Devices, methods, and machine learning systems for determining vehicle speed

Abstract

Devices, methods, and machine learning systems for determining the speed of a vehicle. The method includes: providing inputs to a first model, particularly a first generative model (102), depending on route information (106), probability variables (108), and in particular noise, and providing outputs to a second model, particularly a second physical model (104); determining the output of the first model (102) in response to the inputs of the first model (102), wherein the output of the first model (102) represents the speed (116), wherein the first model (102) includes a first component (102a) trained to map the inputs of the first model (102) determined depending on the route information (106) and probability variables (108) to an intermediate output (110) of the vehicle speed, wherein the first model (102) includes a second component (102b) trained to map the intermediate output (110) to the speed (116) depending on the output of the second model (104), wherein the output of the second model (104) represents a physical constraint (112) for the speed (116) or for the intermediate output (110).

Description

Background Technology

[0001] Vehicle speed can be determined from measurements using either physical or probabilistic models. While physical models are deterministic, the accuracy of speeds determined based on them depends on the observability of the relevant information. Probabilistic models, on the other hand, can model unobservable behavior. However, it is difficult to prove that vehicle speeds predicted by probabilistic models are physically feasible.

[0002] Reliable and accurate determination of vehicle speed is desirable. Summary of the Invention

[0003] This is achieved through a method, device, and machine learning system for determining the speed of a vehicle.

[0004] The method includes: providing inputs to a first model, particularly a first generative model, depending on route information, probability variables, and especially noise, and providing outputs to a second model, particularly a second physical model; determining the output of the first model in response to the inputs of the first model, wherein the output of the first model represents speed, wherein the first model includes a first component trained to map the inputs of the first model, determined depending on the route information and probability variables, to an intermediate output, wherein the first model includes a second component trained to map the intermediate output to speed depending on the output of the second model, wherein the output of the second model represents physical constraints for the intermediate output. The first model is a generative model for the intermediate output. The intermediate output is, for example, acceleration or intermediate speed, which will subsequently be converted into the vehicle's speed. Noise refers to some kind of noise, particularly sampled from a well-known noise distribution (such as a uniform or normal distribution). The first component generating the intermediate speed or acceleration may consist of many layers of an artificial neural network. The second model provides physical constraints for the intermediate output value (such as intermediate speed or acceleration). Therefore, the accuracy and reliability of the speed are significantly improved.

[0005] Technically, a physical model can be provided that takes no input at all and gives an output defined as follows: acceleration can be no less than -100 m / s² and no greater than 100 m / s². This model is still physically plausible, with not very strict boundaries. Preferably, the method includes: providing input to a second model (i.e., the physical model) depending on at least one vehicle state and / or route information; and determining the output of the second model in response to the input. Therefore, the boundaries can vary depending on the input.

[0006] Preferably, the physical constraints for a given time step are determined based on the vehicle speed in the previous time step, the forces applied to the vehicle, and / or the forces exerted by the vehicle. These vehicle parameters provide useful information that the physical model can use to determine limits for acceleration.

[0007] Preferably, the route information includes at least one of the following: geographical characteristics, particularly absolute elevation or road gradient characteristics; traffic flow characteristics, particularly time-related average traffic speed; road characteristics, particularly the number of lanes, road type, and / or road curvature; traffic control characteristics, particularly speed limits, the number of traffic lights, the number of specific types of traffic signs, the number of stop signs, the number of yield signs, and / or the number of pedestrian crossing signs; and weather characteristics, particularly rainfall, wind speed, and / or the presence of fog at a predetermined time. This information is particularly useful for predicting vehicle speeds using a generative model and for training such a generative model.

[0008] Preferably, the method includes: providing input to a third model depending on route information and speed; determining the output of the third model in response to the input of the third model, wherein the output of the third model represents a score indicating the realism of the speed estimate, wherein the third model is trained to map the input of the third model determined depending on the route information and speed to the output of the third model, the output of the third model representing a score indicating the realism of the speed estimate. This score can output true or false information regarding whether the input speed matches a real-world speed with the same route information (such as speed limits and gradients).

[0009] Preferably, the method includes: determining a velocity characteristic over time based on multiple inputs to a first model and multiple inputs to a second model. A sequence of velocity values ​​is determined as the velocity characteristic over time. This allows for the determination of a velocity trajectory with high reliability and accuracy.

[0010] Preferably, the method includes: providing route information as a first sequence of continuous or discrete values ​​over a time period; providing a probability variable as a second sequence of continuous or discrete values ​​over that time period; determining a third sequence of continuous or discrete values ​​for the speed characteristics over time by a first model, particularly a first recurrent neural network, depending on the values ​​of the first and second sequences; and determining a score by a third model, particularly a second recurrent neural network, depending on the values ​​of the first and third sequences. This means that the model processes the same common input values ​​and thus improves the reliability of the score as an indicator of accuracy.

[0011] Preferably, the method includes estimating the vehicle's exhaust characteristics based on its speed characteristics over time and / or a score. Therefore, the exhaust characteristics are based on a highly reliable and accurate speed. The score provides information on whether the speed used to estimate the exhaust characteristics is considered more or less reliable. This can further improve the estimation.

[0012] Preferably, an initial speed is determined, wherein subsequent speeds are determined based on the initial speed. This further improves accuracy.

[0013] Preferably, the initial speed is set to zero, or the initial speed is determined as the output of an initial speed model (particularly an artificial neural network) trained to map route information to an initial speed. Setting the initial speed to zero allows starting from a standstill (e.g., with the ignition switch on). The initial speed model improves the generated intermediate speed and acceleration, and thus improves the speed output of the second model.

[0014] Preferably, in response to training data that defines input data for a first model and a second model, the velocity is determined based on the outputs of the first model and the second model, wherein the output of a third model is determined, the output of the third model representing a score indicating the veracity estimate of the velocity, and wherein at least one parameter of the first model and / or the second model and / or the third model is determined based on the score.

[0015] Preferably, the method includes providing input data comprising speed, route information, intermediate outputs, and at least one vehicle state. The intermediate outputs can characterize the vehicle's intermediate speed or acceleration. This data is particularly useful for training.

[0016] A corresponding device is adapted to perform the steps of this method.

[0017] A machine learning system includes a first model, a second model, and a third model, and is adapted to perform the steps of the method. Attached Figure Description

[0018] Further advantages can be derived from the following description and figures. In the figures: Figure 1 The device used to determine the speed of a vehicle is depicted schematically. Figure 2 A machine learning system is schematically depicted. Figure 3 The steps in the method for determining velocity are illustrated schematically. Detailed Implementation

[0019] Figure 1Device 100 is schematically depicted. Device 100 may include at least one processor and a storage device, which can be adapted to perform the model and method steps described below.

[0020] The device 100 includes a first model, particularly a first generative model 102, and a second model, particularly a second physical model 104.

[0021] The first model 102 is a data-based model. The first model 102 includes a first component 102a, which is adapted to map inputs representing route information 106 and probability variables 108, particularly noise, to an intermediate output 110. The intermediate output 110 may represent the vehicle's intermediate speed or vehicle acceleration. In the examples below, the intermediate output 110 is referred to as intermediate speed or acceleration 110.

[0022] Route information 106 can be determined, for example, based on geospatial coordinates such as Global Navigation Satellite System data (e.g., Global Positioning System GPS data). In this example, route information 106 includes map features. Map features may include speed limits and / or topological information (such as slope). Route information 106 may also include weather conditions.

[0023] The second model 104 is adapted to determine an output that characterizes at least one physical constraint 112. The output of the second model 104 may depend on at least one vehicle state 114 and / or route information 106, particularly speed limits and gradients.

[0024] The first model 102 is a generative model for predicting the output characterizing velocity 116. The second model 104 is a physical model for physically modeling a reasonable velocity. The first model 102 includes a second component 102b, which is adapted to impose at least one physical constraint 112 to determine the velocity 116 in order to limit the velocity 116 to a reasonable velocity.

[0025] In this respect, device 100 is adapted to convert the intermediate output 110 (e.g., intermediate velocity or acceleration) of the first component 102a of the first model 102 into velocity 116, depending on at least one physical constraint 112. The second model 104 enriches the purely data-based generator output of the first model 102 with prior knowledge from physics. This provides a “hybrid” model because it combines a data-based model with a physics-based model.

[0026] Device 100 may include a third model 118. The third model 118 may be a data-based model. The third model 118 is adapted to map inputs representing route information 106 and speed 116 to an output representing a score 120 indicating the realism of the speed 116 estimate. The score 120 may output true or false information regarding whether the speed 116 represented by the output matches the real-world speed of a vehicle for the same speed limit and gradient.

[0027] Device 100 may include a training device 122 adapted to determine at least one parameter of a first model 102 and / or a second model 104 based on a score 120. In this example, to train the first model 102 and / or the second model 104, a gradient descent method (e.g., ADAM) may be iteratively repeated based on training data used to train the first model 102 and / or the second model 104 to map inputs representing route information 106 and probability variables 108 to an output, i.e., speed 116. A third model 118 may also be trained. Input data used for training may include speed 116, route information 106, intermediate outputs 110 (e.g., intermediate speed or acceleration), and / or at least one vehicle state 114. In this context, "training" refers to determining parameters for the first model 102 or the third model 118, or alternately, parameters for both, based on the training data. Optionally, the second model 104 (i.e., the physical model) may be modified to include some parameters that may be trained together with the first model 102.

[0028] In one example, device 100 includes a generative adversarial network (GAN), where a first model 102 and a second model 104 are configured as generators and trained, and a third model 118 is configured and trained to score the authenticity or falsity of a given velocity 116, as is known for Wasserstein GANs. This means that the third model 118 is a critic or discriminator that scores the authenticity or falsity of velocity 116 based on the Wasserstein distance. Jenson-Shannon divergence, and particularly regularized Jenson-Shannon divergence, can alternatively be used in the objective function used to train the GAN. Examples of regularization metrics include spectral normalization, gradient penalty, weight clipping, and layer normalization can be used.

[0029] Generative adversarial networks are configured as follows: Velocity 116 as a univariate time series of length T And given, among which This setting can be easily extended to cases where different time series have different lengths. Route information 106 serves as environmental conditions over time. And given, among which Generative adversarial networks, especially generator networks Learning from unknown distributions Samples drawn from The probability variable 108 in this example is from a known distribution (i.e. Noise extracted from ) To form a time series of noise The known distribution is, for example, a standard normal distribution or a uniform distribution. The time series of intermediate velocity or acceleration 110 is denoted as... .

[0030] Given some real-world data based on velocity x and environmental conditions c In this context, generative adversarial networks learn from distributed systems. Samples are extracted from the generator network. It can be trained to use another distribution. From the distribution Sampling is performed in this other distribution. Approximate real data distribution More specifically, minimizing the difference through adversarial training. The difference in this example is the Wasserstein distance. Instead, Jensen-Shannon divergence, especially regularized Jensen-Shannon divergence, can be used. In a given generator network... g And critique f In this case, it corresponds to the following minimax objective: in This means estimating the divergence, and the generator attempts to minimize that divergence. The Wasserstein distance is, for example, defined as... in F Composed of all 1-Lipschitz functions, i.e. .

[0031] Therefore, the minimum-maximum objective is In one example, it depends on the acceleration trajectory generated by the first model 102. To generate velocity trajectory Then, depending on the acceleration trajectory... acceleration value at time step t , speed trajectory velocity value at time step t Converted into velocity trajectory speed value : in It is the time difference between two consecutive time steps.

[0032] The first model 102, based on the data, predicts the acceleration value at each time step t. Given the initial velocity at the 0th time step, the second model 104 continuously integrates the acceleration values ​​over consecutive time steps to obtain the velocity trajectory.

[0033] For example, the starting speed can be determined based on one of the following three possibilities: We can assume that the vehicle starts from a stationary position, in which case the initial speed is zero.

[0034] An additional starting speed model, such as an artificial neural network, can be trained prior to the starting speed using the same input route and driver features obtained from a generative adversarial network. This additional starting speed model can then be used in the generator to determine the starting speed.

[0035] In addition, an additional starting velocity model can be trained within the training scheme of the generative adversarial network by adjusting the parameters of the starting velocity model, in order to deceive the discriminator of the generative adversarial network.

[0036] In the first model 102, either the acceleration value or the calculated velocity value can be used from time step t to predict the velocity value for time step t+1. This means that the velocity 116 in a time step can be determined from the velocity 116 of the previous time step.

[0037] As a result, acceleration follows a much smaller range than velocity and is centered at 0. The low variance in the magnitude makes it easier for the neural network used as the first model 102 to learn it, leading to faster and more stable training.

[0038] Additionally, the second model 104 may determine at least one physical constraint 112 for speed 116 based on at least one vehicle state 114 and / or route information 106. In this respect, speed 116 is determined based on at least one physical constraint 112.

[0039] The second model 104 is described in further detail below.

[0040] On the one hand, depending on at least one vehicle state 114, the speed to be generated is constrained to a value that is only physically reasonable.

[0041] In one example, the physical forces that a vehicle can exert are calculated, such as those exerted by the vehicle's engine or brakes. In another example, the forces facing the vehicle are calculated, i.e., the forces exerted on the vehicle. In this regard, for example, the forces exerted by air resistance, particularly for a given vehicle at a given speed, can be calculated. This relationship, and other related relationships, are known from physics / engineering literature, for example: Hans-Hermann Braess, Ulrich Seiffert: Vieweg HandbuchKraftfahrzeugtechnik. 2. Auflage, Friedrich Vieweg&Sohn VerlagsgesellschaftmbH, Braunschweig / Wiesbaden, 2001 Bernd Heißing, Metin Ersoy, Stefan Gies: Fahrwerkhandbuch: Grundlagen, Fahrdynamik, Komponenten, Systeme, Mechatronik, Perspektiven. Springer Vieweg2013 Dieter Schramm, Manfred Hiller, Roberto Bardini: Modellbildung undSimulation der Dynamik von Kraftfahrzeugen. Springer, Berlin / Heidelberg 2010 In this example, the force exerted by air resistance F 空气 Depends on: vehicle shape (especially the vehicle's front surface area A), air density ρ 空气 air drag coefficient c w And the current speed v.

[0042] In another example, rolling resistance F 滚动 Depends on: vehicle mass m 车辆 Gravity constant g, rolling resistance coefficient f 滚动 (This depends on the friction between the vehicle tires and the road surface) and the road gradient. .

[0043] In another example, slope resistance F 坡度 Depends on: vehicle massm 车辆 Gravity constant g, road slope .

[0044] In another example, maximum braking force F 制动max Depends on: Braking power p 制动 And velocity v.

[0045] In another example, braking friction F 制动摩擦 Depends on: vehicle mass m 车辆 Gravitational constant g, coefficient of friction µ k ,slope .

[0046] In another example, driving engine force F 驱动引擎 Depends on: Maximum engine power p max Vehicle speed v, tension factor r c The tension factor itself depends on the current speed and some constants that can be calculated for a given vehicle.

[0047] In another example, driving friction F 驱动摩擦 Depends on: vehicle mass m 车辆 Number of drive shafts n 驱动轴 Total number of shafts n 轴 Gravitational constant g, coefficient of friction µ k ,slope .

[0048] The variables mentioned above are, for example, measured variables, given by vehicle specifications, route specifications, or a combination thereof. Vehicle specifications include, for example, vehicle mass. m 车辆 And the front surface area A. Route specifications include, for example, slope. The combination of vehicle specifications and route specifications includes, for example, the coefficient of friction, which depends on the tires, road surface type, and fundamental physics (such as the gravitational constant). If the exact specifications are unknown, it is also possible to estimate vehicle-specific parameters from data of other, particularly similar, vehicles.

[0049] It is possible to calculate the physically plausible range of accelerations at time step t+1. Given the force described above, which is partially dependent on the velocity 116 at time step t, calculate the physically reasonable range of accelerations in the example. .

[0050] Therefore, the acceleration at time step t+1 Limited by the second model 104 in and .

[0051] and The values ​​of are determined by the second model 104, serving as the minimum and maximum accelerations for a given vehicle and its parameters. In one example, this results in: in In this respect, acceleration value The following function can be used as physical constraint 112 to limit: exist Figure 2 The diagram schematically depicts an exemplary machine learning system 200 for determining speed 116. Instead of using a sigmoid compression function... Alternative compression functions can be constructed, such as those parameterized by a neural network or handcrafted. Any compression function can be used here—as long as it maps the input acceleration to an output in the interval [0, 1].

[0052] In this example, route information 106 is defined by a continuous sequence of values ​​over time within time period 202. Instead of a continuous sequence of values ​​over time within time period 202, the sequence of values ​​can be discrete.

[0053] Route information 106 may include at least one of the following: geographical characteristics, particularly absolute altitude or road gradient characteristics 106a; traffic flow characteristics, particularly time-related average traffic speed; road characteristics, particularly the number of lanes, road type, and / or road curvature; traffic control characteristics, particularly speed limit characteristics 106b, the number of traffic lights, the number of specific types of traffic signs, the number of stop signs, the number of yield signs, and / or the number of pedestrian crossing signs; and weather characteristics, particularly rainfall, wind speed, and / or the presence of fog at a predetermined time. In this regard, intermediate output 11, such as intermediate speed or acceleration, may be determined based on route information 106 including route characteristics.

[0054] The first model 102 and the third model 104 include recurrent neural networks. The recurrent neural network can be implemented as a long short-term memory network, a gated recurrent unit, a transformer architecture, or a vanilla recurrent neural network. The first model 102 includes a first component, which is a recurrent neural network adapted to process values ​​in a value sequence over time period 202. In this example, the value of probability variable 108 is determined to be from, in particular, a standard normal distribution or a uniform distribution. P z noise in mid-sample z ~ P z Any other distribution can also be used. The input to the first component 102a is noise. z and inputs for i = 0, ..., T c i , where T is the number of discrete values ​​in time period 202 determined by route information 106. In this respect, noise z and input c i These can be concatenated to form the input of the first component 102a. The output of the first component 102a in response to this input is an intermediate output 110. The second component 102b is adapted to process the intermediate output 110 depending on the output of the second model 104. In this example, at least one physical constraint 112 is provided as the output of the second model 104.

[0055] The third model 118 is adapted to process the values ​​in the value sequence of the time period as input to the third model 118. The third model 118 includes a second recurrent neural network, which is adapted to process the values ​​in the value sequence of the time period 202. The input to the third model 118 is the input... c i ,enter c i It depends on route information 106 and is indicated as x iThe speed is determined by 116. In this respect, the input... c i and x i These can be concatenated to form the input to a third model 118. The third model 118, in response to its input, outputs a score 120, which, for example, indicates the authenticity of velocity 116 by y=1 and the falsity of velocity 116 by y=0. This score is not necessarily binary. It can be a regressive value, for example, positive for true and negative for false. In the Wasserstein generative adversarial network mentioned above, y is a continuous value.

[0056] In this example, the machine learning system 200 is adapted such that both the first model 102 and the third model 118 process the same value of the route information 106 in the value sequence of the time period within the same period.

[0057] Route information 106 can be defined by a continuous or discrete value sequence over time within time period 202. In this example, probability variable 108 and speed 116 are defined by a continuous or discrete value sequence over time within time period 202.

[0058] In this example, time period 202 can be divided into time steps that include the value used to determine velocity 116. In this example, each time step determines a velocity 116.

[0059] In this respect, device 100 is adapted to: provide route information 106 as a first value sequence of continuous or discrete values ​​over time within time period 202; provide probability variables 108 as a second value sequence of continuous or discrete values ​​over time within time period 202; determine a third value sequence of continuous or discrete values ​​of the characteristics of speed 116 over time by a first model 102, in particular a first recurrent neural network, depending on the values ​​of the first and second sequences; and determine a score 120 by a third model 118, in particular a second recurrent neural network, depending on the values ​​of the first and third sequences.

[0060] The following is for reference. Figure 3 The described method assumes the implementation of a trained first model 102. The method includes steps for determining an exemplary period of velocity 116. In the case of using an initial velocity model, it is assumed that the initial velocity model has also been trained. In one aspect, a third model 118 may be present specifically during training. In another aspect, the initial velocity model may also be trained during training. The initial velocity model may be integrated into the first model 102. However, after training, the first model 102 and the second model 104 may be used independently of the third model 118.

[0061] The method for determining speed 116 includes step 302: providing input to a first model 102 based on route information 106 and probability variable 108. In this example, the probability variable is noise, such as white noise. Step 302 may include: providing input to an initial speed model based on route information 106 and speed 116.

[0062] In this example, for route information 106, a continuous first value sequence of road gradient characteristic 106a and speed limit characteristic 106b over time is provided within time period 202.

[0063] In this example, probability variable 108 is provided as a continuous second value sequence over time within time period 202.

[0064] Subsequently, in step 304, velocity 116 is determined based on the first model 102. First component 102a generates acceleration 110. Second component 102b generates velocity 116 based on acceleration 110 and at least one physical constraint 112. In this example, the value of velocity 116 is determined by the acceleration value provided above. The function determines this. To generate a new velocity trajectory, route information 106 and physical model information, i.e., at least one physical constraint 112, are provided one step at a time. The velocity 116 is generated one step at a time. At least one physical constraint 112 at time t and the velocity 116 at time t are used to calculate the acceleration 110 at time t. This information is used to generate the velocity 116 at time t+1. The initial velocity 116 at time t=1 can be given as zero.

[0065] Apply the same process for time t+1 to calculate the velocity 116 at time t+2. Repeat this process to generate the velocity trajectory.

[0066] Alternatively, intermediate velocity can be used instead of acceleration as intermediate output 110. In this case, the intermediate velocity can be limited to a reasonable value in each process step.

[0067] The output of the first model 102 represents the speed 116. In this example, the input of the first model 102, determined from the route information 106 and the probability variable 108, is mapped to the output representing the speed 116.

[0068] The initial speed can also be determined in step 304. In this regard, speed 116 is determined based on the initial speed. In this example, the initial speed is determined as the output of the initial speed model. Specifically, an artificial neural network can map route information 106 to the initial speed. Alternatively, for example, the initial speed can be set to zero when it is assumed that the vehicle is stationary.

[0069] In one aspect, for the characteristics of the velocity 116 over time, a continuous third value sequence is determined by the first model 102, in particular the first recurrent neural network, depending on the values ​​of the first and second sequences.

[0070] The method further includes step 306: providing input to the second model 104. In this example, the input to the second model is at least one vehicle state 114 and / or route information 106.

[0071] Subsequently, in step 308, at least one physical constraint 112 for speed 116 is determined based on at least one vehicle state 114 and / or route information 106. In one example, the physical constraint 112 is determined based on the vehicle speed 116 of the previous cycle, the force applied to the vehicle, and / or the force applied by the vehicle.

[0072] The method may include step 310: providing input to the third model 118 based on route information 106 and speed 116. Values ​​of the first sequence and the third sequence may be provided as input to the third model 118.

[0073] The method may include step 312: in response to the input of the third model 118, determining a score 120 indicating the verisimilitude estimate of the speed 116 based on the output of the third model 118. The output of the third model 118 represents the score 120. In this respect, the third model 118 is trained to map the input of the third model 118, determined based on the route information 106 and the speed 116, to the output representing the score 120.

[0074] In one aspect, the score 120 is determined by the third model 118, particularly the second recurrent neural network, which depends on the values ​​of the first and third sequences.

[0075] The method may include step 314: determining at least one parameter of a first model 102 or a third model 118, or both, based on the score 120. The method may also include determining at least one parameter of the initial velocity model in step 316.

[0076] When the route information 106 includes route features, particularly road gradient characteristics 106a and / or speed limit characteristics 106b, the method may include extracting route features from map information.

[0077] In this example, all the variables required to calculate these physical quantities are given by vehicle specifications (e.g., vehicle mass, gear ratio), route specifications (e.g., gradient), or a combination of both (e.g., vehicle resistance). Vehicle specifications and the laws of physics provide all the physical formulas needed to calculate the dependencies. The second model 104 can be implemented in various ways and is designed to model the various aforementioned aspects in more or less detail. If the exact specifications are unknown, it is also possible to estimate vehicle-specific parameters or dependencies from data from other, particularly similar, vehicles.

[0078] The trained model, device 100, and / or machine learning system 200 can perform the methods described above for predicting vehicle speed characteristics over time.

[0079] Vehicle speed characteristics can be used to simulate vehicle emissions. This simulation can be used for probabilistic judgments on compliance with exhaust limits, and for optimizing the parameterization of vehicle engines or vehicle control. It can also be used for predictive vehicle control, powertrain management, and / or for torque regeneration in vehicle powertrains that include internal combustion engines, electric engines, or both.

[0080] After training, the third model 118 may not exist. When the third model 118 exists after training, it can be used to determine or distinguish safe and unsafe values ​​for the velocity characteristics over time. This ensures higher accuracy through physical compliance, and this in turn ensures higher accuracy in downstream tasks.

[0081] Therefore, downstream tasks can avoid unstable and / or insecure systems, because physically unreasonable inputs can make downstream systems unstable and / or insecure.

[0082] In a preferred embodiment, the speed characteristics over time are determined in response to multiple inputs to the first model 102, depending on multiple inputs to the second model 104. Furthermore, the vehicle's exhaust characteristics can be estimated based on the speed characteristics over time.

Claims

1. A method for determining the speed (116) of a vehicle, characterized in that: The input to the first model (102) is provided based on route information (106), probability variables (108), and the output of the second model (104); wherein the first model (102) is a generative model, and the second model (104) is a physical model, and wherein the route information (106) includes at least one of the following: geographical characteristics; traffic flow characteristics; road characteristics; traffic control characteristics; and weather characteristics, and wherein the probability variables (108) are noise; the output of the first model (102) is determined in response to the input of the first model (102), wherein the output of the first model (102) represents speed (116), wherein the first model (102) includes a first component (102a), a first set of Component (102a) is trained to map the input of a first model (102) determined by route information (106) and probability variables (108) to an intermediate output (110) of vehicle speed, wherein the intermediate output (110) is acceleration, wherein the first model (102) includes a second component (102b) which is trained to map the intermediate output (110) to speed (116) depending on the output of a second model (104), wherein the output of the second model (104) characterizes physical constraints (112) for the intermediate output (110), wherein the physical constraints (112) for the time step are determined based on the vehicle speed (116) in the previous time step, the force applied to the vehicle, and / or the force applied by the vehicle.

2. The method according to claim 1, characterized in that, The input to the second model (104) is provided (306) depending on at least one vehicle state (114) and / or route information (106), and the output of the second model (104) is determined (308) in response to the input to the second model (104).

3. The method according to claim 1, characterized in that: The input to the third model (118) is provided (312) depending on the route information (106) and the speed (116); the output of the third model (118) is determined (314) in response to the input of the third model (118), wherein the output of the third model (118) represents a score (120) indicating the truthiness estimate of the speed (116), wherein the third model (118) is trained to map the input of the third model (118) determined depending on the route information (106) and the speed (116) to the output of the third model (118), the output of the third model (118) representing a score (120) indicating the truthiness estimate of the speed (116).

4. The method according to claim 1, characterized in that, The characteristics of the velocity (116) over time are determined based on multiple inputs to the first model (102) and multiple inputs to the second model (104).

5. The method according to claim 4, characterized in that: Provide route information (106) as a first value sequence of continuous or discrete values ​​over time within a time period (202); provide probability variables (108) as a second value sequence of continuous or discrete values ​​over time within a time period (202); and determine the characteristics of the speed (116) over time by the first model (102) depending on the values ​​of the first and second sequences as a third value sequence of continuous or discrete values. And the score is determined by the third model (118) depending on the values ​​of the first and third sequences (120).

6. The method according to claim 4 or 5, characterized in that, The exhaust characteristics of the vehicle are estimated based on the speed characteristics over time and / or the score (120).

7. The method according to claim 1, characterized in that, Determine the initial velocity (304), where the subsequent velocity (116) is determined based on the initial velocity.

8. The method of claim 7, wherein the starting speed is set to zero, or wherein the starting speed is determined as the output of a starting speed model trained to map route information (106) to the starting speed.

9. The method according to claim 1, characterized in that: In response to training data that defines the input data of the first model (102) and the second model (104), a velocity (116) is determined depending on the output of the first model (102) and the second model (104), wherein the output of the third model (118) is determined to represent a score (120) indicating the truthiness estimate of the velocity (116), and wherein at least one parameter of the first model (102) and / or the second model (104) and / or the third model (118) is determined depending on the score (120).

10. The method according to claim 9, characterized in that, It provides input data including speed (116), route information (106), intermediate output (110) and at least one vehicle status (114).

11. The method of claim 1, wherein the geographical characteristic is at least one of absolute altitude or road slope characteristic (106a); and the traffic flow characteristic is time-related average traffic speed; and the road characteristic is at least one of number of lanes, road type and / or road curvature; and the traffic control characteristic is at least one of speed limit characteristic (106b), number of traffic lights, number of specific types of traffic signs, number of stop signs, number of yield signs and / or number of pedestrian crossing signs; and the weather characteristic is at least one of rainfall, wind speed and / or the presence of fog at a predetermined time.

12. The method according to claim 5, wherein the first model (102) is a first recurrent neural network and the third model (118) is a second recurrent neural network.

13. The method of claim 8, wherein the initial velocity model is an artificial neural network.

14. A device (100) for determining the speed (116) of a vehicle, characterized in that, The device is adapted to perform the steps of the method according to any one of claims 1 to 13.

15. A machine learning system (200) comprising a first model (102), a second model (104) and a third model (118), and adapted to perform the steps of the method according to any one of claims 1 to 13.

16. A computer program product, characterized in that, The computer program product includes computer-readable instructions that, when executed by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 13.

Images