DE102023005203A1 - Sensor fusion method - Google Patents

Abstract

The invention relates to a method for sensor fusion, wherein a vehicle (1) detects its surroundings using at least a first (2.1) and a second sensor (2.2), the first and second sensors (2.1, 2.2) having an at least partially overlapping detection area ( 3), and wherein a computing unit (4) executes a machine learning model for processing sensor data generated by the sensors (2.1, 2.2). The method according to the invention is characterized by the following method steps carried out by the computing unit (4): a) determining one of the sensors (2.1, 2.2) as the main sensor, the computing unit (4) using the knowledge gained from an analysis of the sensor data generated by the main sensor to control a vehicle function, and wherein the computing unit (4) analyzes the sensor data of the main sensor using the machine learning model; b) defining at least one sensor (2.1, 2.2) that deviates from the main sensor as a secondary sensor; ;d) selection of a relevant information area (5) from the sensor data generated by the main sensor depending on the sensor data generated by the at least one secondary sensor by the machine learning model; ande) analyzing the part of the sensor data of the main sensor assigned to the relevant information area (5) to obtain the insights through the machine learning model.

Description

The invention relates to a method for sensor fusion according to the type defined in more detail in the preamble of claim 1.

With increasing digitalization, more and more assistance systems are being integrated into vehicles. This makes it possible to provide a variety of safety and comfort functions. This makes it possible to control a vehicle at least partially automated or even autonomously. This requires the vehicle to detect the surroundings using environmental sensors.

With the help of environmental sensors, the vehicle is able to detect static and dynamic surrounding objects as well as the course of the road. This makes it possible to determine control commands for safe control of the vehicle depending on the situation. For this purpose, the vehicle can have a wide variety of sensors, such as one or more mono and/or stereo cameras, ultrasonic sensors, radar sensors, LiDARe and the like. Each sensor type has a different detection quality for different operating conditions. Images from cameras attached to the vehicle allow particularly differentiated recording and analysis of the vehicle's surroundings. However, in adverse visibility conditions, for example when the sun is low, fog, darkness and the like, visibility is restricted. The detection range of a LiDAR can be reduced due to heavy rain and the like. In order to be able to adequately capture the vehicle's surroundings regardless of the situation, the sensor data from different sensors are merged with one another.

The sensor data is usually merged at the so-called object level. The objects detected using the information from one sensor are merged with the information from the other sensors. However, this approach is inefficient and can create problems. In addition, each type of sensor requires its own methodology for object recognition, which requires the use of comparatively powerful computing hardware.

Artificial intelligence, particularly in the form of artificial neural networks, can also be used for sensor data processing.

From the WO 2016/100814 A1 multimodal sensor data fusion for perceiving systems is known. The publication describes the detection of a positional deviation between two different sensor systems using artificial intelligence. This makes it possible to determine to what extent the detection range of a first and a second sensor differ from one another. To train the artificial intelligence, the AI is supplied with corresponding sensor data for various test cases as well as the positional deviation present. This will gradually enable the AI to correctly estimate the positional deviation between the sensors for various boundary conditions. A position deviation determined by the Kl can be used to control a control actuator of a respective sensor in order to align the two sensors to the same detection area to reduce the position deviation.

The present invention is based on the object of specifying an improved method for sensor fusion, which is characterized by increased efficiency and improved detection reliability.

According to the invention, this object is achieved by a method for sensor fusion with the features of claim 1. Advantageous refinements and further developments result from the requirements that depend on this.

1. a) Festlegung eines der Sensoren als Hauptsensor, wobei die Recheneinheit die aus einer Analyse der vom Hauptsensor generierten Sensordaten gewonnenen Erkenntnisse zum Ansteuern einer Fahrzeugfunktion verwendet, und wobei die Recheneinheit die Sensordaten des Hauptsensors unter Einsatz des Maschinenlernmodells analysiert;
2. b) Festlegen wenigstens eines vom Hauptsensor abweichenden Sensors als Nebensensor;
3. c) Zuführen von dem wenigstens einen Nebensensor erzeugten Sensordaten dem Maschinenlernmodell;
4. d) Auswahl eines relevanten Informationsbereichs aus den vom Hauptsensor erzeugten Sensordaten in Abhängigkeit der von dem wenigstens einen Nebensensor generierten Sensordaten durch das Maschinenlernmodell; und
5. e) Analysieren des dem relevanten Informationsbereich zugeordneten Teils der Sensordaten des Hauptsensors zur Gewinnung der Erkenntnisse durch das Maschinenlernmodell.

A generic method for sensor fusion, wherein a vehicle detects its surroundings by means of at least a first and a second sensor, the first and second sensors having an at least partially overlapping detection range, and wherein a computing unit executes a machine learning model for processing sensor data generated by the sensors, is further developed according to the invention by the following method steps carried out by the computing unit:

1. a) determining one of the sensors as the main sensor, wherein the computing unit uses the knowledge gained from an analysis of the sensor data generated by the main sensor to control a vehicle function, and wherein the computing unit analyzes the sensor data of the main sensor using the machine learning model;
2. b) defining at least one sensor that deviates from the main sensor as a secondary sensor;
3. c) supplying sensor data generated by the at least one secondary sensor to the machine learning model;
4. d) selecting a relevant information area from the sensor data generated by the main sensor depending on the sensor data generated by the at least one secondary sensor by the machine learning model; and
5. e) analyzing the part of the sensor data of the main sensor assigned to the relevant information area to obtain the insights by the machine learning model.

The sensor fusion method according to the invention therefore makes it possible to carry out the fusion of the various sensors not only at the object level, but rather beforehand on the basis of the available raw data. This increases the efficiency of sensor data fusion, since the sensor data generated by the various sensors does not have to be analyzed in detail for object detection.

To do this, one of the vehicle's sensors is selected as the main sensor. Based on the sensor data generated by the main sensor, the machine learning model determines the information relevant for further processing in the vehicle by driver assistance systems. These can, for example, be control commands for at least partially automated or even autonomous control of the vehicle. The vehicle can therefore, in particular, carry out an automated intervention in the longitudinal and/or transverse guidance. It is particularly advantageous to select the sensor system or sensor type as the main sensor that allows the most comprehensive detection and evaluation of the vehicle's surroundings in the respective situation.

At least one of the further sensors of the vehicle with an at least partially overlapping detection range to the main sensor, preferably as many further sensors as possible, are then determined as a secondary sensor. The machine learning model then determines relevant areas of the sensor data from the sensor data generated by the secondary sensors or in the sensor data of the main sensor, which should be processed or analyzed particularly deeply or extensively for further analysis. To put it colloquially, the secondary sensors enable the machine learning model to focus on exactly that part of the area of the environment detected by the main sensor that is relevant to the current driving situation. This means that not all data generated by the main sensor needs to be analyzed extensively, but only the relevant data. This allows the processing effort required to detect and evaluate the environment for the vehicle to be reduced, which ultimately improves efficiency.

For example, a front camera of the vehicle can be selected as the main sensor. An event-based camera and one or more microphones can act as secondary sensors. By analyzing sensor data generated by the corresponding secondary sensors, the machine learning model is able to recognize another vehicle in the vicinity of the ego vehicle. For example, engine noises emitted by the other vehicle can be located relative to the ego vehicle using a microphone array. Due to the known installation position of the respective sensors on the vehicle, the machine learning model can determine in which image section from the front camera selected as the main sensor the respective additional vehicle will be present. Accordingly, the machine learning model only analyzes this section of the camera image in detail.

The machine learning model can be based in particular on the use of so-called transformer-based learning models. Such a transformer architecture in particular uses a so-called attention mechanism, also referred to as an “attention mechanism”. These enable the input of data that quantifies the relevance of other information and thus implicitly a colloquial focus on part of the input data. As a result, during subsequent processing by a machine learning model, an implicit sensor fusion is carried out, which results in parts of the main sensor's input data being weighted by information from other sensors.

To determine the relevant sub-area of the sensor data of the main sensor, not only can the corresponding area be cut out of the sensor data, but a confidence value can also be determined for relevant areas, with those areas being determined as an interesting sub-area whose confidence value is above a predetermined confidence value threshold. The respective confidence values are determined by the machine learning model depending on the sensor data generated by the secondary sensors. A relevant sub-area determined for the sensor data of the main sensor can be contiguous or can also be divided into separate subsections.

In order to enable the machine learning model to take on the tasks assigned to it, the machine learning model was trained before the actual operating phase, for example by the vehicle manufacturer. The training data may have been generated in extensive test campaigns. In particular, it can be supervised learning. Such a training data set describes which subareas of the sensor data of the main sensor are relevant for different sensor systems. In addition, a training data set can contain instructions on how the corresponding data should preferably be processed should. Furthermore, a training data set can also contain a specification as to which sensor system should function as the main sensor and as the secondary sensor.

• Im Fahrzeug seien fünf verschiedene Sensoren verbaut, die in unterschiedlicher Konfiguration ihre jeweiligen Sensordaten erfassen können. Beispielsweise sind ein LiDAR sowie eine jeweilige Time-Of-Flight- und Infrarot-Kamera vorgesehen. Die Trainingskampagne besteht aus einer Vielzahl an Szenarien. Diese Szenarien setzen sich zusammen aus einer Kombination unterschiedlicher Sensoreinstellungen und Szenarien-Eigenschaften sowie dem Szenarien-Label. Die Sensoreinstellungen definieren die unterschiedlichen Konfigurationen, die ein Sensor im Fahrzeug annehmen kann. Beispielsweise sind dies für eine Kamera: Kamerawinkel, Tiefenmessung, Beleuchtung, Schärfe, Auflösung, etc. Die Szenarien-Eigenschaften sind beispielsweise:
  • Geschwindigkeit, Licht- und Sichtbedingungen, Fahrbahnzustand, Fahrwerksdämpfung, Beladung und dergleichen. Auf Basis der Kombination unterschiedlicher Sensor- und Szenarien-Eigenschaften können die jeweiligen Szenarien-Label in der Trainingskampagne vergeben und somit die Sensor-Fusion optimiert werden. Szenarien-Label können hierbei beispielsweise lauten: „Kurvenlagen-Güte“, „ESP-Eingriff“, „Unfallvermeidung“, „Geschwindigkeitsreduktion“, „Ausweichmanöver“. In einer konkreten Sensor- und Szenarien-Eigenschaftskombination ergeben sich beispielsweise während des Trainings Trefferquoten für die Label-Korrektheit (supervised learning) wie „Kurvenlagen-Güte [80%]“, „ESP-Eingriff [90%]“, „Unfallvermeidung [10%]“, „Geschwindigkeitsreduktion [20%]“, „Ausweichmanöver [98%]“. Je höher die Genauigkeit bzw. Trefferquote ist, desto besser ist die Sensorfusion, aus der abgeleitet werden kann, welche die optimale Verknüpfung der Sensor-Eingangsdaten im späteren Serienbetrieb ist.

To illustrate this, here is an example of a training campaign that is carried out during development:

• Five different sensors are installed in the vehicle, which can record their respective sensor data in different configurations. For example, a LiDAR as well as a respective time-of-flight and infrared camera are provided. The training campaign consists of a variety of scenarios. These scenarios are made up of a combination of different sensor settings and scenario properties as well as the scenario label. The sensor settings define the different configurations that a sensor can assume in the vehicle. For example, for a camera these are: camera angle, depth measurement, lighting, sharpness, resolution, etc. The scenario properties are, for example:
  • Speed, light and visibility conditions, road condition, chassis damping, load and the like. Based on the combination of different sensor and scenario properties, the respective scenario labels can be assigned in the training campaign and the sensor fusion can thus be optimized. Scenario labels can be, for example: “Curving quality”, “ESP intervention”, “Accident avoidance”, “Speed reduction”, “Avoidance maneuvers”. In a specific sensor and scenario property combination, for example, hit rates for label correctness (supervised learning) arise during training such as “curve position quality [80%]”, “ESP intervention [90%]”, “accident avoidance [10 %]", "Speed reduction [20%]", "Evasion maneuver [98%]". The higher the accuracy or hit rate, the better the sensor fusion, from which it can be derived which is the optimal combination of the sensor input data in later series operation.

An advantageous development of the method according to the invention further provides that the computing unit discards the part of the sensor data of the main sensor that deviates from the relevant information area before analysis by the machine learning model. As already mentioned, the interesting section of the sensor data from the main sensor can be cut out or removed. The amount of data to be processed by the computing unit is thus reduced, which further increases the efficiency in carrying out the method according to the invention.

According to a further advantageous embodiment of the method according to the invention, the sensor data generated by the at least one secondary sensor are processed before processing by the machine learning model. This increases the data quality of the corresponding sensor data. For example, noise can be removed using appropriate preprocessing algorithms, special filters can be used, and the like. This increases the reliability that relevant information can be obtained from the sensor data generated by the secondary sensors. The data assigned to the interesting subarea of the sensor data of the main sensor can be prepared accordingly.

A further advantageous embodiment of the method according to the invention further provides that the sensor data generated by the at least one secondary sensor are standardized before processing by the machine learning model. This improves computing efficiency even further. In particular, the sensor data generated by different sensor systems can be brought into a uniform format, transferred to the same coordinate system, and the like. This makes it possible to prevent errors from occurring when referencing the respective data to one another due to different unit systems, length ratios and the like.

According to a further advantageous embodiment of the method according to the invention, at least one detection parameter of the main sensor is adapted depending on knowledge gained through an analysis of the sensor data of at least one secondary sensor. From the knowledge gained by the machine learning model by analyzing the sensor data of the secondary sensors, it can also be deduced that, for example, there is a particular operating condition in which the detection quality of the main sensor can be increased if special detection parameters are set for the main sensor. Analogous to previous statements, the machine learning model may have been trained in the development phase to recognize such relationships. In this way, different test scenarios can be carried out, with a set of different detection parameters being set for different detection boundary conditions and respective detections of the environment being carried out. The machine learning model thereby learns to recognize for which detection boundary conditions and with which detection parameters the best results can be achieved. Is this the main sensor? a camera, for example the aperture, exposure time, ISO sensitivity, focal length and the like can be varied for a shot.

A further advantageous embodiment of the method according to the invention further provides that the machine learning model takes into account sensor data generated by at least two secondary sensors to select the relevant information area, with the secondary sensors being assigned a relevance weighting depending on the current detection conditions, and the machine learning model sensor data generated by secondary sensors with a high relevance weighting prioritized for selecting the relevant information area. Like the main sensor, the respective secondary sensors also have detection qualities that depend on different operating conditions. If such a secondary sensor is used to determine the relevant information area in the sensor data generated by the main sensor, which has a reduced detection quality under the current operating conditions, errors can arise. For example, if there is currently heavy rain, a LiDAR sensor could provide a distance value that is too short. There are then different relevance weightings for the different sensor types for different operational conditions. These relevance weights can be assigned to the respective secondary sensors, for example, using a heuristic formula or using artificial intelligence. The machine learning model then takes into account the sensor data from the secondary sensors according to the respective relevance weighting. This increases the reliability that the relevant information range for the sensor data of the main sensor is correctly determined.

According to a further advantageous embodiment of the method according to the invention, in step a) the sensor which has the highest detection quality for the current detection boundary conditions is defined as the main sensor. This increases the reliability that relevant objects are recognized by the machine learning model. Analogous to previous statements in connection with the secondary sensors, a heuristic formula and/or artificial intelligence can also be used here to determine which sensor has the highest detection quality given the current detection boundary conditions.

A further advantageous embodiment of the method according to the invention further provides that the detection quality of the individual sensors is tracked over time and if the detection quality of a specific sensor deteriorates, a warning message is issued to a vehicle occupant. The machine learning model is therefore able to record the degree of wear of the vehicle's different sensors. If the degradation of a specific sensor has progressed to such an extent that reliable environmental detection is no longer possible, a corresponding warning message can be issued. This warning message can be visual, acoustic and/or tactile. For example, a warning light in the vehicle interior can flash, an advisory text can be displayed on a display device, a warning tone can sound and/or a vehicle element touched by the person driving the vehicle, for example a seat and/or the steering wheel, can vibrate. This makes it possible to convey to the vehicle occupant, in particular the person driving the vehicle, that the corresponding sensor should be replaced. The vehicle can be serviced, for example, in an authorized workshop of the vehicle manufacturer. A function can be provided via a corresponding human-machine interface in the vehicle to book corresponding workshop appointments directly from the vehicle.

According to a further advantageous embodiment of the method according to the invention, the computing unit maintains at least two machine learning models that differ in their degree of complexity and uses the less complex machine learning model to analyze the sensor data of the main sensor if, depending on the sensor data of the at least one secondary sensor, there is no relevant information area in the main sensor generated sensor data can be identified. This allows the efficiency of data processing by the computing unit to be increased even further. For example, if the machine learning models are artificial neural networks, they can differ in their depth. The different artificial neural networks therefore have a different number of neurons. The more neurons an artificial neural network contains, the higher the level of complexity. As the level of complexity increases, the computational effort required to execute the corresponding machine learning model also increases. If an artificial neural network with fewer neurons or a smaller depth is used for data analysis, the required computing effort and thus the energy consumption of the computing unit also decreases.

A further advantageous embodiment of the method according to the invention further provides that at least a third sensor has a detection range that deviates from the main sensor, an object detected by the main sensor and identified by the computing unit in the sensor data of the main sensor by the computing unit is tracked and the computing unit determines the third sensor as the new main sensor when the computing unit predicts that the object will pass from the detection range of the main sensor into the detection range of the third sensor. This allows the efficiency of the sensor fusion method according to the invention to be increased even further. In this way, the appropriate sensor can be quickly selected as the main sensor for the respective operating conditions. If a relevant object leaves the detection range of the current main sensor and enters the detection range of a said third sensor, the respective sensor can be promptly selected as the new main sensor, so that the corresponding relevant object can continue to be tracked by the machine learning model or the computing unit with sufficient detection quality can.

In general, it is also possible for several sensors in the vehicle to function simultaneously as the main sensor and/or secondary sensor, depending on the operating conditions.

For example, a front camera aligned in the direction of the vehicle's longitudinal axis and an event-based camera that is offset therefrom and directed laterally away from the center axis of the vehicle can function as two main sensors at the same time. The front camera monitors a first object and the event-based camera monitors a second object. At the same time, the event-based camera can function as a secondary sensor for the centrally positioned and aligned front camera and vice versa.

• Dabei zeigt 1 eine Draufsicht auf ein das erfindungsgemäße Verfahren ausführendes Fahrzeug in einer Fahrsituation.

Further advantageous embodiments of the method according to the invention for sensor fusion also result from the exemplary embodiment, which is described in more detail below with reference to the figure.

• This shows 1 a top view of a vehicle carrying out the method according to the invention in a driving situation.

1 shows a vehicle 1 traveling on a road 6 in a direction F in a top view. This is a driving situation in which there is heavy fog 7, which is indicated by hatching. A tractor 8 approaches the vehicle 1 in the opposite lane. The vehicle 1 comprises a plurality of sensors, with at least a first sensor 2.1 and a second sensor 2.2 being present, which have an at least partially overlapping detection area 3. In the exemplary embodiment shown, the first sensor 2.1 is a front camera and the second sensor 2.2 is a microphone array. Sensor data generated by the respective sensors 2.1, 2.2 are processed by a computing unit 4, which executes a machine learning model. The machine learning model is trained in the sense of the method according to the invention. This training can, for example, have been carried out by the vehicle manufacturer in special measurement campaigns or training campaigns. 1 also shows a camera image 9 taken by the front camera at the current moment. Due to the dense fog 7, the tractor 8 cannot yet be clearly recognized in the camera image 9. However, the position lights 10 of the tractor 8 shine through the fog 7. The computing unit 4 sets the first sensor 2.1 as the main sensor, for example based on a standard setting, depending on the extent to which information can be obtained by analyzing corresponding sensor data, or based on the current detection quality in the respective driving situation. Accordingly, the second sensor 2.2, here the microphone array, is defined as a secondary sensor. The respective sensor data are fed to the machine learning model, whereby the machine learning model determines a relevant information area 5 for the camera image 9 by determining a position relative to the vehicle 1 derived from the engine noise of the tractor 8. This relevant information area 5 is then extensively or profoundly analyzed by the machine learning model. In particular, the position lights 10 can be found and tracked more easily through appropriate image analysis. Thus, the detection probability for the tractor 8 is improved by analyzing the camera image 9. The computing effort required to analyze the corresponding camera image 9 is therefore also reduced, which improves the efficiency of the computing unit 4 when carrying out the method according to the invention or when analyzing data.

In a further example (not shown), an at least partially automated vehicle uses a camera-based system to detect the environment. In addition, infrared sensors are installed in the vehicle's headlights for adaptive headlight control. When driving on a country road, a vehicle in front is not initially recognized by the camera-based system due to its color being very similar to the background. However, since the information from the infrared sensors can detect the heat signature of the vehicle in front and this is fed to the machine learning model, the machine learning model is able to tell the camera system which camera image section is relevant for further data analysis by determining the relevant information area 5. In other words, the focus is directed to the relevant information area 5.

In a final example (also not shown), an at least partially automated vehicle also monitors its surroundings using a camera-based system. The vehicle also has an event-based camera that allows camera images to be captured with higher temporal resolution and higher dynamic range. The vehicle is driving towards the exit of a tunnel at high speed. The front camera cannot image the strong contrast due to the difference in brightness between the inside and outside of the tunnel and therefore cannot easily detect vehicles outside the tunnel. However, since the event-based camera has a higher dynamic bandwidth, it can already detect the oncoming vehicle. With appropriate data analysis by the machine learning model, a corresponding image section can be determined for the front camera and, if necessary, the focusing of the front camera can be adjusted. As soon as the lighting conditions improve for the front camera, the oncoming vehicle can be seen much more quickly.

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

• WO 2016/100814 A1 [0006]

Claims (10)

Method for sensor fusion, wherein a vehicle (1) detects its surroundings by means of at least a first (2.1) and a second sensor (2.2), the first and second sensors (2.1, 2.2) having an at least partially overlapping detection area (3), and wherein a computing unit (4) executes a machine learning model for processing sensor data generated by the sensors (2.1, 2.2), characterized by the following method steps carried out by the computing unit (4): a) determining one of the sensors (2.1, 2.2) as the main sensor, wherein the computing unit (4) uses the knowledge gained from an analysis of the sensor data generated by the main sensor to control a vehicle function, and wherein the computing unit (4) analyzes the sensor data of the main sensor using the machine learning model; b) defining at least one sensor (2.1, 2.2) that deviates from the main sensor as a secondary sensor; c) supplying sensor data generated by the at least one secondary sensor to the machine learning model; d) selecting a relevant information area (5) from the sensor data generated by the main sensor depending on the sensor data generated by the at least one secondary sensor by the machine learning model; and e) analyzing the part of the sensor data of the main sensor assigned to the relevant information area (5) to obtain the insights by the machine learning model. Procedure according to Claim 1 , characterized in that the computing unit (4) discards the part of the sensor data of the main sensor that deviates from the relevant information area (5) before analysis by the machine learning model. Procedure according to Claim 1 or 2 , characterized in that the sensor data generated by the at least one secondary sensor are prepared before processing by the machine learning model. Procedure according to one of the Claims 1 until 3 , characterized in that the sensor data generated by the at least one secondary sensor are normalized before processing by the machine learning model. Procedure according to one of the Claims 1 until 4 , characterized in that at least one detection parameter of the main sensor is adjusted depending on knowledge gained through an analysis of the sensor data of at least one secondary sensor. Procedure according to one of the Claims 1 until 5 , characterized in that the machine learning model takes sensor data generated by at least two secondary sensors into account for selecting the relevant information area (5), a relevance weighting being assigned to the secondary sensors depending on the current detection boundary conditions, and the machine learning model sensor data generated by secondary sensors with a high relevance weighting for the selection of the relevant information area (5). Procedure according to one of the Claims 1 until 6 , characterized in that in step a) the sensor (2.1, 2.2) is defined as the main sensor which has the highest detection quality for the current detection boundary conditions. Procedure according to Claim 7 , characterized in that the detection quality of the individual sensors (2.1, 2.2) is tracked over time and if the detection quality of a specific sensor (2.1, 2.2) deteriorates, a warning message is issued to a vehicle occupant. Procedure according to one of the Claims 1 until 8th , characterized in that the computing unit (4) maintains at least two machine learning models that differ in their degree of complexity and uses the less complex machine learning model to analyze the sensor data of the main sensor if, depending on the sensor data of the at least one secondary sensor, there is no relevant information area (5) in the from Sensor data generated by the main sensor can be identified. Procedure according to one of the Claims 1 until 9 , characterized in that at least a third sensor has a detection range that deviates from the main sensor, an object detected by the main sensor and identified by the computing unit (4) in the sensor data of the main sensor is tracked by the computing unit (4) and the computing unit (4) the third Sensor is defined as the new main sensor when the computing unit (4) predicts that the object will move from the detection area (3) of the main sensor into the detection area of the third sensor.

Images