DE102022209561A1 - System for detecting the pedestrian intention ahead of a pedestrian in an environment, vehicle and traffic monitoring system - Google Patents

Abstract

The invention relates to a system (1) for detecting the pedestrian intention of a pedestrian ahead in an environment of the system (1), comprising a sensor system (3) for recording 3D environmental data of objects in the environment as a single image, and an extraction module (4), which for extracting pedestrians from the 3D environmental data as 3D object data, a processing unit (5) being provided which is designed to generate a simulation of a respective pedestrian as a skeleton model from the 3D object data, and a determination unit (14 ) is provided, which is designed to determine at least the posture of a respective pedestrian from the skeleton model, and a behavior estimator (6) is provided, which is designed to determine a pedestrian intention ahead based on the posture.

Description

The invention relates to a system for detecting the pedestrian intention of a pedestrian ahead in an environment of the system, comprising a sensor system for acquiring 3D environmental data of objects in the environment as a single image, and an extraction module which is used to extract pedestrians from the 3D environmental data as 3D -Object data is formed. The invention further relates to a vehicle and a traffic monitoring system with such a system and a method.

Avoiding collisions between vehicles and pedestrians is an important requirement for automated driver assistance systems (ADAS) and future self-driving vehicles. For this purpose, appropriate sensors are usually provided with which the road user is measured.

However, autonomous driving has the challenge of reliably and constantly predicting how road traffic situations will develop in the next 1 to 2 seconds. This task is particularly non-trivial when it comes to predicting pedestrian intentions.

Many previous studies and approaches to solving this problem record past data, such as the speed and trajectory of a pedestrian, based on a sequence of images over the last few seconds in order to predict the near future. If you evaluate the simulation values with the real data, these approaches achieve a reliability of around 70-80%. However, this requires a sequence of images and requires a correspondingly long evaluation.

However, for use in a system of an autonomously operated vehicle, a much higher reliability of the prediction in conjunction with rapid evaluation is necessary.

The DE 10 2013 203239A1 discloses a grid-based method and apparatus for predicting the position of an object in the environment of a vehicle, comprising: providing a first grid comprising cells each representing a portion of the environment; determining the probability for each cell of the first grid that the object or a portion thereof is located in the portion of the environment represented by the respective cell; providing a second grid comprising cells each representing a portion of the environment; Determining the probability for each cell of the second grid that the object or a portion thereof will be in the portion of the environment represented by the respective cell at a predetermined time in the future.

It is an object of the invention to provide an improved and faster system for recognizing the pedestrian intention of a pedestrian ahead. Another task is to specify a vehicle and a traffic monitoring system with such a system as well as a method.

The task is solved by a system with the features of claim 1 and a vehicle with the features of claim 10 and a traffic monitoring system with the features of claim 13 and a method with the features of claim 15.

The subclaims list further advantageous measures that can be suitably combined with one another to achieve further advantages.

The task is solved by a system for detecting the pedestrian intention of a pedestrian ahead in an environment of the system comprising a sensor system for recording 3D environmental data of objects in the environment as a single recording,
and further comprising an extraction module which is designed to extract pedestrians from the 3D environmental data as 3D object data,
wherein a processing unit is provided which is designed to generate a simulation of a respective pedestrian as a skeleton model from the 3D object data,
and a determination unit is provided which is designed to determine at least the posture of a respective pedestrian from the skeleton model,
and a behavior estimator is provided, which is designed to determine a pedestrian intention ahead based on body posture.

Such a sensor system, for example in an autonomous or highly assisted vehicle, can include at least two mono cameras and a radar, which provide the image data and the necessary depths as 3D object data as a single image.

A single image is a single image, for example a single overall image of the environment from the existing sensors; In contrast to this is the sequence of recordings.

The 3D object data can be used to generate a simulation of the pedestrian as a skeleton model the. The determination unit uses this to determine the posture from the skeleton model.

Furthermore, using the behavior estimator, a pedestrian intention ahead, for example going straight, stopping, turning, is generated based on the body posture.

In particular, the skeletal model may not be a precise anatomical model, but rather a schematic skeletal model.

By determining just one skeleton model based on a single image, a quick evaluation of the 3D object data is possible. According to the invention, the body posture can now be used as an empirical biomechanical value to predict with a very high degree of probability which movements a pedestrian will perform in the near future. For example, it is known from biomechanics that with certain postures, for example a certain angle of the upper body or certain positions of the leg, a forward movement is unavoidable and can therefore be clearly determined. Thanks to the current posture of all pedestrians in the area, the movements/intentions of the pedestrians ahead can now be reliably recognized with the necessary speed. This is important, for example, if the system is used in an autonomous vehicle, especially a vehicle with autonomy level four or higher.

Due to the system according to the invention, it is not necessary to record a movement sequence of a pedestrian to determine the pedestrian intention. In particular, the system can also be retrofitted with existing sensor systems, i.e. the extraction module, the processing unit, the determination unit and the behavior estimator can be designed as a software module, which can be easily retrofitted.

In a further embodiment, the skeleton model can be generated from the 3D object data using a neural network, for example a convolutional neural network, which is often used for image processing.

In further training, the behavior estimator is designed as an artificial machine learning method, which is designed to determine a pedestrian intention of a respective pedestrian based on the recognized posture. Such an artificial machine learning method can be, for example, an artificial neural network that has been trained with appropriate training data. These are easy to create, for example, using simulation. In further training, such a neural network can be further trained while driving, for example, and thus constantly improved.

In a further embodiment, the behavior estimator further comprises a storage unit in which different reference pedestrian intentions are stored for different reference postures, wherein the reference pedestrian intentions include at least intended pedestrian position changes, and wherein the behavior estimator comprises an evaluation unit which is designed to compare the posture of the pedestrian with the reference posture Determination of pedestrian intentions ahead.

This makes a quick comparison possible, for example through a simple target/actual comparison with the recognized posture with a reference posture. In particular, the machine learning method and such a behavior estimator can also interact with a storage unit and an evaluation unit, so that, for example, non-stored reference postures are recognized using the neural network, for example, or that postures are assigned to a stored reference posture using the neural network, for example with blurred detection .

In a further embodiment, the reference posture includes at least a cell phone posture or the machine learning method is designed to identify such a cell phone posture based on the recognized posture. Such a cell phone posture usually involves a bent back/tilted head and a raised forearm. Cell phone posture refers to a posture in which the pedestrian looks at the display of a mobile device while walking or standing.

In further development, the system is designed to assign reduced attention to the corresponding pedestrian when determining the cell phone posture. In particular, a vehicle or a driver assistance system or perception system of a vehicle that can be operated at least partially autonomously can react accordingly to the reduced level of attention, for example reducing the driving speed / preparing (emergency) braking maneuvers or drawing attention to itself using a warning tone.

In further training, the system is designed to assign different recognized skeleton model sizes to different pedestrian ages. Such skeleton model sizes can be generated by the processing unit the skeleton models from the 3D object data are automatically supplied. In particular, a threshold value can be set, for example a small height of up to 1.20 m, based on which the pedestrian is assigned the status “child”.
For example, an autonomous vehicle with such an integrated system can react accordingly when such a “child” status is returned, for example reducing the speed, making preparations for braking, etc.

In a further embodiment, the skeleton model comprises at least various skeleton model features, wherein at least the legs, the arms, the torso and the head and the neck are schematically included as skeleton model features, and wherein the behavior estimator is designed to determine the posture of a respective pedestrian based on the skeleton model features . Such skeletal model features can be used to identify fast posture; For example, a precise anatomical representation such as hands or a precise representation of all body elements etc. can be dispensed with, which in turn contributes to the speed of the system.

In particular, a face facing or facing away from the system can be detected. Based on such a facing face, for example, if the system is integrated in a vehicle, it can be recognized whether the pedestrian has recognized/perceived the vehicle. This information can be used, for example, in an autonomously driving vehicle to generate the trajectory, for example with regard to the speed at bus stops/S-Bahn stations, where pedestrians often cross the street inattentively in order to reach the bus, for example.

In a further embodiment, the processing unit is designed to carry out skeleton model recognition that is robust to measurement errors/measurement inaccuracies, for example in the form of a blurred image. The measurement errors can include noise in the measurement data and outliers. For this purpose, special filters such as a reduction filter or median filter, which work in real time, or processing algorithms such as convolutions, which work in real time, can be used. Such measurement errors can increase with distance to the sensor system due to inaccuracies in recording, for example. This means that a simulation of the pedestrian as a skeleton model can be reliably generated from the 3D object data, even from blurred 3D environmental data and 3D object data.
In a further embodiment, the sensor system is designed to capture a sequence of 3D environmental data of objects in the environment as a sequence of individual images, and the extraction module is designed to extract a respective pedestrian as a sequence of the corresponding 3D object data from the 3D environmental data , and wherein the processing unit is designed to generate a sequence of skeleton models from the sequence of 3D object data and wherein the determination unit is designed to determine a sequence of postures from the sequence of skeleton models, and wherein the behavior estimator is designed to do so to determine a pedestrian intention ahead based on the sequence of body postures.

This allows, for example, the accuracy of the prediction to be further increased, for example with regard to a prediction over a longer period of time.

Furthermore, the object is achieved by a vehicle with a system as described above, the vehicle being designed as a vehicle that can be operated autonomously or semi-autonomously. In particular, the vehicle has an autonomy level equal to or higher than four. The system can act as an overall perception framework of an AD (Automated Driving) system and provide input to other modules for the optimal driving strategy of the vehicle that can be operated autonomously or semi-autonomously.

The integration of such a system into the perception and driving function of an autonomously operated vehicle enables, in addition to the pure biomechanical evaluation of the pedestrian's intention, the pedestrian to interact with the vehicle. This means that in situations where a human driver also needs interaction, the autonomously operated vehicle can use the detected posture of the pedestrian to carry out a safe driving maneuver. Examples include situations of a standing pedestrian at a zebra crossing/bus stop/S-Bahn station. Here, using the system integrated in the vehicle, it could be assumed, for example, that the vehicle was not noticed by detecting the facing away, or a safe driving strategy could be generated by detecting the position of the cell phone.
In further training, the vehicle can have a driver assistance system and/or an assisted driving system, which is designed to evaluate pedestrian intentions recognized by the system with regard to a possible collision with the vehicle.

In particular, the driver assistance system and/or the assisted driving system can be designed to respond to a detected collision drive to emit a warning tone and/or initiate collision-avoidance measures such as braking or evasive maneuvers. Therefore, such a system in an autonomously/semi-autonomously operated vehicle increases safety for pedestrians as well as drivers/other road users in particular.

Furthermore, the object is achieved by a traffic monitoring system for traffic monitoring in a predetermined traffic area with a system as described above, wherein the traffic monitoring system is designed to detect vehicles in the traffic area, and wherein the traffic monitoring system is further designed to detect pedestrian intentions recognized by the system with regard to a to evaluate a possible collision with one of the detected vehicles.

Streets with densely parked vehicles in residential areas or confusing intersections already pose a high risk for pedestrians. Children playing can often unexpectedly jump onto the street between parked cars or suddenly cross a street.

With such a system provided, for example in traffic lights/lampposts, the traffic monitoring system can, for example, carry out measures when a risk of collision is detected, for example emit a warning tone/give light signals etc. or send out a V2X signal.

Furthermore, the traffic monitoring system can be designed to take into account signals from traffic control systems in the specified traffic area, for example traffic lights, when evaluating a pedestrian intention recognized by the system with regard to a possible collision with one of the detected vehicles. For example, when switching from green to red at a pedestrian traffic light, a movement forward that is recognized as a pedestrian intention, for example “running off”, can be recognized as a danger.

Furthermore, this information can be transmitted as information to a turning vehicle, for example a truck, so that pedestrians are better protected at dangerous traffic points.

• - Erfassen von 3D-Umfelddaten von Objekten in dem Umfeld als Einzelaufnahme durch ein Sensorsystem,
• - Extrahieren von Fußgängern aus den 3D-Umfelddaten als 3D-Objektdaten durch ein Extraktionsmodul,
• - Generieren einer Simulation eines jeweiligen Fußgängers als Skelettmodell aus den 3D-Objektdaten durch eine Verarbeitungseinheit,
• - Bestimmen zumindest der Körperhaltung eines jeweiligen Fußgängers aus dem Skelettmodell,
• - Bestimmen einer vorausliegenden Fußgängerintension anhand der Körperhaltung durch einen Verhaltensschätzer.

Furthermore, the task is solved by a method for recognizing the pedestrian intention of a pedestrian ahead in an environment comprising the steps:

• - Capturing 3D environmental data from objects in the environment as a single image using a sensor system,
• - Extracting pedestrians from the 3D environmental data as 3D object data through an extraction module,
• - Generating a simulation of a respective pedestrian as a skeleton model from the 3D object data by a processing unit,
• - determining at least the posture of a respective pedestrian from the skeleton model,
• - Determining a pedestrian's intention ahead based on body posture using a behavior estimator.

The advantages and advantageous configurations of the system can also be transferred to the method. In particular, the method is designed to be carried out on the system according to the invention.

• 1: ein erfindungsgemäßes Fahrzeug mit einem erfindungsgemäßen System,
• 2: eine Handy-Haltung,
• 3: verschiedene Skelettmodellgrößen und deren Zuordnung,
• 4: ein erfindungsgemäßes Verkehrsüberwachungssystem,

Further properties and advantages of the present invention emerge from the following description with reference to the accompanying figures. It shows schematically:

• 1 : a vehicle according to the invention with a system according to the invention,
• 2 : a cell phone stance,
• 3 : different skeleton model sizes and their assignment,
• 4 : a traffic monitoring system according to the invention,

1 shows a vehicle 2 according to the invention with a system 1 according to the invention.

The system 1 has a sensor system 3, which is designed to record depth information and 3D/information. For example, the sensor system 3 can include two mono cameras and a radar, which are already installed in autonomous or highly assisted vehicles 2 for recording depth/3D information.

The sensor system 3 captures 3D environmental data of objects in an environment of the vehicle 2 as a single recording. Such objects can be vehicles/pedestrians/houses etc.

There is also an extraction module 4, which is designed to extract pedestrians from the 3D environmental data as 3D object data. For this purpose, the extraction module 4 can be designed as a software module that is executed on a corresponding computer.

Furthermore, a processing unit 5 is provided, which is designed to generate a simulation of the pedestrian as a skeleton model from the 3D Generate object data. In particular, the processing unit 5 can be designed as a software module.

Furthermore, the processing unit 5 can be designed to carry out skeleton model recognition that is robust against measurement errors/measurement inaccuracies. The measurement errors can include noise/blur in the measurement data as well as outliers. For this purpose, the processing unit 5 can use special filters such as a reduction filter or a median filter, which work in real time, or processing algorithms such as convolutions, which work in real time. Such measurement errors/blur can increase, for example, with the distance to the sensor system 3 due to inaccuracies during recording. Using such filters/processing algorithms, a simulation of a respective pedestrian can be reliably generated as a skeleton model from the 3D object data, even from noisy data.

The skeleton model can be very schematic, for example consisting only of simple skeleton model features such as torso, arms, legs, head and neck from which a posture can be reliably recognized. An exact anatomical representation such as hands etc. is not required. Even facial features such as the eyes do not have to be simulated as a skeleton model. It is therefore sufficient to recognize whether the pedestrian has turned his face towards the vehicle 2 or away.

Furthermore, a determination unit 14 is present. This extracts the posture from the skeleton model. In particular, the determination unit 14 can be designed as a software module which, for example, recognizes the posture from the skeleton model using an extraction algorithm.

Using a single individual image and the skeleton model, it is possible to determine the current posture of the individual pedestrian in real time.

Based on the current posture, a pedestrian intention ahead can now be determined using a behavior estimator 6.

For this purpose, the behavior estimator 6 can be designed as an artificial machine learning method. In particular, this can be an artificial neural network, preferably a deep neural network or CNN network, which has been trained in advance to determine a pedestrian intention ahead based on body posture.

The artificial neural network used can use the empirical biomechanical values of “posture” to predict with very high probabilities what intentions a pedestrian will have in the near future. For example, it is known that with certain postures, for example a certain angle of the upper body, a forward movement is unavoidable so that the immediate intention can be clearly determined in real time.

Furthermore, such a neural network can be continuously improved on-board, i.e. during operation, for example by checking later whether the predicted pedestrian intention has occurred, so that ever better pedestrian intentions can be predicted.

Additionally or alternatively, the behavior estimator 6 can also include a storage unit in which different reference pedestrian intentions for different reference postures are stored. The reference pedestrian intentions can include intended pedestrian position changes, such as forward movements.

Furthermore, the behavior estimator 6 can have an evaluation unit for comparing the detected current posture of the pedestrian with the reference posture in order to determine pedestrian intentions ahead. This comparison can, for example, be a simple target/actual comparison. This means that a pedestrian intention can be quickly recognized. Such a storage unit and an evaluation unit can also interact with a neural network, for example for validation or supplementation.

The reference posture can include at least one cell phone posture 9 and/or the machine learning method can be designed to identify such a cell phone posture 9 based on the recognized posture.

2 shows such a cell phone posture 9. Such a cell phone posture 9 includes a bent back 10 and a partially strongly inclined head 11 and a raised forearm 12. The cell phone posture 9 includes a posture with which when walking/or Standing and looking at a display of a mobile device.

The system 1 can be designed to assign reduced attention to such an identified pedestrian. Based on this knowledge, the autonomous/semi-autonomously operable vehicle 2 can now, for example, adapt its driving strategy, such as reducing speed and making braking preparations, in order to avoid a possible collision, taking into account the detected collision to prevent reduced attention. For example, if the risk of collision is too high, further measures can be taken, for example a warning signal can be issued by system 1 or vehicle 2.

Furthermore, the system 1 can be designed to assign different skeleton model sizes to different ages of the pedestrians. For example, skeletal model sizes less than 1 m or 1.20 m can be assigned to children, who can also be assigned reduced attention.

3 shows such an assignment. This shows a child K, on the left of the picture, and an adult woman F with a typical forward movement. Furthermore, typical postures with a walker can be recognized from the skeleton model, for example, which indicate the slow movement of an elderly person M, which is of particular interest for vehicles 2 at intersections with traffic lights, since, for example, the intersection cannot be reached in time when the green light is turned on by elderly people M can be crossed with a walker.

Particularly on busy streets/intersections, the system 1 can also transmit all information to other vehicles via a V2V connection, for example, and thus contribute efficiently to traffic safety.

Based on the recognized posture, it can also be determined whether the pedestrian's face is facing the vehicle 2 or facing away. Especially at transitions, a facing face indicates that the vehicle 2 is registered, while an averted face or a cell phone posture 9 indicates that the vehicle 2 is not registered.

In particular, the system 1 can be integrated into an overall perception frame of an AD system 7 ( 1 ) be integrated in the vehicle 2 and provide other modules of the vehicle 2 with input for an optimal driving strategy.

Furthermore, the Assisted Driving system 7 ( 1 ) based on the pedestrian intentions recognized by the system 1, evaluate them with regard to a possible collision with the vehicle 2 and, if there is a risk of collision, issue a warning tone and / or cause the vehicle 2 to take collision-avoiding measures such as braking preparations.

With such a system 1, a better, higher reliability of the prediction of pedestrian intentions can be achieved in the long term. As a result, the safety of an autonomously or semi-autonomously operated vehicle 2 can be significantly increased, especially in city traffic, for example in front of intersections, zebra crossings, schools, etc.

4 shows a traffic monitoring system 8 for traffic monitoring in a predetermined traffic area with a system 1 according to the invention, the traffic monitoring system 8 being arranged on a traffic light pole 13. The traffic monitoring system 8 is designed to detect vehicles in the traffic area, for example an intersection. Such a traffic monitoring system 8 can be arranged, for example, at busy or confusing intersections.

Such a traffic monitoring system 8 can be designed to evaluate the pedestrian intentions recognized by the system 1 with regard to a possible collision with one of the detected vehicles. If such a collision is imminent, suitable measures can be taken, for example; For example, the traffic monitoring system 8 can use V2X communication to transmit a collision warning to the autonomously or semi-autonomously operated vehicle 2, which can take appropriate measures to prevent a collision.

Furthermore, the traffic monitoring system 8 can be designed to take into account signals from traffic control systems, for example traffic lights, in the specified traffic area when evaluating a pedestrian intention recognized by the system 1 with regard to a possible collision with one of the detected vehicles. This can, for example, prevent children from unexpectedly running into the intersection when the light is red. The traffic monitoring system 8 can also transmit the pedestrian intention to turning vehicles.

Reference symbol list

1

system

2

vehicle

3

Sensor system

4

Extraction module

5

Processing unit

6

Behavior estimator

7

AD system

8th

Traffic monitoring system

9

cell phone posture

10

hunched back

11

tilted head

12

raised forearm

13

Traffic light poles

14

Determination unit

QUOTES INCLUDED IN THE DESCRIPTION

This list of documents listed by the applicant was 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 102013203239 A1 [0006]

Claims (15)

System (1) for detecting the pedestrian intention of a pedestrian ahead in an environment of the system (1), comprising a sensor system (3) for acquiring 3D environmental data of objects in the environment as a single image, and an extraction module (4) which is used to extract Pedestrians from the 3D environmental data are designed as 3D object data, characterized in that a processing unit (5) is provided, which is designed to generate a simulation of a respective pedestrian as a skeleton model from the 3D object data, and a determination unit (14 ) is provided, which is designed to determine at least the posture of a respective pedestrian from the skeleton model, and a behavior estimator (6) is provided, which is designed to determine a pedestrian intention ahead based on the posture. System (1) after Claim 1 , characterized in that the behavior estimator (6) is designed as an artificial machine learning method, which is designed to determine a pedestrian intention of a respective pedestrian based on the recognized posture. System (1) after Claim 1 or 2 , characterized in that the behavior estimator (6) further comprises a storage unit in which different reference pedestrian intentions are stored for different reference postures, the reference pedestrian intentions comprising at least intended pedestrian position changes, the behavior estimator (6) further comprising an evaluation unit which is designed to do so Compare the pedestrian's posture with the reference posture to determine pedestrian intentions ahead. System (1) according to one of the preceding Claims 2 or 3 , characterized in that the reference posture comprises at least one cell phone posture (9) and/or the machine learning method is designed to identify such a cell phone posture (9) based on the recognized posture. System (1) after Claim 4 , characterized in that the system (1) is designed to assign reduced attention to the corresponding pedestrian when determining a cell phone posture (9). System (1) according to one of the preceding claims, characterized in that the system (1) is designed to assign different skeleton model sizes to different ages of the pedestrians. System (1) according to one of the preceding claims, characterized in that the skeleton model comprises at least various skeleton model features, wherein at least the legs, the arms, the torso and the head and the neck are included as skeleton model features, and wherein the behavior estimator (6) is included is designed to determine the posture of a respective pedestrian based on the skeletal model features. System (1) according to one of the preceding claims, characterized in that the processing unit (5) is designed to carry out skeleton model recognition which is robust against measurement errors/measurement inaccuracies. System (1) according to one of the preceding claims, characterized in that the sensor system (3) is designed to record a sequence of 3D environmental data of objects in the environment as a sequence of individual images, and wherein the extraction module (4) is designed to extract a respective one Pedestrian is designed as a sequence of the corresponding 3D object data from the 3D environmental data, and wherein the processing unit (5) is designed to generate a sequence of skeleton models from the sequence of the 3D object data, and wherein the determination unit (14) does so is designed to determine a sequence of postures from the sequence of skeleton models, and wherein the behavior estimator (6) is designed to determine a pedestrian intention ahead based on the sequence of postures. In the 1 Some of the modules shown can also be integrated into other modules. Vehicle (2) with a system (1) according to one of the preceding claims, characterized in that the vehicle (2) is designed as a vehicle (2) that can be operated autonomously or semi-autonomously. vehicle (2). Claim 10 , characterized in that the vehicle (2) has a driver assistance system and / or an assisted driving (AD) system (7), which is designed to detect pedestrian intentions recognized by the system (1) regarding a possible collision with the vehicle (2 ) to rate. vehicle (2). Claim 11 , characterized in that the driver assistance system and/or the assisted driving system (7) is designed to respond to a detected risk of collision to issue a warning tone and/or initiate collision-avoidance measures. Traffic monitoring system (8) for traffic monitoring in a predetermined traffic area with a system (1) according to one of the preceding Claims 1 until 9 , characterized in that the traffic monitoring system (8) is designed to detect vehicles (2) in the traffic area, and wherein the traffic monitoring system (8) is further designed to detect pedestrian intentions recognized by the system (1) regarding a possible collision with one of the to evaluate detected vehicles. Traffic monitoring system (8). Claim 13 , characterized in that the traffic monitoring system (8) is designed to take into account signals from traffic control systems in the specified traffic area when evaluating a pedestrian intention recognized by the system (1) with regard to a possible collision with one of the detected vehicles. Method for detecting the pedestrian intention of a pedestrian ahead in an environment characterized by the steps: - Acquiring 3D environmental data of objects in the environment as a single image by a sensor system (3), - Extracting pedestrians from the 3D environmental data as 3D object data an extraction module (4), - generating a simulation of a respective pedestrian as a skeleton model from the 3D object data by a processing unit (5), - determining at least the posture of a respective pedestrian from the skeleton model, - determining a pedestrian intention ahead based on the posture using a behavior estimator (6).

Images