maneuvers influence, for example, the last recognized driving maneuver and GPS information about the street type, Number of tracks, etc. are used as parent nodes in the network brought in. If new sensors are available, relevant Information from these sensors can be added as new child nodes. When new factors for the selection of driving maneuvers available are (for example, driver state), the new information be added as a node in the network.

The 2 allows a diagram of an exemplary Bayesian network to determine a driving maneuver (driving maneuver) in dependence of the following state variables detect, which are derived in part from the current vehicle position (location, eg with satellite positioning GPS, GALILEO, etc.) derived:
Status of the direction indicator ("turn signal off", "turn signal left", "turn signal right"), steering angle (Steering), vehicle speed, lane, distance to the vehicle in front, previously detected driving maneuver (previous maneuver).

The 3 shows a probability table for a Bayesian network, by which the probability of the existence of a particular maneuver is defined by a fixed dependency tree with nodes and edges and the probabilities associated with the nodes. The probabilities are defined in the example of the relationships of the likelihood of driving maneuvers and the state of a turn signal ("turn signal off"("nil"),"left turn signal" (left), "turn signal right" (right).) The driving maneuvers are:
Vehicle stopped (Stopped), Vehicle follows the road (Follow Road), Vehicle follows a preceding vehicle (Follow Car), Turn Left, Turn Right, Thread Left (Merge Left), Thread Right (Merge Right), Overtake, Merge onto Freeway, Exit Freeway, Cross Intersection.

The 4 shows a section of a probability table for the Bayesian network with relationships of the probability of previously detected driving maneuvers, selected vehicle positions and the above enumerated driving maneuvers. The vehicle positions are driving on the road (road), intersection (intersection) or ramp of a motorway or highway (ramp).

With a compilation of such probability tables and the system of equations ( 4 ) the Bayesian network is defined, so that from uncertain or missing state variables probabilities of the presence of the predetermined driving maneuvers from the system of equations ( 4 ), the state variables and the probabilities.

Claims (10)

Method for automatic recognition of driving maneuvers of Vehicles with the help of detected state variables, characterized by - Define a Bayesian network formed by knots and edges for Description of the likelihood of certain driving maneuvers being given dependence of state variables, at initialized with a local probability distribution Node a state size and the Edges causal dependencies between the nodes, and - To calculate the likelihood of driving maneuvers the Bayesian network by entering detected or derived State variables in the Bayesian network. Method according to claim 1, characterized by self-learning Adapting the local probability distributions. Method according to claim 2, characterized in that that the self-learning adaptation depending on the respective Driver of the vehicle takes place and the probability distributions as driver-related parameters are stored. Method according to one of the preceding claims, characterized characterized in that the Bayesian network main node (parent node) for the selection of state variables for factors which influence the driver in the selection of a driving maneuver and secondary nodes (Children's knot) for Has state variables, as a result of driving maneuvers arise. Method according to one of the preceding claims, characterized by recognizing lane changes, overtaking, turning, motorway ramps and / or motorway exits as driving maneuvers. Method according to one of the preceding claims, characterized characterized in that state variables of Operating state of a direction indicator, braking action, Vehicle speed, acceleration, vehicle position, road type, Viewing direction and / or number of available lanes and / or are state variables derived from the vehicle position. Facility ( 1 ) for the automatic detection of driving maneuvers of vehicles characterized by - sensors ( 5 ) for the acquisition of state variables, - data memory ( 3 ) for storing a Bayesian network formed of nodes and edges for describing the probability of having certain driving maneuvers in predetermined dependence on state quantities, in which nodes initialized with a local probability distribution represent a state variable and the edges represent causal dependencies between the nodes, and one with Sensors ( 5 ) and the data memory ( 3 ) associated evaluation unit ( 2 ) to Be calculation of the likelihood of driving maneuvers with the Bayesian network as a function of those with the sensors ( 5 ) detected state variables as input variables of the Bayesian network. Facility ( 1 ) according to claim 7, characterized in that the evaluation unit ( 2 ) is set up for self-learning adaptation of the probability distributions. Facility ( 1 ) according to claim 7 or 8, characterized in that data storage ( 6 ) is provided for storing a database with data for describing road types, number of lanes and / or speed limits of roads with assigned location positions and the evaluation unit ( 2 ) is set up to calculate the road type, the number of lanes and / or speed limits as state variables using the database and the current vehicle position. Facility ( 1 ) according to one of claims 7 to 9, characterized in that the evaluation unit ( 2 ) is set up to perform a calculation even if partially missing or unsafe information.