Method and apparatus for fusion of state hypotheses

Abstract

The invention relates to a device for determining the state of a vehicle, said device being arranged to determine K different state hypotheses in relation to the state of the vehicle, where Kgt; 1. The apparatus is further arranged to determine a feature vector describing a driving condition of the vehicle, and to determine K weights for the respective K state hypotheses based on the feature vector and in accordance with the machine learning unit. The apparatus is further arranged to fuse the K state hypotheses into a fused state hypothesis related to the state of the vehicle based on the respective K weights.

Description

Technical Field

[0001] The present invention relates to a method and a corresponding device for fusing state hypotheses from different sensors of a vehicle. Background Art

[0002] A vehicle may include a plurality of different sensors, which are, for example, arranged to detect sensor data related to the vehicle environment and / or vehicle dynamic measurement parameters. The sensor data of the plurality of sensors can be evaluated to determine the state of the vehicle, where the state can in particular be related to the position of the vehicle in the road network on which the vehicle is traveling (e.g., related to the position in a digital map of the road network).

[0003] Sensor data from different sensors may lead to a plurality of different state hypotheses regarding the current state of the vehicle. This document relates to the technical task of fusing these different state hypotheses in an efficient and robust manner to achieve particularly precise state determination. Summary of the Invention

[0004] This is considered to be solved by each independent claim. Advantageous embodiments are described in particular in the dependent claims. It should be noted that additional features of claims dependent on an independent claim may form an independent invention that is independent of the combination of all features of the independent claim, either without the features of the independent claim or only in combination with some of the features of the independent claim, and this invention may be the subject of an independent claim, a divisional application, or a later application. This also applies to the technical teachings described in the specification, which may constitute inventions independent of the features of the independent patent claims.

[0005] According to one aspect, a device for determining the state of a (motor) vehicle is described. The state of the vehicle may include the position of the vehicle, in particular the position on the road on which the vehicle is traveling and / or the position in a digital map of the road network on which the vehicle is traveling.

[0006] The device is arranged to determine K different state hypotheses related to the state of the vehicle, where K>1 (e.g., K>2, or K>3).

[0007] The device may be arranged to determine K different state hypotheses based on the sensor data of a plurality of sensors of the vehicle. Exemplary sensors are

[0008] · one or more environment sensors (such as one or more cameras, radar sensors, lidar sensors, and / or ultrasonic sensors), each of the one or more environment sensors being arranged to detect sensor data related to the environment of the vehicle; and / or

[0009] · One or more vehicle sensors (e.g., speed sensors, wheel speed sensors, inertial measurement units, etc.), where the one or more vehicle sensors are respectively configured to detect sensor data related to driving dynamic measurement parameters of the vehicle (e.g., driving speed, wheel speed, acceleration, steering angle, etc.); and / or

[0010] · A receiver of a global navigation satellite system (GNSS), in particular a GPS receiver.

[0011] Different state hypotheses can be determined respectively based on sensor data of at least some different sensors among a plurality of sensors of the vehicle. Alternatively or additionally, different methods can be used to determine different state hypotheses.

[0012] K different state hypotheses can include, for example:

[0013] · A state hypothesis determined based on sensor data of the GNSS receiver (where the sensor data of the GNSS receiver is evaluated, for example, in combination with a digital map of the road network on which the vehicle travels); and / or

[0014] · A state hypothesis determined based on lane markings on the road on which the vehicle travels (where the lane markings are identified, for example, based on camera data of a camera of the vehicle); and / or

[0015] · A state hypothesis determined based on one or more landmarks in the environment of the vehicle (where the one or more landmarks are shown in a digital map and / or identified based on camera data of a camera of the vehicle); and / or

[0016] · A state hypothesis determined based on mileage data of the vehicle (e.g., driving speed and / or steering angle of the vehicle).

[0017] Therefore, different state hypotheses can be determined for the current time point based on the corresponding current sensor data K. In a corresponding manner, for a series of successive time points, K current state hypotheses corresponding to the current time point can be respectively determined based on the corresponding current sensor data.

[0018] The device is further configured to determine a feature vector that describes the (corresponding current) driving condition of the vehicle. The feature vector can be determined based on sensor data of a plurality of sensors of the vehicle. In addition, the feature vector can include a plurality of features (e.g., respectively as vector components). Exemplary features are:

[0019] · Characteristics related to the driving speed and / or acceleration of the vehicle; and / or

[0020] · Characteristics related to the traffic density on the road on which the vehicle travels; and / or

[0021] · Features related to one or more weather conditions in the vehicle environment; and / or

[0022] · Features related to light conditions in the vehicle environment; and / or

[0023] · Features related to the degree and / or type of precipitation (e.g., rain or snow) in the vehicle environment; and / or

[0024] · Characteristics related to the road type (e.g., highway, urban road, rural road, etc.) of the lane on which the vehicle is traveling; and / or

[0025] · Features related to the presence or absence of obstacles to GNSS (especially GPS) signals; exemplary obstacles being tunnels, relatively tall buildings (e.g., in the city center) and / or bridge underpasses.

[0026] Thus, based on sensor data from one or more sensors of the vehicle, a feature vector can be determined that describes the currently existing driving situation of the vehicle. The feature vector can be updated at a series of successive time points to describe the currently existing driving situation.

[0027] The device is also arranged to determine K weights for the corresponding K state hypotheses based on the feature vector and according to the machine learning unit. Here, the K weights can be normalized, for example such that the sum of the K weights corresponds to a normalization value (e.g., 1).

[0028] The device can in particular be arranged to determine K confidence values for the corresponding K state hypotheses based on the feature vector and according to the machine learning unit. Here, the K confidence values can indicate or show the credibility and / or quality of the corresponding K state hypotheses. The K confidence values may not be normalized. The K confidence values can be normalized according to a normalization unit to determine the corresponding K weights. The normalization unit can use the Softmax function to normalize the K confidence values. Thus, K normalized weights can be determined particularly efficiently and precisely for the fusion related to the driving situation of the K state hypotheses.

[0029] The device is also arranged to fuse the K state hypotheses into a fusion state hypothesis related to the vehicle state based on the corresponding K weights. The device can in particular be arranged to weight the K state hypotheses with the corresponding K weights. The fusion state hypothesis can be determined precisely and effectively as the weighted average of the K state hypotheses.

[0030] Thus, a device is described that is designed to effectively implement the fusion related to the driving situation of K different state hypotheses. Thus, a particularly precise fusion state hypothesis can be provided.

[0031] The machine learning unit may include one or more artificial neural networks. Additionally, the machine learning unit may have been trained based on training data, where the training data may include, for example, multiple training data sets. Each training data set may respectively have training feature vectors to describe corresponding existing determined driving situations. Additionally, each training data set may respectively include K training confidence values and / or K training weights, which are used to fuse K state hypotheses in the corresponding existing determined driving situations. Alternatively or additionally, each training data set may respectively give K training state hypotheses and fused training state hypotheses for the determined driving situations. The training data may be determined through experiments (in the context of the driving of one or more vehicles) and / or based on simulations.

[0032] To train the machine learning unit, based on the training feature vectors and K individual training state hypotheses, the machine learning unit may determine a predicted, fused current state hypothesis from the training data set. This may be compared with the fused training state hypothesis in the training data set. This may be implemented for multiple different training data sets from the training data in a corresponding manner to determine the value of an error function (e.g., the average (possibly squared) deviation between each current state hypothesis and the corresponding training state hypothesis). Based on the backpropagation algorithm, the error function may be propagated back to each parameter of the machine learning unit in order to adapt the parameters and thus (iteratively) train the machine learning unit.

[0033] Thus, supervised learning of the machine learning unit may have been performed. The machine learning unit may be trained to give the credibility and / or quality (where the credibility and / or quality are described by trust values) of different state hypotheses for corresponding existing driving situations (which are respectively described by feature vectors).

[0034] By using such a trained machine learning unit, a particularly precise driving situation-related fusion can be achieved.

[0035] The device may be set to determine, for a series of successive time points, K different state hypotheses respectively corresponding to the current time point, and to determine a feature vector by which the driving situation of the vehicle at the corresponding current time point is described (for this purpose, sensor data existing at the corresponding current time point of one or more sensors of the vehicle may be used).

[0036] The device may also be set to determine K weights based on the feature vector and according to the machine learning unit for the corresponding K state hypotheses. The K state hypotheses may be fused into a fused state hypothesis corresponding to the current time point based on the corresponding K weights.

[0037] Thus, the device can be set up, in particular iteratively, to update the state of the vehicle at a series of successive time points based on a fusion state hypothesis determined for a corresponding time point. This can be achieved using a Kalman filter.

[0038] Thus, an exact fusion of the state hypotheses can be continuously achieved during the operation of the vehicle.

[0039] The individual (single) state hypotheses may relate to different (intermediate) time points. It can be caused (e.g., for the respective K state hypotheses based on K Kalman filters) that the individual state hypotheses are predicted at a common time point respectively. In other words, it can be caused (e.g., with the respective Kalman filters) that the K state hypotheses to be fused are synchronous respectively (and thus each relate to a respective common time point). The K time-synchronized state hypotheses can be precisely fused into a fusion state hypothesis (for the corresponding common time point) based on the respective K weights.

[0040] Alternatively or additionally, a Kalman filter can be used to update the state of the vehicle at a series of successive time points. Here, based on the state at the previous time point, the predicted state at the subsequent time point can be determined using a model. Furthermore, the fusion state hypothesis at the subsequent time point can be used as a measurement update to determine the updated state at the subsequent time point based on the predicted state and based on the fusion state hypothesis.

[0041] The device can also be set up to operate the vehicle according to the fusion state hypothesis (at a series of successive time points), in particular for driving functions for the automatic longitudinal and / or lateral guidance of the vehicle. Thus, a particularly reliable automated operation of the vehicle can be caused.

[0042] According to another aspect, a (road) motor vehicle (in particular a passenger car or a truck or a bus or a motorcycle) is described which includes the device described herein.

[0043] According to another aspect, a method for determining the state of a (motor) vehicle is described. The method includes determining different K state hypotheses of the state of the vehicle, where K>1. Furthermore, the method includes determining a feature vector describing the driving situation of the vehicle, and determining K weights for the corresponding K state hypotheses based on the feature vector and according to a machine learning unit. The method also includes fusing the K state hypotheses into a fusion state hypothesis for the vehicle state based on the respective K weights.

[0044] According to another aspect, a software (SW) program is described. The SW program can be designed to be executed on a processor (e.g., on a vehicle controller) and thereby execute the method described in this document.

[0045] According to another aspect, a storage medium is described. The storage medium may include a software program configured to execute on a processor and thereby perform the methods described herein.

[0046] It should be noted that the methods, apparatuses, and systems described herein may be used alone or in combination with other methods, apparatuses, and systems described herein. In addition, any aspect of the methods, apparatuses, and systems described herein may be combined in various ways. In particular, the features of the claims may be combined in various ways. In addition, the features listed in parentheses should be understood as optional features. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The present invention will be described in more detail below with reference to embodiments. The accompanying drawings illustrate

[0048] Figure 1a exemplary components of a vehicle;

[0049] Figure 1b exemplary state assumptions;

[0050] Figure 2 an exemplary apparatus for determining a fused state assumption of a vehicle state; and

[0051] Figure 3 a flowchart of an exemplary method for determining a fused state assumption of a vehicle state. DETAILED DESCRIPTION

[0052] As mentioned at the beginning, this document relates to the effective and accurate determination of vehicle conditions. The state of a vehicle may be described by a state vector, which may have one or more state parameters. Exemplary state parameters for describing the state of a vehicle include:

[0053] · The position of the vehicle (e.g., a position within a digital map describing the road network on which the vehicle is traveling); and / or

[0054] · The speed of the vehicle; and / or

[0055] · The orientation of the vehicle; and / or

[0056] · The (longitudinal and / or lateral) acceleration of the vehicle.

[0057] Figure 1a An exemplary vehicle 100 with a plurality of different environmental sensors 102 is shown. Exemplary environmental sensors 102 are one or more cameras, one or more radar sensors, one or more ultrasonic sensors, one or more lidar sensors, etc. The vehicle 100 also includes a plurality of vehicle sensors 103, such as speed sensors, acceleration sensors, inertial measurement units, steering sensors, wheel speed sensors, etc.

[0058] The (control) device 101 of the vehicle 100 is configured to evaluate the sensor data of a plurality of sensors 102, 103 of the vehicle 100 to determine the state of the vehicle 100. Here, this state, in particular Figure 1b the state hypothesis 111 shown in, can be updated recursively and / or iteratively at a series of successive time points based on the sensor data currently available. A Kalman filter can be used for this purpose.

[0059] The device 101 can also be configured to manipulate one or more longitudinal and / or lateral guidance actuators 104 of the vehicle 100 based on the determined state of the vehicle 100, for example to automatically longitudinally and / or laterally guide the vehicle 100. For example, the determined state of the vehicle 100 can indicate the position of the vehicle 100 within the lane traveled by the vehicle 100. One or more actuators 104 can be manipulated according to the determined state in order to automatically longitudinally and / or laterally guide the vehicle 100 within the determined lane of the road.

[0060] Figure 1b An exemplary lane 110 in which the vehicle 100 is arranged is shown. Based on the sensor data of the plurality of sensors 102, 103 available at the determined time point n, a plurality of different state hypotheses 111 (i.e., a plurality of different state hypotheses 111) related to the state of the vehicle 100 can be determined at the time point n. The different state hypotheses 111 can already be determined here based on the sensor data of different sensors 102, 103. Exemplary state hypotheses 111 related to the position of the vehicle 100 can be:

[0061] · A state hypothesis 111 that is determined based on the GNSS (Global Navigation Satellite System) coordinates of the vehicle 100, in particular based on GPS coordinates; and / or

[0062] · A state hypothesis 111 that is determined based on odometer data (e.g., steering sensor and / or speed sensor and / or wheel speed sensor); and / or

[0063] · A state hypothesis 111 that is determined based on one or more landmarks in the environment of the vehicle 100 (wherein one or more landmarks can be identified and / or located based on the sensor data of an environmental camera and / or a radar sensor and / or a lidar sensor); and / or

[0064] · A state hypothesis 111 that is determined based on one or more lane markings on the lane 110 traveled by the vehicle 100 (wherein one or more lane markings can be identified and / or located based on the sensor data of an environmental camera).

[0065] Thus, for K time points n, different state hypotheses 111 can be determined (where K > 1). Different K state hypotheses can be fused to determine a fused state hypothesis, which can be used, for example, in the context of a Kalman filter (as a measurement update) to determine the current state of vehicle 100 (at time point n).

[0066] The quality, in particular the accuracy and / or reliability, of the different state hypotheses 111 can depend on the respective prevailing driving conditions. For example, the quality of a state hypothesis 111 based on GNSS coordinates can depend on the weather conditions currently prevailing. In addition, the quality of a state hypothesis 111 depending on the camera data of the environmental camera 102 can depend on the prevailing lighting conditions. In addition, the quality of a state hypothesis 111 depending on the sensor data of the environmental sensor 102 can depend on the current traffic density.

[0067] The device 101 can be set to fuse the K different state hypotheses 111 according to the currently prevailing driving conditions of vehicle 100. For this purpose, the sensor data of multiple sensors 102, 103 can be evaluated to describe the currently prevailing driving conditions, where, for example, the driving conditions can be described by a feature vector having a set of determined vector components for different features of the driving conditions. Exemplary features of the driving conditions are:

[0068] · The driving speed of vehicle 100; and / or

[0069] · The traffic density in the environment of vehicle 100; and / or

[0070] · The lighting conditions in the environment of vehicle 100; and / or

[0071] · The degree and / or type of precipitation; and / or

[0072] · The presence of obstacles to GNSS, in particular GPS signals, in the environment of vehicle 100 (such as bridges, tunnels or buildings); and / or

[0073] · The type of road 110 on which vehicle 100 is traveling.

[0074] The device 101 can also be set to determine a driving-condition-specific fusion method for fusing the different K state hypotheses 111 based on the feature vector used to describe the driving conditions (present at the current time point n). Then, the driving-condition-specific fusion method can be used to determine the combined and / or fused state hypothesis.

[0075] The driving-condition-specific fusion method can be determined based on a machine learning unit. Here, the machine learning unit may have been pre-trained based on training data.

[0076] Figure 2 An exemplary apparatus 200 for fusing K different state hypotheses 111 into a fused state hypothesis 204 is shown. The apparatus 200 includes a fusion unit 213 that is configured to fuse the K different state hypotheses 111 into the fused state hypothesis 204 using K weights 203 for the respective K state hypotheses 111. Here, within the fusion unit 213, the fused state hypothesis 204 can be determined as a weighted average of the K state hypotheses 111 (using the respective K weights 203).

[0077] Thus, as a fusion method, calculating a weighted average of the K state hypotheses 111 can be used. The K weights 203 for weighting the respective K state hypotheses 111 can be adapted according to the corresponding existing driving conditions, in particular according to the feature vector 201 for describing the corresponding existing driving conditions. Thus, driving condition-specific weights 203 and thus a driving condition-specific fusion method can be used.

[0078] The K weights 203 can be determined based on a machine learning unit 211, where the machine learning unit 211 includes, for example, one or more artificial neural networks. The machine learning unit 211 can be configured to determine K confidence values 202 for the K different state hypotheses 111 based on the feature vector 201. The confidence value 202 of a state hypothesis 111 can give the quality and / or credibility of the state hypothesis 111. Here, the quality and / or credibility can increase as the confidence value 202 increases.

[0079] The K confidence values 202 can be normalized in a normalization unit 212 to determine the respective K weights 203, the sum of which results in a normalized value, for example, 1. The normalization unit 212 can use the Softmax function for this purpose.

[0080] Training data can be used to train the machine learning unit 211, where the training data can include a plurality of (e.g., 1000 or more, or 10000 or more) training data sets. Each training data set can respectively have a training feature vector 201 (for describing a determined driving condition) and a related set of training confidence values 202. The training confidence values 202 can be measured under the determined driving conditions. In other words, the training confidence values 202 can indicate what quality and / or credibility the different state hypotheses 111 actually have when a determined driving condition exists. The training data can be determined experimentally during the operation of one or more vehicles 100.

[0081] Based on the training data, supervised learning of the machine learning unit 211 can be performed (e.g., using the backpropagation algorithm). Here, the machine learning unit 211 can be trained based on the training data for the purpose of reducing the determination of the error function. The error function can depend on the training confidence value 202 shown in the training data and the actual confidence value 202 determined by the machine learning unit 211 (based on the training feature vector 201 shown in the training data) (e.g., depending on the (average, possibly squared) deviation between the training confidence value 202 and the corresponding actual confidence value 202).

[0082] Alternatively or in addition, the machine learning unit 211 can be trained together with the entire fusion device 200. The training data can be used here together with training data sets, each of which has a pair of training feature vectors 201 and a fusion training state hypothesis 204 (as a ground truth reference). Each training data set can also show K individual training state hypotheses 111 (which are determined based on the corresponding sensor subsets 102, 103). In the context of the supervised learning process for training the machine learning unit 211, a predicted actual state hypothesis 204 can be determined for the training feature vector 201 respectively (based on the fusion device 200 and taking into account the K individual training state hypotheses 111), and compared with the corresponding training state hypothesis 204 to determine the value of the error function (e.g., the average (possibly squared) deviation). Based on the error function, the parameters of the machine learning unit 211 can be adapted (by using the backpropagation algorithm).

[0083] Therefore, in order to train the machine learning unit 211, the ground truth fusion training state hypothesis 204 can be used, which is provided, for example, by a (very precise) sensor for positioning (such as DGPS (Differential Global Positioning System) or RTK (Real - Time Kinematics)).

[0084] The error function that enables the training of the machine learning unit 211 can be determined by comparing the actual state hypothesis 204 predicted by the unit 211 and the fusion training state hypothesis 204 according to the ground truth reference. Therefore, the end - to - end training of the machine learning unit 211 can be performed within the fusion device 200. According to backpropagation, the error can be passed back to each weight 203, each confidence value 202, and finally to the parameters of the machine learning unit 211, thereby adapting the parameters of the machine learning unit 211 during training.

[0085] Since the error function acts on the same part as the final prediction (i.e., the fusion state hypothesis 204), the quality of the final prediction can be ensured in a particularly reliable manner. In addition, a credibility test of each of the K confidence values 202 and / or weights 203 with respect to precise positioning can be achieved.

[0086] Alternative or supplementary methods for generating ground truth references (i.e., for determining the training state hypothesis 204) include, for example, post-processing optimization methods, which can be applied globally after a test drive of the vehicle 100 to optimize the original position estimate.

[0087] Thus, during operation of the vehicle 100, a feature vector 201 can be determined for a determined time point n, which describes the driving situation present at the determined time point n. Additionally, a set of confidence values 202 can be determined based on the machine learning unit 211, and this set of confidence values is converted into a corresponding set of (normalized) weights 203 by the normalization unit 212. This set of weights 203 can be used to fuse different state hypotheses 111 to determine a fused state hypothesis 204 for a determined state. The fused state hypothesis 204 can be used to determine the current state of the vehicle 100 at the determined time point n. Here, the fused state hypothesis 204 can be used, for example, as a measurement update in a Kalman filter.

[0088] It should be noted that a set of confidence values 202 determined by the machine learning unit 211 can be (possibly independently of the fusion described in this document) used to evaluate a corresponding set of (individual) state hypothesis sets 111. For example, the vehicle 100 can have a driving function that operates based on one or more (individual) state hypotheses 111 (possibly without accessing the fused state hypothesis 204). The driving function can take into account the confidence values 202 and / or weights 203 determined by the machine learning unit 211 for one or more (individual) state hypotheses 111. Thus, the quality and / or reliability of the driving function can be improved.

[0089] Therefore, the device 101 described herein can be set to determine different K state hypotheses 111 related to the state of the vehicle 100, where K > 0 (possibly K = 1), and to determine a feature vector 201 for describing the (current) driving situation of the vehicle 100. The device 101 can also be set to determine K weights 203 and / or K confidence values 202 based on the feature vector 201 and according to the machine learning unit 211 for the corresponding K state hypotheses 111. Additionally, the device 101 can also be set to operate the driving function of the vehicle 100 based on the K state hypotheses 111 and based on the corresponding K confidence values 202 and / or weights 203 (possibly without causing fusion of the K state hypotheses 111). Thus, the quality and / or reliability of the driving function can be improved.

[0090] Figure 3A flowchart showing a method 300 (possibly implemented by a computer) for determining the state of a (motor) vehicle 100 is shown. The state can be related to the position of the vehicle 100. Additionally, the state can be described by a state vector having one or more state parameters. The method 300 includes determining 301 K different state hypotheses 111 related to the state of the vehicle 100, where K > 1. Each state hypothesis 111 can be determined based on sensor data from one or more sensors 102, 103 of the vehicle 100. Here, for different state hypotheses 111, at least partially different sensors 102, 103 and / or at least partially different methods can be used to determine the corresponding state hypothesis 111.

[0091] The method 300 further includes determining 302 a feature vector 201 that describes the driving situation of the vehicle 100. The feature vector 201 can include one or more features for describing the driving situation in which the vehicle 100 is in. Each feature can be determined based on sensor data from one or more sensors 102, 103 of the vehicle 100.

[0092] Furthermore, the method 300 includes determining 303 K weights 203 based on the feature vector 201 and according to the use of a machine learning unit 211 for the corresponding K state hypotheses 111. The machine learning unit 211 may have been pre-trained based on training data. Here, in particular, the parameters of one or more neural networks can be trained.

[0093] The method 300 also includes fusing 304 the K state hypotheses 111 into a fused state hypothesis 204 for the state of the vehicle 100 based on the corresponding K weights 203. Here, in particular, a weighted average of the K state hypotheses 111 can be determined (using the K weights 203) to determine the fused state hypothesis 204.

[0094] The measures described herein can lead to a particularly efficient and accurate fusion of different state hypotheses 111. Thus, the quality of state determination and the quality of the autonomous driving function of the vehicle 100 (using the fused state hypothesis 204) can be improved.

[0095] The present invention is not limited to the illustrated embodiments. In particular, it should be noted that the description and the drawings are only used to illustrate the principles of the proposed methods, devices, and systems.

Claims

1. An apparatus (101) for determining the state of a vehicle (100); wherein the apparatus (101) is configured to - determine K different state hypotheses (111) related to the state of the vehicle (100), where K > 1; - determine a feature vector (201) that describes the driving condition of the vehicle (100); - based on the feature vector (201) and according to a machine learning unit (211), determine K weights (203) for the corresponding K state hypotheses (111); and - based on the corresponding K weights (203), fuse the K state hypotheses (111) into a fused state hypothesis (204) related to the state of the vehicle (100).

2. The apparatus (101) according to claim 1, wherein - the apparatus (101) is configured to determine the different K state hypotheses (111) based on sensor data of a plurality of sensors (102, 103) of the vehicle (100); wherein the plurality of sensors (102, 103) particularly includes - one or more environmental sensors (102), each of the one or more environmental sensors being configured to detect sensor data related to the environment of the vehicle (100); - one or more vehicle sensors (103), each of the one or more vehicle sensors being configured to detect sensor data related to driving dynamic measurement parameters of the vehicle (100); and / or - a receiver of a global navigation satellite system; and - determine different state hypotheses (111) based on sensor data of at least some different sensors (102, 103) among the plurality of sensors (102, 103) of the vehicle (100).

3. The apparatus (101) according to any one of the preceding claims, wherein the apparatus (101) is configured to determine the feature vector (201) based on sensor data of a plurality of sensors (102, 103) of the vehicle (100).

4. The apparatus (101) according to any one of the preceding claims, wherein the apparatus (101) is configured to - based on the feature vector (201) and according to the machine learning unit (211), determine K confidence values (202) for the corresponding K state hypotheses (111); wherein the K confidence values (202) indicate the credibility and / or quality of the corresponding K state hypotheses (111); - normalize the K confidence values (202) according to a normalization unit (212) to determine the corresponding K weights (203); wherein the normalization unit (212) particularly uses a Softmax function to normalize the K confidence values (202).

5. The apparatus (101) according to any one of the preceding claims, wherein the apparatus (101) is configured to - weight the K state hypotheses (111) with the corresponding K weights (203); and - Determine the fused state hypothesis (204) as a weighted average of the K state hypotheses (111).

6. The apparatus (101) according to any one of the preceding claims, wherein - The machine learning unit (211) includes one or more artificial neural networks; and / or - the machine learning unit (211) has been trained based on training data including a plurality of training data sets; wherein the training data sets respectively include: - training feature vectors (201) for describing determined driving situations; and - K training state hypotheses (111) and a fused training state hypothesis (204) for the determined driving situation.

7. The apparatus (101) according to any one of the preceding claims, wherein the apparatus (101) is configured to, for a series of successive time points respectively, - determine different K state hypotheses (111) for a corresponding current time point; - determine a feature vector (201) that describes the driving situation of the vehicle (100) at the corresponding current time point; - determine K weights (203) for the corresponding K state hypotheses (111) based on the feature vector (201) and according to the machine learning unit (211); and - fuse the K state hypotheses (111) into the fused state hypothesis (204) for the corresponding current time point based on the corresponding K weights (203).

8. The apparatus (101) according to any one of the preceding claims, wherein the apparatus (101) is configured to iteratively update the state of the vehicle (100) at a series of successive time points based on the fused state hypothesis (204) determined for the corresponding time points, in particular using a Kalman filter.

9. The apparatus (101) according to any one of the preceding claims, wherein - the feature vector (201) includes a plurality of features; and - the plurality of features particularly includes, - Characteristics related to the driving speed and / or acceleration of the vehicle (100); and / or - features related to the traffic density on the lane (110) on which the vehicle (100) is traveling; and / or - features related to the weather conditions in the environment of the vehicle (100); and / or - features related to the light conditions in the environment of the vehicle (100); and / or - features related to the degree and / or type of precipitation in the environment of the vehicle (100); and / or - features related to the road type of the lane (110) on which the vehicle (100) is traveling; and / or - features related to obstacles to signals for a global navigation satellite system.

10. The apparatus (101) according to any one of the preceding claims, wherein the different K state hypotheses (111) include: - a state hypothesis (111) determined based on sensor data of a global navigation satellite system; and / or - a state hypothesis (111) determined based on lane markings on the lane (110) on which the vehicle (100) is traveling; and / or - a state hypothesis (111) determined based on one or more landmarks in the environment of the vehicle (100); and / or - a state hypothesis (111) determined based on the mileage data of the vehicle (100).

11. The device (101) according to any one of the preceding claims, wherein the state of the vehicle (100) includes the position of the vehicle (100), in particular the position on the lane (110) on which the vehicle (100) is traveling.

12. A method (300) for determining the state of a vehicle (100); wherein the method (300) comprises: - determining (301) K different state hypotheses (111) related to the state of the vehicle (100), where K > 1; - determining (302) a feature vector (201) that describes the driving condition of the vehicle (100); - determining (303) K weights (203) for the corresponding K state hypotheses (111) based on the feature vector (201) and according to a machine learning unit (211); and - fusing (304) the K state hypotheses (111) into a fused state hypothesis (204) related to the state of the vehicle (100) based on the corresponding K weights (203).

Images